okto-nexus
OfficialAn MCP coordination hub for teams of AI agents working in one repository: identity, presence, messages, handoffs, artifacts, events, and monitoring — over HTTP at /mcp with an agent nxs_ key.
Identity & presence —
agent_whoamifor your profile/role/communication contract,agent_registerto update your own profile (self-only, catalog-validated capabilities),agent_list/agent_get/capability_listfor reachability-scoped discovery.Sessions —
session_openreturns asession_id+ one-timesession_secret;session_heartbeatkeeps you present for broadcasts;session_closeends it (idempotent).Workspaces —
workspace_resolvemaps an absoluteproject_rootto a deterministicworkspace_id;workspace_listis admin/permission-gated with paths hidden by default.Messages & inbox —
message_createpersists a message and fans out per-recipient inbox deliveries (direct, capability, role, tag, broadcast, or mixed targets);inbox_count/inbox_peek/inbox_pull/inbox_ack/inbox_extend/inbox_historygive leased, at-least-once pull/ack with redelivery;message_statustracks per-recipient delivery lanes.message_get/message_list/message_waitare MIGRATED shims returning replacements.Channels —
channel_create/channel_listfor organizational labels (not ACLs, not delivery).Handoffs (executable work) —
handoff_createwith a required target and visibility (public/eligible/private), optional payload, trace_id, acceptance_criteria (feature_verification) and depends_on (feature_dag);handoff_list_available(with optional long-poll),handoff_claim(single winner, returns confidential payload),handoff_complete,handoff_verify(pass/fail with feedback, claimant can't self-verify),handoff_reject,handoff_cancel,handoff_getfor status/result.Events & monitoring —
event_get(cursor-paginated, filterable by type/agent/task/handoff/trace, profiles default/summary/full),event_cursorfor an O(1) pre-flight anchor,event_waitfor explicit long-poll viatimeout_seconds>0.Artifacts —
artifact_putregisters file/text/json/markdown/html content (or imports a workspace-contained path) into managed storage;artifact_getretrieves under audience-scoped visibility (unauthorized reads returnNOT_FOUND).Derived view —
shared_md_renderatomically rewrites a workspaceshared.mdwith agents/sessions, open tasks, open handoffs, and recent events (never handoff payloads).Health & metadata —
coordination_healthwindowed report (1h/24h/7d; needsfeature_health+health.read),nexus_infofor package/schema/surface revisions and live feature flags,tag_listfor the operator-managed tag catalog.Ephemeral monitor tokens —
poll_token_issue/poll_token_renew/poll_token_revokemint short-lived read-onlynxsept_bearers for helper monitors (events + inbox read only, no MCP or mutation).
Note: this surface is the HTTP MCP feature configuration — memory (feature_memory) and harness runtime tools are absent here; several capabilities (verification, DAG, HITL, health, trace) are gated behind opt-in flags that default off.
Okto Nexus
Local-first coordination for teams of AI agents.
Okto Nexus is an MCP server and operator hub for agents working in the same repository. It gives them durable identities, presence, messages, inboxes, handoffs, artifacts, an event log, governance controls, and a live dashboard without requiring a cloud broker.
okto-nexus serve exposes the complete hub on one port:
/mcp— MCP over streamable HTTP, authenticated as an agent;/api/v1— operator REST APIs, read-only monitor endpoints, and SSE;/— the bundled React dashboard.
MCP is available through HTTP. Coordination records live in one
SQLite database in WAL mode. Optional metrics, referenced workspace files, and
the derived shared.md view live outside that database.
Release fact | Value |
Package |
|
Python |
|
MCP surface | Use |
MCP resources | 12 versioned reference resources |
MCP prompts | 0 |
Surface revision | 62 |
Database schema | Read the installed revision through |
Storage | local SQLite/WAL catalog + adapter-backed artifact payloads |
Contents
Related MCP server: Nexus Memory
Why Nexus
One coordination space per project. An absolute
project_rootis canonicalized and hashed into a deterministicworkspace_id. Every client that resolves the same real path joins the same workspace.Durable delivery. Messages fan out into per-recipient inbox lanes with leases, redelivery, acknowledgements, delivery status, and optional read receipts.
Single-winner work dispatch. Handoffs support atomic claim, leases, rejection, cancellation, optional verification, and optional dependency graphs.
Explicit identity and presence. Operators create identities and API keys; agents open sessions and heartbeat to remain present.
Targeted routing. Direct, capability, role, tag, broadcast, mixed, and direct-with-fallback strategies share one validated grammar.
Governed communication. Permissions, communication scopes, versioned policies, quotas, guardrails, groups, and human approval can restrict writes without exposing the control plane to agents.
Observable by design. Monotonic event IDs, cursor reads, long-poll, replay export, SSE, health aggregates, and a live dashboard expose what the team is doing.
Local-first and fail-closed. Configuration, target grammars, catalogs, workspace paths, API keys, and state transitions are validated before writes.
Token-aware MCP docs. First-use guidance remains resident; deeper reference material is available through versioned MCP resources on demand.
R4 execution status
For current local and remote execution setup, use the
R4 operations guide. Nexus embeds
Core locally; the Connector application is required only on remote execution
hosts. Agents connect directly to HTTP /mcp. Legacy native launch factories
and Nexus MCP stdio have been retired.
This is a development release. Four adapter catalog entries do not establish four qualified native integrations. The acceptance ledger and evidence index distinguish installed contract tests from final provider/platform/multi-host qualification. Final G0–G3 acceptance remains open.
Install
The recommended install includes the HTTP hub, dashboard, local embedding provider, and tokenizer:
uv tool install "okto-nexus[serve]"
okto-nexus serveEquivalent with pipx:
pipx install "okto-nexus[serve]"Available extras:
Extra | Includes | Use it when |
none | CLI and coordination library | You are embedding Nexus components |
| FastAPI, Uvicorn, Core, dashboard | You need HTTP without Torch/model dependencies |
| sentence-transformers | You want the local embedding provider separately |
| HTTP stack, Core, embeddings, tokenizer | You want the complete supported hub |
| pytest, FastAPI, Uvicorn, httpx | You are developing or testing Nexus |
The published wheel and sdist contain the compiled dashboard. Node.js is only needed when rebuilding the frontend from source.
From a checkout:
git clone https://github.com/OktoLabsAI/okto-nexus.git
cd okto-nexus
uv sync --extra dev
# Complete HTTP build:
uv sync --extra serve --extra devStart the hub
okto-nexus serveDefaults:
listen address:
0.0.0.0:8202(all IPv4 interfaces);dashboard:
http://127.0.0.1:8202/;MCP:
http://127.0.0.1:8202/mcp;data directory:
~/.okto_nexus;database:
~/.okto_nexus/nexus.db;initial workspace context: the current directory. The dashboard keeps a saved selection when present and otherwise may open the all-workspaces view.
Useful variants:
okto-nexus serve --project-root /absolute/path/to/project
okto-nexus serve --host 127.0.0.1 --port 8202 # loopback-only access
okto-nexus serve --trust-mode strict
okto-nexus serve --embedding-mode localFrom another computer, use http://<server-network-IP>:8202/. Remote dashboard
access requires an operator login configured locally in Settings → Operator access.
MCP and Connector clients continue to authenticate with agent API keys.
--host overrides OKTO_NEXUS_HOST; either can restrict the listening address.
On first use, open Agents → New agent in the dashboard. Create one identity
per participant and copy its nxs_... key immediately: Nexus stores only the
hash and shows plaintext only at creation or regeneration. Capability and tag
values must first exist in Registry before an identity can use them.
Run harness teams
A worker that runs under a harness needs four things before the hub starts its work from a broadcast handoff. Each item below is a separate gate.
Discovery roots. Pass them to
serve:--harness-root Papproves one installation root. Repeat the flag for each trusted directory. The hub accepts at most 32 roots.--pi-install-root Pand--pi-node Pset the Pi installation. Use both flags together. Use--pi-nodeonly for a Node executable.Each path must be absolute and must exist. The hub resolves symbolic links and uses the resolved path. A path that does not resolve is refused at startup. The flags do not add directories to PATH.
Execution grant. The agent needs a current grant with the
execute_workaction. Finding an installation does not give authority.Agent opt-in. Set
broadcast_handoff_auto_claimfor the agent, or for the global default, in the runtime policy. The default is off for each agent.Startup gate. The hub starts auto-claim only when both
runtime_handoff_auto_claim_enabled(default true; envOKTO_NEXUS_RUNTIME_HANDOFF_AUTO_CLAIM_ENABLED) andfeature_harness_integrations(default true) are on. If one is off, the hub does not start new work. Recovery of existing work continues.
Details: R4 operations guide.
Connect an MCP client
Configure the selected MCP client to use this streamable-HTTP endpoint:
http://127.0.0.1:8202/mcpSend the existing agent key in the Authorization: Bearer nxs_... header.
The HTTP endpoint still accepts older query-key configurations during
migration, but new configurations should keep the key out of the URL.
Generic client configuration shape (use the fields your client supports):
{
"mcpServers": {
"okto-nexus": {
"url": "http://127.0.0.1:8202/mcp",
"headers": {"Authorization": "Bearer nxs_REPLACE_ME"}
}
}
}Agent pre-flight
Every authenticated agent should do this on its first turn:
Call
agent_whoami(), use the returnedagent_idconsistently, and treat itsroleandcommunication.contentas the default operating contract.Call
workspace_resolve(project_root=<absolute cwd>).Call
session_open(agent_id=<you>, workspace_id=<resolved id>)and retain the returnedsession_idand one-timesession_secret.Check
inbox_count(agent_id=<you>); pull and acknowledge backlog.Anchor monitoring with
event_cursor(project_root=..., agent_id=<you>, stream="workspace").
The full procedure is available at
okto-nexus://reference/preflight.
The role guides responsibilities, operating perspective, and decision boundaries. The communication block guides tone, format, language, verbosity, structure, and agent-to-agent as well as user-facing communication. Follow both unless the user explicitly directs otherwise for the current task or interaction. That task-scoped override does not modify the Nexus profile or bypass permissions, policies, guardrails, approvals, communication scope, safety rules, or higher-priority host instructions. Capabilities are routing claims, not authorization or persona.
Direct message
{
"project_root": "/absolute/path/to/project",
"from_agent_id": "researcher",
"subject": "API findings",
"body": "The endpoint is idempotent; details are attached.",
"target": {
"strategy": "direct",
"agent_id": "implementer"
},
"from_session_id": "ses_...",
"session_secret": "..."
}The response names the resolved recipients and delivery count. The recipient uses:
inbox_count(agent_id="implementer")
inbox_pull(agent_id="implementer", session_id="ses_...", session_secret="...")
inbox_ack(agent_id="implementer", message_ids=[...],
session_id="ses_...", session_secret="...")Messages are delivered through the inbox. event_get and event_wait are
observability tools, not delivery.
Long-poll
event_wait is a snapshot when timeout_seconds is omitted, null, or 0.
Long-poll is explicit:
event_wait(
project_root="/absolute/path/to/project",
agent_id="researcher",
stream="workspace",
cursor=123,
timeout_seconds=25,
profile="summary"
)Always continue from next_cursor. The waiter uses SQLite
PRAGMA data_version plus bounded sleep polling; HTTP runs the blocking wait
in a worker thread so it does not block the shared event loop.
Architecture
MCP HTTP REST / SSE / SPA CLI
\ | /
+----------- inbound adapters and transport auth ----------+
|
application services
identity · messages · inbox · handoffs · events · artifacts
permissions · policies · approvals · guardrails · memory · health
|
domain models and pure rules
|
+------------------------- outbound ports ---------------------------+
| SQLite repositories | files/shared.md | waiter | telemetry | embed |
+--------------------------------------------------------------------+bootstrap() resolves configuration, creates the store, applies migrations,
wires repositories, telemetry, embeddings, and approval execution, then seeds
the reserved operator, backfills the capability catalog, and creates built-in
permission/communication presets. create_server() lazily imports FastMCP and
registers the effective tools and resources. serve wraps the same composition
in FastAPI/Uvicorn; tail and admin are separate CLI adapters.
Important boundaries:
domain code contains state machines, routing, IDs, and invariants;
application services own the core coordination use cases; operator CRUD/maintenance routes may drive repositories and units of work directly;
inbound adapters translate MCP, HTTP, SSE, and CLI calls;
outbound adapters implement SQLite, files, telemetry, tokenization, embeddings, and waiting;
coordination truth is durable in SQLite; bounded process caches are implementation details, not authoritative state.
HTTP surfaces and authentication
Surface | Direct loopback client | Remote client |
SPA shell/assets, | Public | Public |
Dashboard REST data/control plane | Keyless operator trust | Human operator session |
MCP | Active | Active |
EPT monitor endpoints | Scoped | Scoped |
MCP-over-HTTP connections always represent an agent. The dashboard/REST loopback trust path represents
the local operator. Browser-origin checks protect mutating operator routes, and
binding to 0.0.0.0 preserves direct loopback access. Forwarded requests never
inherit loopback trust. Agent REST/MCP credentials remain separate from UI sessions.
For remote UI access, open the dashboard locally at http://127.0.0.1:8202,
then set a username and password under Settings → Operator access. Remote
browsers use Operator sign-in, not an agent API key. Passwords require at
least 12 characters and are salted and hashed. Sessions expire after eight
hours and use HttpOnly, SameSite cookies (Secure over HTTPS). Sign out revokes
the session; changing the password revokes all sessions. Local access can reset
a forgotten password; remote changes require the current password.
The separate operator-auth.db in the Nexus home stores credential hashes,
session hashes, and bounded human-action audit records, without agent keys.
Startup no longer issues an agent key for dashboard login. HTTPS protects
credentials and session traffic when using the dashboard across a network.
Permanent agent keys can authenticate REST and MCP, but helper monitors should
receive only a short-lived ephemeral poll token (nxsept_...). EPTs are bound
to the issuing session, agent, and workspace and are accepted only as
Authorization: Bearer nxsept_...; query api_key and x-api-key are rejected
for this token type. The bearer is valid only on:
GET /api/v1/eventsandGET /api/v1/events/cursor;GET /api/v1/inbox/countandGET /api/v1/inbox/peek.
They cannot call MCP or mutate state.
Dashboard
The bundled dashboard provides:
Graph — toggle between detailed agent cards and compact activity-sized circles, with profile colours, live presence status badges, recent message flow, unread traffic, open handoffs, and claimed relationships;
Messages — inbox lanes, peer conversations, undelivered targeting outcomes, receipts, and optional semantic search;
Handoffs — a six-column Kanban including
VERIFYING, claim details, dependency state, verification, cancellation, and results;Meta-harness — an OpenAI-style chat for private or broadcast messages and handoffs, with agent filtering, aggregated acknowledgement flags, and replies/results in one timeline;
Artifacts — paginated deliverables with list/Explorer-style grid views, type-aware icons, rich previews, metadata, and managed-payload downloads;
Events — filtered event history, trace navigation, and live SSE updates;
Memory — durable memory browse/search/curation when
feature_memoryis enabled;Workspaces — sessions, analytics, and coordination health;
Agents — identities, keys, activation, roles, capabilities, metadata, colors, permissions, tags, inbound/outbound audiences, communication style, and steering;
Registry — operator-managed capability and tag vocabularies;
Policies — versioned policies and per-agent bindings;
Guardrails — groups, versioned content rules, assignments, and scrubbed denial audit;
Communication — versioned communication presets and bindings;
Approvals — pending, decided, and archived human-in-the-loop actions. Use Archive to dismiss a stale or expired request and remove it from the pending badge. The Archived tab retains the original request, decision, result, archive time, and operator. Archiving does not grant permission, send a native answer, cancel an already submitted action, or resume a waiting harness. New decisions on archived requests are blocked. The operator-only
POST /api/v1/approvals/{approval_id}/archiveendpoint is idempotent;Settings — runtime-manageable settings, feature flags, retention, and database maintenance; metrics use their own header-menu panel.
Semantic search requires embedding_mode=stub or local. off returns
EMBEDDINGS_UNAVAILABLE on the REST search endpoint. The local provider
needs the embeddings extra; stub is deterministic and is intended for tests
or demonstrations.
Coordination model
Workspaces, agents, and sessions
Agents are global identities; workspaces represent canonical project roots.
Most coordination tools accept
project_root.session_openandshared_md_renderconsume a resolvedworkspace_id.An authenticated agent can update only its own profile with
agent_register, subject toidentity.update_profileandidentity.update_capabilities. Operators create identities.Authenticated discovery is reachability-scoped.
agent_listandagent_gethide unreachable peers;capability_listreturns the complete catalog but filters owner identities.workspace_listis permission-gated; absolute paths require a separate permission.Cross-workspace errors are intentionally operation-specific:
WORKSPACE_MISMATCHfor ownership guards,NOT_FOUNDfor hidden artifact or memory reads, andDEPENDENCY_NOT_FOUNDfor dependency creation.
Presence is explicit. A session is considered present while its heartbeat is
within presence_ttl_seconds. A trust-sensitive write advances the heartbeat
only when it authenticates with that session's credentials; in
trust_mode=open, a credential-free write advances no session. Read-only tools
do not heartbeat. Call session_heartbeat during long read-only or idle periods
and session_close when finished.
Messages and inboxes
message_create persists one message and resolves recipients at send time.
Each recipient gets a durable delivery row in its global inbox. The lanes are:
Lane | Meaning |
| Available to pull |
| Pulled and protected by an in-flight lease |
| Acknowledged |
| Dead-lettered after exhausting delivery claims; not redelivered automatically |
An expired delivered lease becomes pullable again. Delivery is therefore
at-least-once until acknowledgement. message_status lets the sender inspect
each recipient's lane. Every pull/redelivery emits message.delivered; ack
emits message.read. By default, ack also sends one synthetic read-receipt
message to the sender; receipts do not recursively create receipts.
Channels are organizational labels, not ACLs or delivery mechanisms. Access still intersects with permissions, policy, guardrails, and communication reachability. Message retention can remove aged messages and their deliveries, including unread or in-flight rows, so durability is bounded by configured retention and explicit database reset.
Communication intent
Choose the coordination mechanism by intended outcome: if another agent is expected to perform work or produce a deliverable, create a handoff. If the recipient only needs to know something or reply, use a message.
Handoff — executable work. Any request to execute, investigate, change, build, test, review, validate, or otherwise produce a deliverable must use
handoff_create; a direct message must not be its sole record. Target the intended assignee directly when known, or use a capability, role, tag, mixed, broadcast, or direct-with-fallback target when the first eligible claimant should own the work. The handoff is the canonical, operator-visible record of ownership, lifecycle, result, and verification. Include the objective, context, scope, constraints, and expected deliverable; add acceptance criteria, dependencies, and a verifier when applicable and the correspondingfeature_verification/feature_dagflag is enabled. Those fields are rejected while their feature is off.Broadcast message — shared alignment. Use it for shared context, decisions, announcements, discoveries, risk or blocker alerts, and general alignment across every selected reachable recipient. It informs recipients but assigns no owner and creates no task lifecycle. If anyone is expected to act, create one or more handoffs.
Direct message — conversation and informal coordination. Use it for status checks, questions, clarifications, acknowledgements, focused context exchange, and informal coordination. It may discuss an existing handoff, but new executable work requires a handoff; reference that handoff in the conversation. Use
handoff_getfor canonical lifecycle state and direct messages for contextual updates or blocker explanations.
A broadcast message is informational fan-out to many recipients. A handoff
with a broadcast target is a claim pool for one executor. If several agents
must produce independent results, create separate handoffs.
Routing
There are seven strategies:
Strategy | Descriptor | Notes |
|
| One named identity |
|
| One or any of several registered capabilities |
|
| Exact role match |
|
| Registered tag selector |
|
| Present workspace agents for messages; globally registered eligible agents for handoffs |
|
| Non-empty union of non-broadcast rules |
| direct plus | Handoffs only |
Messages support direct, capability, role, tag, broadcast, and mixed. Omitting a message target means broadcast. Handoffs require an explicit target and additionally support direct-with-fallback.
Tag selectors use AND across keys and OR across values. Rich
In/NotIn/Exists/DoesNotExist expressions are also supported.
Capability and tag names fail closed against operator-managed catalogs.
Target resolution is intersected with presence where applicable and with the caller's effective communication reach/audience. Separate enforcement layers then allow, deny, limit, or intercept the write:
per-agent permissions and recipient/rate limits;
versioned policy action rules and quotas;
content guardrails;
optional HITL approval interception.
A channel itself adds no ACL. Visibility controls who may see an item; eligibility controls who may claim it.
Event log
Events are immutable after insertion during normal operation. event_id is
globally monotonic and never reused. Retention may delete old rows, so retained
history can contain gaps.
Streams are workspace, agent, and handoff. Supported filters are
type, agent_id, task_id, handoff_id, and trace_id. Authenticated event
reads require events.read and omit actors outside the caller's communication
reach, except the caller's own and system events.
Response profiles:
Profile | Behavior |
| Safe trim: preserves contextual fields while removing empty or duplicated data; oversized event payloads can yield a follow-up hint |
| Aggressive projection: omits heavy bodies/payloads and returns follow-up hints |
| Raw debugging escape hatch |
event_get is non-blocking. event_cursor returns the current end in O(1).
event_wait long-polls only when timeout_seconds > 0. Dashboard SSE already
provides operator UI updates; agent monitoring remains MCP polling/long-poll or
the EPT read-only REST plane.
Handoffs
OPEN -> CLAIMED
OPEN -> REJECTED | CANCELLED
CLAIMED -> COMPLETED
CLAIMED -> VERIFYING when acceptance criteria exist
CLAIMED -> REJECTED claimant rejects
CLAIMED -> OPEN lease expires
VERIFYING -> COMPLETED verifier passes
VERIFYING -> CLAIMED verifier fails; executor reworksOnly one agent wins handoff_claim. A claim returns the confidential payload.
Available-list responses omit payload; handoff_get includes it only for the
claimant. Events and synthetic notifications never carry the payload.
With feature_verification=true, creation may include
acceptance_criteria and verify_by. The executor cannot verify its own
result. A failed verdict persists feedback, returns the handoff to CLAIMED,
and renews the lease.
With feature_dag=true, creation may include depends_on. The persisted
status remains OPEN, but blockedness is derived. Blocked items are excluded
from handoff_list_available and claim returns DEPENDENCY_NOT_MET until every
dependency is COMPLETED.
Artifacts and shared.md
Artifacts are published as text/json/markdown/html content or imported from a
workspace-contained file. Path containment is checked after canonical
resolution; escapes return PATH_OUTSIDE_WORKSPACE. Text content is bounded by
max_inline_bytes.
Payload bytes and free-form metadata do not live in SQLite. The database keeps
only the minimal searchable and authorization catalog. The default local
ArtifactStore adapter writes each payload and its manifest.json below
~/.okto_nexus/artifacts/<workspace>/<agent>/<artifact-id>/. Imported files
are copied into that managed location, so the artifact survives later changes
to the original workspace file. The dashboard's Artifacts screen pages
through the managed payload catalog, filters by one or more producing agents
and a production date interval, and previews or downloads the selected payload.
The publisher's effective outbound audience is frozen on the artifact. The
same audience controls artifact.created visibility and artifact_get.
Unauthorized, cross-workspace, and missing reads all return NOT_FOUND.
shared_md_render atomically rewrites a derived workspace shared.md with four
fixed sections: relevant agents/sessions, open tasks, open handoffs, and recent
events. It never becomes the source of truth and never renders handoff payloads.
Authenticated callers need shared_md.render.
Permissions, policies, guardrails, and approvals
The operator control plane manages:
permission presets and per-agent effective permissions;
capability/tag catalogs and communication scopes;
versioned attachable policies with deny-overrides and quota windows;
versioned communication guidance returned privately by
agent_whoami;agent groups and versioned guardrails assigned by scope and priority;
one-shot HITL approval decisions and operator steering.
Guardrail denials persist scrubbed metadata, not rejected raw content. When
feature_hitl=true and a policy requires approval, message_create or
handoff_create may return status: "pending_approval". Do not resend the
action; watch the returned approval ID.
Opt-in features
All seven flags default to off:
Flag | Effect |
| Accept and project |
| Enable new |
| Enable handoff acceptance criteria and verdicts |
| Enable handoff dependencies |
| Register three memory MCP tools and show Memory UI; restart required |
| Enable |
| Enable REST replay export |
Nuances:
approval history and decisions remain available when interception is off;
disabling verification or DAG blocks new contracts/dependencies, while already persisted workflow state remains enforceable and decidable;
workspace health REST remains available when the MCP health flag is off;
CLI replay export is operator-shell access and is not gated by
feature_replay;memory REST supports operator curation independently, but MCP memory tools are registered only when the flag is on at startup.
Chain limits
Each root message has limits for the chain that it starts. A new root copies
the values when the hub creates it. A change in Settings applies to new
roots only. A cap error returns QUOTA_EXCEEDED to the caller.
Setting | Default | Range | Env var | Flag |
| 4 | 0 to 64 |
|
|
| 16 | 1 to 1024 |
|
|
| 1800 | 1 to 86400 |
|
|
max_relay_depthstops agents that relay work to each other without end.max_executions_per_rootlimits the work that one root can start.root_deadline_secondscloses a chain that runs too long.
MCP surface
HTTP MCP publishes the tools enabled by the startup configuration. Use
tools/list and nexus_info for the effective surface and resource revisions.
Memory adds three tools; enabled harness integration adds the eight runtime
tools listed below. Nexus MCP stdio is no longer supported.
Area | Tools |
Metadata |
|
Identity/workspace |
|
Events |
|
Messages/channels |
|
Inbox |
|
Handoffs |
|
Artifacts |
|
Derived view |
|
Health |
|
Catalog |
|
Ephemeral monitor tokens |
|
Optional memory |
|
Optional runtime |
|
message_get, message_list, and message_wait are intentional migration
shims. They return MIGRATED with replacements in the inbox/event surface
instead of failing as unknown tools.
coordination_health stays registered but returns VALIDATION_ERROR while
feature_health is off. memory_put/get/search are absent until
feature_memory is enabled and the server restarts.
Session credentials are trust-sensitive on:
message_create;handoff_claim/complete/verify/reject/cancel;inbox_pull/ack/extend;memory_putwhen published.
In trust_mode=open they are optional but validated if supplied. In
trust_mode=strict they are required. poll_token_issue/renew/revoke always
require a valid session_id and session_secret in both modes.
Versioned MCP resources
URI | Version |
| 4 |
| 3 |
| 5 |
| 5 |
| 3 |
| 2 |
| 2 |
| 4 |
| 5 |
| 4 |
| 2 |
| 2 |
nexus_info reports package/schema/surface versions, the URI-to-version map,
and effective feature flags. Use that live metadata instead of assuming a
cached surface.
Response envelope
Successful tools return:
{
"ok": true,
"data": {}
}Failures return:
{
"ok": false,
"error": {
"code": "VALIDATION_ERROR",
"message": "Human-readable explanation"
}
}Transient SQLite lock/busy failures use DB_ERROR with
details.retryable=true.
Resident token footprint
For the 0.1.10 default surface:
Component | Characters |
Server instructions | 3,992 |
Tool docstrings | 6,563 |
Parameter schemas/descriptions | 15,720 |
Cuttable resident surface | 26,275 (~6,568 tokens) |
Total measured surface | 37,345 (~9,336 tokens) |
Deep explanations live in resources so clients load them only when needed. That historical measurement reported a cuttable reduction of about 44.3% against its frozen baseline; it is not a measurement of the R4 surface.
Configuration
One-shot runtime sessions and MCP presets (0.3.0)
Select one_shot in the runtime session policy to create one single-use session
per call. Global capacity defaults and independent agent overrides configure
parallel executions, unused warm instances, queue size, overflow rejection,
queue timeout and execution timeout. Set the parallel limit to 0 for no
pool-specific ceiling; the host resource budget still applies. Warm instances
must be fewer than a finite parallel limit: for example, 10 parallel and 5 warm. Starting and closing
instances still occupy capacity; a successful response remains successful if
resource cleanup needs another attempt. Queued calls run before replenishment.
Each instance is claimed independently. Pool openings have a separate bounded dispatch budget, so slow replenishment cannot hold up ready sessions. Capacity is released only after confirmed disposal, never merely because a call ended or timed out. The connection wizard saves these settings on Next.
Warm-up reserves the full deficit in one maintenance pass and starts independent opens in parallel, within host and transport credits. Calls arriving during warm-up reserve capacity without binding to a cold process: the oldest compatible caller receives the next ready instance. A slower opening cannot pin that caller while another compatible instance sits idle. No productive turn is sent until the resource is ready; cancellation and waiting deadlines remain durable.
The pool and queue use hexagonal boundaries: domain/resource_pool.py contains
the rules, application/resource_pool.py coordinates them through the ports in
application/resource_pool_ports.py, and adapters/outbound/sqlite/resource_pool.py
provides the internal durable implementation. bootstrap/resource_pool.py
wires these adapters. Local and remote runtimes use the same allocation service;
process operations continue through the authorized execution outbox.
A future distributed adapter must atomically couple dequeue, capacity claims
and outcomes with the execution outbox, preserving idempotency and fenced
disposal. Replacing storage does not change the pool rules; a broker alone is
not a substitute for this transactional contract.
The agent settings show occupied capacity, queued calls, recent outcomes and cancellation controls. Capacity refusals return directly to the caller; broadcast responses identify refused recipients separately. Later failures create a correlated private inbox message for the original caller, including execution stage, possible effects and retry safety. These notices are durable coordination messages; they do not automatically start a new agent turn.
Runtime harness settings also accept an MCP preset with stdio or http
servers. Presets compose with the existing inherit global MCPs setting:
an entry replaces the same inherited name, a disabled entry removes it, and
Nexus integrations remain reserved. Use approved execution-host env_refs
or header_refs for credentials, rather than literal secrets. Paths resolve
on that host, including when using a remote Connector. Codex and Claude use
their native MCP integration; Pi receives tools through a session-owned MCP
bridge. A preset change applies to new sessions and retires incompatible unused
warm instances. Resetting history with agent preservation keeps these settings.
Implementation evidence is tracked in plans/one_shot_mcp/ledger.json.
For serve settings managed by the runtime catalog, effective precedence is:
CLI flag > environment variable > stored dashboard override > defaultWithin the shared bootstrap there is no stored layer, so precedence is
CLI > env > default. Unknown flags, missing values, invalid enums, and
out-of-range numbers fail closed with CONFIG_ERROR. Boolean CLI flags take
an explicit value such as --feature-trace true.
Core runtime
Environment | CLI | Default | Notes |
|
|
| Runtime data directory |
|
|
| SQLite database |
|
| 5000 | Minimum 0 |
|
| 200 | Waiter interval; minimum 1 |
|
| 1000 | Longest sleep of the serve runtime loops when no work is pending; 100 to 60000 |
|
| 30 | Server wait ceiling; minimum 0 |
|
| 300 | Minimum 1 |
|
| 65536 | Inline artifact/content limit |
|
| 300 | Minimum 1 |
|
| 60 | Derived stale threshold |
|
| 1800 | Broadcast/tag presence window |
|
| 86400 | Opportunistic stale close |
|
| 1000 | Render ceiling |
|
| 1000 | Event page ceiling |
|
| 3600 | Minimum 60 |
|
|
|
|
|
|
|
|
|
|
| Sender inbox receipts |
|
|
|
|
|
|
| Operator REST/dashboard path disclosure |
|
|
| One bounded best-effort startup pass |
Retention
Environment | CLI | Default | Minimum |
|
| 30 | 0 |
|
| 14 | 0 |
|
| 7 | 0 |
|
| 30 | 7 |
Pruning removes aged events, read deliveries, closed sessions, and messages
older than the message window. Message retention is pure-age: it can remove
unread, in-flight, or parked messages and cascades to their deliveries and
embeddings. Handoffs and other non-message live rows are not age-pruned.
Metrics
Environment | CLI | Default | Notes |
|
|
|
|
|
|
| Local telemetry JSONL/state |
|
|
| Used only in beacon mode |
|
|
|
|
Settings → Transport security controls the same global policy for the UI,
HTTP APIs, MCP, and runtime WebSocket links. HTTP and WS are allowed over LAN
by default; credentials and binding approvals are still required. https_only
rejects plaintext requests, including loopback, with TLS_REQUIRED (HTTP 403).
Configure TLS at the ASGI server before enabling this mode. This setting does
not install certificates or enable TLS on the listener. It takes effect for new
requests and connections; restart to close existing streams. Only the trusted
ASGI scheme is used: raw client forwarding headers cannot bypass this policy.
| OKTO_NEXUS_METRICS_RETENTION_DAYS | --metrics-retention-days | 30 | Minimum 0 |
| OKTO_NEXUS_METRICS_PUBLISH_INTERVAL_SECONDS | --metrics-publish-interval-seconds | 3600 | Minimum 60 |
Metrics are opt-in. Local mode stores bounded per-event metadata; beacon mode
publishes aggregate hourly counts only. Message bodies, prompts, workspace file
paths, coordination IDs, keys, tokens, URLs, and stack traces are excluded.
Local telemetry JSONL is not currently pruned automatically; operators must
manage those files even though metrics_retention_days is validated and
exposed in configuration.
Feature flags
Environment | CLI | Default |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
feature_memory changes tool registration and requires restart/reconnect.
The other flags gate live behavior.
Serve-only and transport-specific settings
Environment | CLI | Default | Scope |
|
| 8202 |
|
|
|
|
|
|
|
|
|
— |
|
| Initial dashboard workspace |
| — | unset | Suppress serve banner |
Operations
CLI commands
Command | Purpose |
| Start MCP HTTP, REST, SSE, and dashboard |
| Show CLI help |
| Operator NDJSON follower over the event service |
| Enforce retention, optionally vacuum |
| Add keys to legacy keyless identities |
| Export a workspace replay stream as NDJSON |
| Migrate one selected Nexus stdio config entry to HTTP |
Use --help on every command for the full argument grammar.
Tail
okto-nexus tail \
--project-root /absolute/path/to/project \
--agent-id observer \
--stream workspace \
--from latesttail applies per-agent event visibility. An optional --cursor-file belongs
to that consumer only; corrupted checkpoints fail closed.
Retention
Start with a dry run:
okto-nexus admin prune \
--project-root /absolute/path/to/project \
--dry-runThen execute:
okto-nexus admin prune \
--project-root /absolute/path/to/project \
--messages-keep-days 30 \
--vacuumRetention spans the whole shared store even though --project-root is
validated as the command anchor. --vacuum is the only option that compacts
freed pages on disk. There is no always-running coordination reaper;
auto_prune_on_start is a bounded opportunistic pass.
Issue legacy keys
okto-nexus admin issue-keys \
--project-root /absolute/path/to/projectThis is additive and idempotent. Existing keys are never rotated. Newly issued plaintext keys are printed once.
Replay export
okto-nexus admin export \
--project-root /absolute/path/to/project \
--trace-id trc_... \
--output nexus-events.ndjsonThe first line is a manifest; subsequent lines are raw events ordered by
event_id. CLI export is operator-shell access and remains available even
when the REST replay flag is off.
Selected MCP config migration
Set OKTO_NEXUS_MCP_KEY in the operator environment to the selected agent's
existing nxs_ key. The command never issues or rotates a key. Plan a change
to one explicitly chosen entry:
okto-nexus admin migrate-mcp-entry --config /path/to/.mcp.json \
--entry okto-nexus --url http://127.0.0.1:8202/mcpReview the redacted diff summary and its full expected_sha256, then apply
with --apply --expected-sha256 <reviewed-hash>. The command backs up the
original file and refuses concurrent edits. Other mcpServers entries retain
their values. It rejects entries with environment values or unknown fields
that need a manual migration.
Ephemeral monitor token
An authenticated agent can call poll_token_issue, give only the returned
nxsept_... token and base URL to a read-only helper, renew it before expiry,
and revoke it on teardown. The raw token is returned only on issue/renew.
Data model and migrations
The core coordination tables include:
Area | Tables |
Core coordination |
|
Settings/security/catalogs |
|
Search and memory |
|
Governance/workflows |
|
Not every table is workspace-scoped: agents and catalogs are global, inbox deliveries are keyed by recipient identity, and bindings/control-plane records have their own ownership rules.
Migrations are embedded in the package and applied in order. The following list describes the early coordination migrations; it is not the complete R4 schema:
001–008: core schema, close metadata, handoff payload/result, presence, durable inbox deliveries, leases, and session secrets;
009–015: API keys, settings, permissions, embeddings, tags/scopes, capability catalog, and trace IDs;
016–021: governance, approvals, verification, dependencies, memory, and health/event indexes;
022–026: versioned attachable policies, communication presets, display colors, groups/guardrails, and ephemeral poll tokens.
Automatic runtime recovery is enabled by default under Settings → Global runtime defaults.
Recovery scheduling and delivery defaults come from Core's RuntimeAutomation
API, also used by Connector CLI. Nexus retains its transactional inbox and
authorization checks; the embedded host does not maintain a separate retry policy.
Connector clients can test reachability without credentials using GET /v1/reach
or okto-nexus-connector reach --server <base-url>. The public, read-only endpoint
returns the installed server_version, installed server_core_version (or
null if Core is absent), and minimum_cli_version. The minimum CLI version is
maintained in application/reach.py with the supported management contract.
On restart, Nexus reconciles retained local sessions against Core ownership,
resource-release and event-history proofs before admitting new work. A dispatch
refusal that provably never called Core does not require a nonexistent session
journal. Previously submitted messages are never automatically replayed.
New messages arriving during recovery stay unclaimed in a durable queue. The Meta-harness shows Waiting for runtime recovery; MCP can still consume them. Before admission, Nexus revalidates the original sender credentials and current execution authorization. Recovery uses five retries with bounded backoff. An unresolved message moves to Runtime recovery needs attention after bounded queue checks. Inspect Execution log for recovery diagnostics. Disabling the global option stops automatic retries and deferred delivery; it does not grant permission to reuse uncertain sessions. Changes to the Core installation or its inventory may still require reviewing the connection configuration.
Nexus refuses to run against an unsupported newer schema. Runtime SQLite databases and their WAL/SHM/journal sidecars are ignored and must not be committed.
Errors
The domain/MCP contract has a closed catalog of 29 canonical codes:
Area | Codes |
Workspace |
|
Validation/identity |
|
Catalog/control plane |
|
Governance |
|
Dependencies/state |
|
Handoffs |
|
Content/path |
|
Infrastructure |
|
Compatibility/internal |
|
REST adapters also use transport-specific codes such as AUTH_FAILED,
CROSS_ORIGIN_BLOCKED, INVALID_PARAM, INVALID_WINDOW,
INVALID_SETTING, EMBEDDINGS_UNAVAILABLE, and INTERNAL.
Development
uv sync --extra dev
uv run pytest -qFor the complete HTTP/embedding environment:
uv sync --extra serve --extra dev
uv run pytest -qFrontend:
cd frontend
npm ci
npm run buildThe build writes packaged static assets under
src/okto_nexus/adapters/inbound/http/static/.
Release checks:
uv lock --check
uv build --out-dir dist/release-0.3.0
uvx twine check \
dist/release-0.3.0/okto_nexus-0.3.0-py3-none-any.whl \
dist/release-0.3.0/okto_nexus-0.3.0.tar.gzPublish only explicitly named current-version artifacts. The top-level
dist/ may contain older builds.
Project layout
src/okto_nexus/
adapters/
inbound/
cli/ serve, tail, admin
http/ FastAPI, REST, SSE, packaged SPA
mcp/ HTTP server, resources, projections, configured tools
outbound/
sqlite/ repositories and migrations adapter
embedding/ optional semantic provider
file/ sharedmd/ artifact and derived-view I/O
telemetry/ tokenizer/ metrics support
application/ use cases and ports
domain/ entities, routing, state machines, policies
migrations/ 001 through 027
testing/ reusable test/replay harnesses
frontend/ React dashboard source
tests/ unit, contract, integration, replay tests
docs/design/ architecture and design records
pyproject.toml package metadata and extras
uv.lock reproducible dependency lockContributing
See CONTRIBUTING.md for the supported Python and frontend development environments, validation commands, and pull-request workflow.
Troubleshooting
event_wait returns immediately
Pass timeout_seconds > 0. Omitted, null, and zero are snapshots by design.
An agent misses broadcasts
Check that it opened a session in the correct resolved workspace and continues to heartbeat. Read-only event/inbox checks do not advance presence.
A direct peer is missing from discovery
Authenticated discovery is filtered by communication reachability. Inspect the caller's outbound and the peer's inbound communication scopes/tags in the dashboard.
Capability or tag targeting fails
Create the capability/tag value in Registry first. Catalog validation is fail-closed.
Memory tools are absent
Set feature_memory=true and restart/reconnect. Unlike live behavior flags,
this flag changes MCP tool registration.
Semantic search is unavailable
Use embedding_mode=stub or install the embeddings/serve extra and use
embedding_mode=local. off intentionally disables search.
DB_ERROR reports a lock
If details.retryable=true, retry the same call after the competing writer
commits. Avoid opening the runtime SQLite file with tools that hold long write
transactions.
The hub says another server already owns the home
serve holds {home}/nexus.serve.lock and refuses a second server using that
same home. Stop the other hub or choose a different --home; changing only
--db-path does not change the lock scope.
Workspace path is rejected
Pass an existing absolute path. Nexus canonicalizes the real path before deriving the workspace ID and validating artifact containment.
A monitor cannot mutate state
That is expected for nxsept_ tokens. They are intentionally read-only and
accepted only on the four monitor endpoints.
Cached docs appear stale
Call nexus_info and compare surface_revision and resource_versions before
reusing cached MCP reference content.
Security and limitations
Report vulnerabilities privately according to SECURITY.md. Do not include vulnerability details, credentials, tokens, or private workspace data in a public issue.
Nexus is designed for local or controlled single-tenant coordination, not as a public multi-tenant broker.
It does not terminate TLS. Put an authenticated TLS reverse proxy in front of a remote bind.
The dashboard shell and public health/info/license assets remain public; data/control REST requires authentication outside loopback.
API keys are hash-only at rest and shown once. Regeneration invalidates the old key immediately.
Session secrets are stored in plaintext in the local SQLite database; anyone who can read that file is inside the session trust boundary.
Channels are labels, not security boundaries.
SQLite is the only built-in coordination store. There is no Redis, PostgreSQL, or cloud broker adapter.
There is no always-running coordination scheduler/reaper. Expiry is checked opportunistically and retention runs manually or at startup when enabled.
The HTTP server does use worker threads/tasks for operational needs such as blocking waits and telemetry; “no scheduler” does not mean “no threads.”
Message durability is bounded by message retention and explicit reset.
Memory is experimental and changes the MCP surface at startup.
Artifact file references are confined to the canonical workspace.
Avoid committing runtime databases, sidecars, metrics output, and local secrets to version control.
Future direction
Likely extension points are additional durable-store adapters, a push-backed waiter that removes internal sleep polling, stronger remote deployment packaging, and further generated documentation from the live MCP schemas. The delivered surface already includes permissions, catalogs, communication scopes, policies, guardrails, HITL, trace correlation, verified/DAG handoffs, memory, health, replay, ephemeral poll tokens, embeddings, metrics, REST, SSE, and the dashboard.
Release notes
0.1.10 — previous release
Validation-hardening release: lands the external PR backlog and closes the open validation bugs (#26–#30). The MCP contract remains at surface revision 33 and the latest database migration remains 028.
Bumped the package and distribution metadata from 0.1.9 to 0.1.10.
Capped message artifact references at 20 (
MAX_ARTIFACTS) with{count, max}diagnostics on both MCP and REST; the dashboard Meta-harness route now defers to the domain cap (the legacy 10-attachment fence was removed) andVALIDATION_ERRORdetails survive the REST envelope.Rejected exact-duplicate artifact references; the error names the offending index and the duplicated reference.
Capped tag selector value lists at 20 values per key with set-based de-duplication, removing an O(n²) scan reachable from
message_create/handoff_createtag targets.Added length caps to the routing target grammar identifier fields (
agent_id/role/capability, 256 chars) and to eachdepends_onid (128 chars).Resynced the dashboard ColorPicker draft when the
valueprop changes while the component stays mounted.Consolidated the duplicated bounded-list count check into the shared domain helper
check_list_size.
0.1.9
Dashboard presentation release. The MCP contract remains at surface revision 33 and the latest database migration remains 028.
Bumped the package and distribution metadata from 0.1.8 to 0.1.9.
Added list/grid switching to the Artifacts catalog, including an Explorer-style grid with type-aware file icons, names, and payload sizes.
Collapsed Meta-harness read-receipt notifications into acknowledgement flags on their original messages by default. Grey flags identify incomplete target sets; green flags mean every target acknowledged, and their detail modal identifies pending recipients and delivery/read timestamps.
Added the
meta_harness_receipt_displayinterface setting so operators can restore separate receipt messages with thetimelinemode.
0.1.8
Artifact storage and catalog release. The MCP contract is at surface revision 33 and the latest database migration is 028.
Moved artifact payloads and free-form metadata out of SQLite into an adapter-backed store, with a local filesystem adapter by default.
Imported path-based artifacts into managed storage so they remain available independently of the original workspace file.
Added the Artifacts catalog with paginated producer/date filtering, detail previews, expanded Raw/Rich rendering, and managed-payload downloads.
Added rendered metadata fields plus safe Rich previews for Markdown, JSON, and HTML artifacts.
0.1.7
Dashboard usability and operator-interaction release. The MCP surface remains unchanged; the latest database migration is 027.
Added the Meta-harness chat with independent Message/Handoff and Private/Broadcast controls.
Combined outgoing turns, incoming agent messages, and handoff completion or rejection outcomes in a live, agent-filterable timeline.
Rendered structured chat content as readable labels and lists instead of raw JSON.
Reworked Guardrails group composition and rule authoring, including agent auto-completion, capability-scoped assignments, regex assistance, and stricter validation.
Added completed-agent responses and rejection reasons to Handoff cards and details.
Removed nested page scrolling from the dashboard shell.
0.1.6
Startup compatibility maintenance release. The MCP contract and database schema are unchanged: surface revision remains 32 and the latest migration remains 026.
Rebuilt the MCP v1 FastMCP settings model after import, eliminating the unresolved
lifespanforward-reference warning introduced bypydantic-settings2.15.Applied the compatibility path to both stdio and streamable-HTTP server construction.
Constrained the MCP SDK dependency to
mcp>=1.0,<2so adopting the breaking v2 API requires an explicit migration.Added a regression test that promotes the startup warning to an error and verifies the settings model is complete.
Verified the installed executable with local MiniLM warm-up, a live health request, and the complete test suite.
0.1.5
Live MCP smoke-test maintenance release. The MCP contract and database schema are unchanged: surface revision remains 32 and the latest migration remains 026.
Restored the real stdio smoke test on clean stores after capability registration became fail-closed.
Documented when and how to run the isolated two-agent smoke flow on Unix and Windows, including its
LIVE E2E RESULT: PASScompletion signal.Added semantic cards for structured
kindmessages in the Graph conversation drawer, with distinct read-receipt and handoff lifecycle treatments instead of raw JSON.Synchronized
pyproject.toml,uv.lock, and release commands at version 0.1.5.
0.1.4
Dashboard, observability, and coordination-guidance release. The MCP guidance contract advances to surface revision 32; the database schema is unchanged and the latest migration remains 026.
Added detailed and compact activity-based agent graph modes, profile colours, live status badges, richer relationship context, and graph conversations.
Added an event timeline, expanded event and handoff filters, message detail hydration, workspace display names, and catalog import/export workflows.
Improved dashboard views for agents, approvals, communication, events, guardrails, handoffs, messages, policies, and workspaces.
Extended observability APIs and repositories with message lookup, filtered handoff/event queries, and bucketed event timeline data.
Completed PyPI project URLs, keywords, and supported-Python classifiers, with a focused metadata regression test.
Added contributor and security policies, structured GitHub issue forms, and a dedicated documentation-assets location for product screenshots.
Clarified that agents follow the role and communication profile returned by
agent_whoamiunless the user supplies a task-scoped override, without bypassing platform governance.Defined handoffs as the required, traceable mechanism for executable work, broadcasts as shared alignment or information fan-out, and direct messages as status, clarification, and informal coordination channels.
Rewrote this README against the current CLI, transport/authentication model, dashboard, 43/46-tool surface, seven routing strategies, message retention, governance features, verification/DAG workflows, 34-table schema, 29-code error catalog, operations, testing, and limitations.
Synchronized
pyproject.tomland the root project entry inuv.lockat version 0.1.4, and aligned the package/dashboard license label with the addendum that is actually included inLICENSE.Removed tracked runtime SQLite databases and added ignore coverage for database files and journal/WAL/SHM sidecars.
Verified release archives contain the dashboard and migrations without runtime SQLite databases or sidecars.
0.1.2
Documentation-accuracy sweep at surface revision 31.
Corrected pre-flight, monitoring, heartbeat, inbox, trace, policy, HITL, artifact-audience, and target-grammar documentation.
Kept 12 deep reference resources versioned and reduced the resident cuttable surface to about 26.3k characters.
Corrected the token-reduction gate to include experimental-surface growth.
0.1.1
Hardened authenticated self-only identity/session rules and permission checks.
Added permission-gated workspace paths, shared view, health, and memory.
Added configurable aggregate metrics telemetry.
0.1.0
Added ephemeral remote-monitor tokens and read-only monitor endpoints.
Added attachable policies, communication presets, guardrail/group administration, enriched agent graph cards, and loopback trust hardening.
Made memory a registration-time experimental MCP surface.
0.0.x
Built the MCP reference-resource system, HTTP hub, live dashboard, inbox receipts, semantic search, monitoring guidance, and dashboard observability waves.
Automatic inventory revalidation
Authenticated host inventory publications automatically revalidate configured runtime connections. Core version changes and changes to unrelated installations do not require repeating agent setup when the selected installation evidence and runtime contract remain identical. Original bindings, permissions, grants, and running-session identities are preserved. Fresh inventory is still required.
The Execution log shows INVENTORY_REVALIDATED or INVENTORY_REVIEW_REQUIRED per
connection. Changed or missing installations, changed execution capabilities, or
missing historical evidence require review. Legacy embedded connections can use
their retained, digest-verified local installation record as the baseline.
Version observations survive a Core upgrade only for the exact same host source
and platform; current Core still evaluates runtime support and containment.
Remote machine approval and connection status
Remote approvals are scoped to the agent and its registered Connector machine
identity (connector_id / executor_id), not its hostname or IP address.
Settings → Remote machine policy defaults to Require approval when
another machine requests the same agent. The operator can select Deny another
machine or Accept and revoke previous machine instead. The first machine
always requires approval. A reviewed replacement revokes the previous machine's
endpoints, grants, binding tickets and session capabilities atomically.
Automatic replacement retains the approved harness preferences and the original grant expiry and consumed action count. It cannot widen tool access or renew an exhausted authorization. Ambiguous prior connections require manual review.
Connector onboarding is asynchronous: after submitting the request, the CLI
returns and its daemon completes the connection after approval, including after
a daemon restart. No operator key is needed on the remote computer. Use
okto-nexus-connector status (or --json status) for each configured connection's
current state. The authenticated GET /v1/connections/status endpoint reports
only connections belonging to the calling agent, without credentials or paths.
License
Copyright 2026 Okto Labs.
Okto Nexus is distributed under the Elastic License 2.0 together with the project's SaaS, competing-service, internal-use, and branding addendum. It permits internal and qualifying single-tenant use and prohibits the specified multi-tenant, white-label/OEM, competing, and large-scale internal-platform uses; applicable notices and attribution remain required.
Read the complete
LICENSE before use
or redistribution. It is also included in the source distribution and served
by the running hub at GET /api/v1/license.
Global harness MCPs
See global harness MCP inheritance for global defaults, agent/harness overrides and local/remote host behavior.
Available Tools
43 toolsagent_getA
Return one agent's details incl. last_seen_at. Scoped by reachability: an agent outside your comm scope reads as NOT_FOUND, indistinguishable from a non-existent agent_id.
Inputs: agent_id (required). Returns: object
agent_listA
List registered agents (global), each with role/capabilities and last_seen_at. Authenticated callers see only agents their comm scope can reach (plus themselves); anonymous callers see all.
Inputs: no named parameters. Returns: object
agent_registerA
Update YOUR OWN profile (role/capabilities/metadata); SELF-ONLY (else PERMISSION_DENIED). Capabilities are fail-closed against the central catalog. Docs: okto-nexus://reference/tool-docs/identity.
Inputs: agent_id (required), role, metadata, capabilities. Returns: object
agent_whoamiA
Return YOUR OWN profile: agent_id, role, capabilities, metadata, permissions, effective_policies + governance, plus communication style when set. Docs: okto-nexus://reference/tool-docs/identity.
Inputs: no named parameters. Returns: object
artifact_getA
Retrieve an artifact by id within the workspace resolved from project_root. Reads are audience-scoped: a caller outside the frozen audience gets NOT_FOUND (indistinguishable from a missing id).
Inputs: artifact_id (required), project_root (required). Returns: object
artifact_putC
Register a file/text/json/markdown/html artifact in the resolved workspace. Docs: okto-nexus://reference/tool-docs/artifacts.
Inputs: project_root (required), artifact_type (required), name, path, +2 more. Returns: object
capability_listB
List the capability catalog merged with owners: every registered name (with description; agent_count 0 if unowned), agents scoped to your comm reach. Normalised as capability routing matches.
Inputs: no named parameters. Returns: object
channel_createA
Create a channel by name (idempotent; created=false if it already existed). Channels are organizational labels, not ACLs.
Inputs: name (required), project_root (required). Returns: object
channel_listB
Return the workspace channels (general is seeded by default).
Inputs: project_root (required). Returns: object
coordination_healthA
Windowed coordination-health report for the workspace: aggregated ok|warn status, 7 metric blocks and thresholds. Requires feature_health and health.read.
Inputs: project_root (required), window, agent_id. Returns: object
event_cursorA
Return the stream's CURRENT END as a cursor (O(1), no scan) - the pre-flight monitor anchor so you see only events appended after now. Returns {cursor} (0 for empty).
Inputs: stream (required), agent_id (required), project_root (required). Returns: object
event_getA
Read a cursor-paginated page of the event log (non-blocking). Actors outside your comm scope are omitted; yours and system events always show. Docs: okto-nexus://reference/tool-docs/events.
Inputs: stream (required), agent_id (required), project_root (required), limit, +3 more. Returns: object
event_waitA
Read the event log; optionally long-poll (0/omitted/null = snapshot; >0 blocks). Scoped like event_get (actors outside your comm scope omitted). Patterns: okto-nexus://reference/monitoring.
Inputs: stream (required), agent_id (required), project_root (required), limit, +4 more. Returns: object
handoff_cancelA
Creator-only OPEN -> CANCELLED; retract a handoff nobody should take (e.g. a pool target matching zero agents). Only OPEN handoffs cancel. In strict mode pass session creds.
Inputs: agent_id (required), handoff_id (required), project_root (required), reason, +2 more. Returns: object
handoff_claimA
Atomically claim an OPEN handoff; single winner, others get a structured error. Returns the payload + claimed_by/lease_expires_at. In strict mode pass session_id + session_secret.
Inputs: agent_id (required), handoff_id (required), project_root (required), session_id, +1 more. Returns: object
handoff_completeA
Owner-only delivery of a CLAIMED handoff: -> COMPLETED, or -> VERIFYING when acceptance_criteria were set (verifier decides via handoff_verify). In strict mode pass session creds.
Inputs: agent_id (required), handoff_id (required), project_root (required), result, +2 more. Returns: object
handoff_createA
Create an OPEN handoff (validates target/visibility); emit handoff.created. After creating, poll handoff_get for status/result. Full docs: okto-nexus://reference/tool-docs/handoff.
Inputs: target (required), visibility (required), project_root (required), from_agent_id (required), +6 more. Returns: object
handoff_getA
Read a handoff by id: status, claimant, payload, result/rejected_reason + verification/dependency fields if set. The creator's path to the outcome. Full docs: okto-nexus://reference/tool-docs/handoff.
Inputs: agent_id (required), handoff_id (required), project_root (required). Returns: object
handoff_list_availableA
Expire leases, then list OPEN handoffs visible+eligible to the caller (paginated); entries are metadata-only until claimed.
Inputs: agent_id (required), project_root (required), limit, cursor, +1 more. Returns: object
handoff_rejectA
Reject a handoff (owner CLAIMED->REJECTED or direct-target OPEN->REJECTED). In trust_mode=strict pass session_id + session_secret.
Inputs: agent_id (required), handoff_id (required), project_root (required), reason, +2 more. Returns: object
handoff_verifyB
Verifier-only verdict on a VERIFYING handoff: 'pass' -> COMPLETED (verified_by), 'fail' -> CLAIMED for rework (feedback + renewed lease). In strict mode pass session creds.
Inputs: verdict (required), agent_id (required), handoff_id (required), project_root (required), +3 more. Returns: object
inbox_ackA
Acknowledge messages into history (read). Returns {acknowledged, read_message_ids}. Emits a message.read receipt to each sender.
Inputs: agent_id (required), message_ids (required), session_id, session_secret. Returns: object
inbox_countA
Return your inbox lane sizes {unread, in_flight, read}. Cheap READ-ONLY between-turns check (pull when unread > 0). Expired in-flight leases count as unread; parked excluded.
Inputs: agent_id (required). Returns: object
inbox_extendA
Renew the lease on in-flight messages you pulled but have not finished (now + extend_seconds). All-or-nothing: if any id is not in-flight the call fails per-message and nothing is extended.
Inputs: agent_id (required), message_ids (required), extend_seconds (required), session_id, +1 more. Returns: object
inbox_historyA
List your acknowledged (read) messages, newest-first, keyset-paginated (stable pages even while you keep acknowledging). READ-ONLY.
Inputs: agent_id (required), limit, cursor, profile. Returns: object
inbox_peekA
Triage pending messages (unread + in-flight) WITHOUT consuming. READ-ONLY, envelope-only by default (body_preview + body_bytes). include_parked/include_bodies opt in.
Inputs: agent_id (required), limit, profile, include_bodies, +1 more. Returns: object
inbox_pullA
Take your unread messages into in-flight and return them WITH body (index-free; no cursor). At-least-once: unacked pulls are redelivered. Docs: okto-nexus://reference/tool-docs/inbox.
Inputs: agent_id (required), limit, profile, session_id, +2 more. Returns: object
message_createA
Persist a message and fan it out to recipient inboxes; emit message.created. The response confirms delivery (recipients + delivered_count). Full docs: okto-nexus://reference/tool-docs/messages.
Inputs: body (required), subject (required), project_root (required), from_agent_id (required), +7 more. Returns: object
message_getA
MIGRATED (S3): replaced by inbox_pull / inbox_peek / inbox_history. Always returns ok:false code=MIGRATED with the replacement call.
Inputs: no named parameters. Returns: object
message_listA
MIGRATED (S3): replaced by inbox_peek / inbox_history (your messages) and event_get (bus traffic). Always returns ok:false code=MIGRATED.
Inputs: no named parameters. Returns: object
message_statusA
Track a message you SENT: per-recipient delivery states {recipient, status, attempts, read_at} (unread/delivered/read/parked). READ-ONLY.
Inputs: message_id (required). Returns: object
message_waitA
MIGRATED (S3): replaced by inbox_count polling (cheap) or event_wait (explicit blocking). Always returns ok:false code=MIGRATED.
Inputs: no named parameters. Returns: object
nexus_infoA
Report server versions: package_version, schema_version, surface_revision, resource_versions, features (read-only {feature_*: bool}). Call when behaviour disagrees with cached schemas.
Inputs: no named parameters. Returns: object
poll_token_issueA
Issue an ephemeral read-only monitor bearer (nxsept_...) for this authenticated agent's session workspace. Store only in the background monitor; never persist the raw token.
Inputs: session_id (required), session_secret (required). Returns: object
poll_token_renewA
Rotate and extend the active ephemeral poll token for this session. The previous raw nxsept_ bearer stops working immediately.
Inputs: session_id (required), session_secret (required). Returns: object
poll_token_revokeB
Revoke the active ephemeral poll token for this session.
Inputs: session_id (required), session_secret (required). Returns: object
session_closeB
Close a session (idempotent); repeating returns ok and stays closed.
Inputs: session_id (required), workspace_id. Returns: object
session_heartbeatA
Advance a session heartbeat and report the derived status; keeps you PRESENT (in the broadcast audience) and clear of the stale-session reaper.
Inputs: session_id (required), workspace_id. Returns: object
session_openA
Open a session bound to (agent_id, workspace_id); returns a per-session session_secret (ONLY here - keep it; required by sensitive verbs in strict mode). Heartbeat to receive broadcasts.
Inputs: agent_id (required), metadata, workspace_id. Returns: object
tag_listA
List the global operator-managed tag catalog (keys + values) that agent tags, comm_scope and tag targets are validated against fail-closed.
Inputs: no named parameters. Returns: object
workspace_listA
GLOBAL-ADMIN: enumerate ALL workspaces. Paths OMITTED by default (include_paths=true is an admin/ops opt-in). For discovery use agent_list / capability_list.
Inputs: agent_id, include_paths. Returns: object
workspace_resolveB
Resolve a project_root to its deterministic workspace_id and upsert it.
Inputs: project_root (required), display_name. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 43 tools
The tool set covers many distinct sub-domains (handoffs, inbox, events, sessions, etc.), and most tools have clearly differentiated purposes. However, some read-oriented tools overlap (event_get/event_wait, inbox_pull/inbox_peek) and three legacy message_* tools are deprecated but still present, which could momentarily confuse an agent.
All 43 tools use consistent snake_case with a domain_action pattern (e.g., handoff_create, inbox_pull, poll_token_issue). The only variation is minor compound forms like handoff_list_available, but the convention is predictable and uniform.
43 tools is well above the typical 3–15 range and even excluding the 3 deprecated message tools leaves 40. This heavy surface likely burdens agents and exceeds what the server’s purpose requires.
The surface covers CRUD-like lifecycle for handoffs, inbox, sessions, events, artifacts, agents, workspaces, and poll tokens, with no dead ends in core workflows. Minor gaps exist (no artifact list/delete, no agent/channel delete), but agents can work around them.
Maintenance
Related MCP Connectors
Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
MCP gateway and cloud hosting for MCP servers, agents and apps like RAG, managed from your agent.
One MCP for any agent: your connected apps, shared memory, web search, social data and media tools.
Related MCP Servers
AlicenseNot gradedqualityDmaintenancePersistent memory and handoff intelligence layer for MCP agents. Most memory servers retrieve text — Memory Nexus compounds operational context, learning from usage and progressively synthesizing observations into higher-order intelligence across sessions and tools.MIT- AlicenseAqualityBmaintenanceNexus Memory gives every MCP-compatible agent one persistent, self-hosted shared memory with hybrid retrieval, drift detection, and anti-poisoning features.193MIT
- AlicenseNot gradedqualityCmaintenanceA real-time inter-agent switchboard, delivered as one centralized streamable-HTTP MCP server. Any MCP-capable agent can message, coordinate, and stay ambiently aware of others.1AGPL 3.0
- AlicenseNot gradedqualityCmaintenanceAn event-driven MCP server that enables agents to share context streams, publish and subscribe to events, manage tasks, and follow protocols, keeping a fleet of agents mutually context-aware in real time.1MIT