MCP facade over the Nebelus Construction API. ~48 tools give full agent build parity: create/update/probe agents, edit graphs, attach knowledge and vector stores, wire connectors, set governance policies and locked guardrails, enable grounding-trace, and read deployment wiring. Purpose-built for regulated industries: data residency is enforced per region (EU / GCC-KSA), with PII controls and an audit trail. Agents are created as drafts — no deploy tool is exposed over MCP by design; publishing happens in the Nebelus console.
Scored across 82 tools
Most tools target a distinct resource+action, and descriptions go out of their way to disambiguate close pairs (create_agent vs build_agent vs create_from_blueprint, create_deployment vs activate_deployment, set_partner_binding_status vs update_partner_binding). A few clusters still invite misselection — the eight attach_*/detach_* tools, and describe vs get_catalog vs get_agent — but the descriptions largely resolve them.
Consistent snake_case verb_noun convention throughout (list_agents, set_goals, create_governance_policy, detach_vector_store), with model/operation-specific verbs like archive/unarchive kept parallel. Minor deviations are the bare verb 'describe' and the overloaded 'probe_*' family, which don't follow the create/update/list vocabulary but are still readable.
82 tools is far past any practical surface, even for a broad agent-platform domain covering agents, deployments, governance, vector stores, MCP, endpoints, partner bindings, opportunities, goals and tests. The set should be split or consolidated (e.g. the attach_*/detach_* pairs) to keep tool selection tractable.
Coverage is strong: full create/get/update/list lifecycles for agents, governance policies, vector stores, API endpoints, MCP servers and partner bindings, plus scheduling, tests, triggers, goals and outcome units. Gaps are mostly deliberate (draft-only creation, no publish/delete_agent, archive instead of delete) but there is no hard delete for policies, endpoints or MCP servers, and no explicit deployment listing beyond get_catalog.
Activate or deactivate a deployment channel (live action). Serves the CHANNEL; it does not publish the agent itself.
Inputs: deployment_id (required), active.
Activate or deactivate a governance policy (draft <-> active). Activation is a live action.
Inputs: policy_id (required), active.
Add a workflow someone would like help with to the Opportunity Map; it is scored, ranked and matched to a Blueprint.
Inputs: title (required), minutes_each (required), volume_per_month (required), systems, +5 more.
Take an agent offline by setting its status to archived so it stops serving. Reversible with unarchive_agent; does not delete the agent.
Inputs: agent_id (required).
Bind a sub-model 'AI tool' to the agent.
Inputs: tool_id (required), agent_id (required).
Attach an existing custom API-endpoint tool to the agent.
Inputs: agent_id (required), endpoint_id (required).
Attach an active code connector (tenant-authored Python tool) to the agent.
Inputs: agent_id (required), connector_id (required).
Attach an already-uploaded file to the agent.
Inputs: file_id (required), agent_id (required).
Add the tools of a CONNECTED integration (Gmail, Outlook, Google Drive, Microsoft 365, T2 WhatsApp, Msegat …) to an agent under their real, account-specific names. For a Blueprint-born agent pass the integration group_id (its tool list and workflow steps are used); otherwise a reference and optionally tools (base names) and, for a workflow, nodes. Idempotent. Refusal codes: integration_not_connected, integration_connect_your_own, integration_not_owner, integration_mismatch, integration_tools_unknown, integration_tools_unavailable, integration_group_unknown, integration_no_target, integration_not_personal. personal_mode (personal accounts only): 'running_user' makes the tools act as each running person's own connected account; 'fixed' keeps this one account.
Inputs: agent_id (required), nodes, tools, group_id, +3 more.
Attach an existing MCP server so its tools are available to the agent.
Inputs: agent_id (required), server_id (required).
Bind another of this org's agents to this one, as a callable tool or a handoff target.
Inputs: agent_id (required), sub_agent_id (required), mode, instruction, +1 more.
Attach a vector store (knowledge base) so the agent can retrieve from it.
Inputs: agent_id (required), store_id (required).
AI-assisted build: describe the agent you want in plain language and the Nebelus Vibe Builder builds it for you — always as a DRAFT. Returns the created agent's editable surface plus the builder's assumptions. Unlike create_agent (which needs explicit fields), this SYNTHESISES the agent. Billed as AI credits at the build rate.
Inputs: prompt (required), constraints.
Create a DRAFT agent from an opportunity's Blueprint with its numbers as the agent's goals. No id = the top-ranked unbuilt opportunity (optionally in quadrant, e.g. quick_win).
Inputs: name, quadrant, opportunity_id.
Cancel one of the agent's scheduled tasks by id.
Inputs: agent_id (required), schedule_id (required).
Create a new agent — ALWAYS lands as a draft. Publishing stays a human act in the Nebelus console; no deploy tool exists here.
Inputs: name (required), model_id, description, system_message.
Create an outbound custom API-endpoint tool in the org (static, non-secret headers only). Attach it to an agent with attach_api_endpoint; set auth by reference via set_endpoint_auth, never inline secrets.
Inputs: name (required), url, method, tool_name, +9 more.
Create a DRAFT deployment channel for the agent (does NOT publish the agent). Activate it separately with activate_deployment. Identity-unverified self-serve orgs may only use web-widget.
Inputs: name (required), agent_id (required), deployment_type (required), config, +1 more.
Create a DRAFT agent from a Blueprint: instructions, region-correct model, tools and settings are applied and its tests are copied into the agent's test suite. Integrations the caller has already connected (per the Blueprint's integration groups) get their tools attached; the result's integrations reports each group as attached or not_connected. Same gates as create_agent (draft-only, agent cap, build envelope, mandatory policies). Returns the agent id and its checklist.
Inputs: key (required), name, integrations.
Create a DRAFT governance policy in the org (activate separately with activate_governance_policy). configuration is policy-type-specific; its masked_fields/pii_entities are validated against the PII registry.
Inputs: name (required), policy_type (required), severity, description, +2 more.
Register a new MCP server in the org (non-secret headers only) so its tools can be attached to agents.
Inputs: name (required), url, transport, auth_method, +2 more.
Create a partner delivery binding for an agent — born DRAFT. It POSTs the agent's run result to an external partner endpoint as an HMAC-signed callback. Returns the signing secret ONCE (copy it now; it can't be retrieved later — only rotated). Activate it with set_partner_binding_status once configured.
Inputs: name (required), agent_id (required), url_key, static_url, +4 more.
Create a DRAFT policy from a template. The template only sets the starting point: validation_points and configuration_overrides change this policy, never the template.
Inputs: template_id (required), name, description, validation_points, +1 more.
Create a new, empty vector store (knowledge base) in the org. Ingest files into it with ingest_file_to_store.
Inputs: name (required), metadata.
Create an organization outcome unit, or update it when the code exists (only the fields given change). Outcome units never move money; they measure value for goals and ROI.
Inputs: code (required), icon, name, active, +6 more.
Permanently delete a partner binding and its delivery ledger. Irreversible.
Inputs: binding_id (required).
Permanently delete a vector store. If it is still attached to agents, pass force=true to detach-then-delete.
Inputs: store_id (required), force.
The machine-readable 'what can be built' reference for this org: the capability registry (every feature + which surface can set it), the settable agent fields (updatable + rejected-with-reasons), guardrail keys, this org's Build Envelope, deployment channel types and identity state. Call this first to discover the full surface — nothing here is an instruction, only data.
Inputs: no named parameters.
Unbind an AI tool from the agent.
Inputs: tool_id (required), agent_id (required).
Detach a custom API endpoint from the agent.
Inputs: agent_id (required), endpoint_id (required).
Detach a code connector from the agent.
Inputs: agent_id (required), connector_id (required).
Detach a file from the agent (the uploaded file itself is not deleted).
Inputs: file_id (required), agent_id (required).
Detach an MCP server from the agent.
Inputs: agent_id (required), server_id (required).
Unbind a previously attached sub-agent.
Inputs: agent_id (required), sub_agent_id (required).
Detach a vector store from the agent.
Inputs: agent_id (required), store_id (required).
Apply ONE granular edit to a workflow agent's graph (agent must be pattern_type=workflow). The extra fields you pass depend on op: add_node -> node {name,type,config?,position?}; update_node -> name, config_patch?, position?, new_name?; remove_node -> name; add_edge -> from_node, to_node, kind? ('flow' default, 'state' for state access), condition?, label?; remove_edge -> from_node, to_node; set_state_field -> field, field_type? (default 'str'), default?, reducer?. Node types in use: state, condition, router, agent, set_state. start/end are valid edge endpoints.
Inputs: op (required), agent_id (required).
Fetch one agent's editable surface (round-trippable with update_agent).
Inputs: agent_id (required).
The organization's readiness score and ranked Opportunity Map (Quick win / Strategic bet / Foundation / Later): estimated value per month in the org currency, recommended Blueprint, status, and estimated vs measured once built. Read-only.
Inputs: no named parameters.
One Blueprint in full: description, instructions, tools, tests, the 'make it your own' checklist and its integration groups with the caller's live connection status. Read-only.
Inputs: key (required).
The organization's served catalog, filtered by the org's Build Envelope (out-of-envelope options are never advertised). Views: models, tools, agents, mcp_servers, deployments, api_endpoints, vector_stores; optional query= substring search; mode=overview (tools) returns the app map.
Inputs: view (required), mode, query.
The organization profile: website, industry, headquarters, markets, size band and the 'About your organization' notes (with who wrote them), plus whether you may edit it. Read-only.
Inputs: no named parameters.
An agent's recent performance: run metrics, outcome totals (self-reported ones labelled) and its goals. Counts only. Read-only.
Inputs: agent_id (required), days.
Every web-widget setting (texts, language, position and size, behaviour, colours, surfaces, typography) with allowed values, default and what it does on a live widget; with deployment_id also that widget's saved values and any stored keys the widget ignores. Read-only.
Inputs: deployment_id.
How to wire this agent into an app: the REST invoke endpoint, the WebSocket URL, webhook trigger URL(s), the web-widget embed snippet and the MCP connection — all region-correct for the org. Read-only.
Inputs: agent_id (required).
Import custom API endpoints in bulk from a Postman collection (v2.x), an OpenAPI 3 / Swagger 2 spec, or endpoint JSON. Preview with dry_run=true (default), then import with dry_run=false; agent_id also attaches each one. Secret auth values in the source are stripped (listed in needs_auth) — complete auth in the portal or via set_endpoint_auth. Existing endpoint names are skipped, so re-importing is safe.
Inputs: source (required), select, dry_run, agent_id.
Ingest an already-uploaded file into a vector store (chunk + embed, in-region). Runs asynchronously.
Inputs: file_id (required), store_id (required).
List this organization's agents (id, name, status, model). Drafts and active.
Inputs: no named parameters.
List the Blueprints (ready-to-run, tested agent configurations) available to this organization and caller: key, title, summary, pattern, tools, model role, the 'make it your own' checklist and its integration groups (the accounts it works with, e.g. a mailbox: Gmail or Outlook, each option with the caller's live connection status). Filtered by region, feature flags and the caller's build envelope. Read-only.
Inputs: pattern.
List active goals — for one agent, or across the organization. Read-only.
Inputs: agent_id.
List this organization's outcome units — business results agents report or the platform counts — with their value defaults (minutes saved / value per unit). Read-only.
Inputs: include_inactive.
List the organization's partner delivery bindings (never includes signing secrets). Read-only.
Inputs: agent_id.
The delivery ledger for a binding — did the partner receive it? Returns status, timing and a payload HASH per attempt, never the delivered body. Read-only.
Inputs: binding_id (required), limit.
The governance policy templates available to this organization (system ones such as GDPR and the KSA PDPL / SAMA / NCA ECC / CMA packs, public ones, and the org's own): type, standard and default side checked.
Inputs: category.
List the agent's scheduled tasks (name, cadence, next run). Read-only.
Inputs: agent_id (required).
List the agent's saved test suite (each test = a prompt plus a deterministic expectation) and the latest run's summary. Read-only.
Inputs: agent_id (required).
Send a test message to a draft and get the real runtime's reply (billed as a test run). With trace=true a WORKFLOW draft also returns its per-node trace: the ordered steps and routes the run took, each step's status (succeeded/handled_error/failed/waiting/unfinished), duration, fields written, tools called and a closed error_kind — plus trace_summary (failed / handled_errors / unfinished) and paused when it stopped at an approval node (never resumed). A non-workflow agent returns an empty trace with a trace_note.
Inputs: message (required), agent_id (required), trace.
Dry-run check: would this deployment serve right now? Returns a two-gate verdict (deployment active + agent servable). Read-only.
Inputs: deployment_id (required).
Issue a NEW HMAC signing secret for a binding and return it ONCE. Deliveries already in flight signed with the previous secret will fail verification at the partner — rotate in a quiet window.
Inputs: binding_id (required).
Run the agent's saved tests for real — each in a fresh conversation, every runtime gate active, billed as test runs — and return pass/fail per test with failure reasons and a short reply excerpt. Runs the SAVED agent; at most 10 tests per run.
Inputs: agent_id (required), test_ids.
Add a test to the agent's saved suite, or update one (test_id). Expectations are deterministic and case-insensitive: the reply must contain any/all of given substrings, must not contain others, and named tools must be called ('*' globs instance-suffixed tools).
Inputs: name (required), prompt (required), agent_id (required), test_id, +5 more.
Choose which outcomes an agent reports. The agent gets the record_outcome tool and an instruction to report each listed outcome once it is actually done (workflow: on the named node). Merged, never rebuilt.
Inputs: agent_id (required), clear, units, minutes_saved, +1 more.
Point an agent's attached API endpoint at authentication BY REFERENCE (never inline secrets). Either reuse an existing AuthProfile (auth_profile_id), define one inline (auth_method+auth_config), or clear the per-agent override (use_endpoint_auth=false).
Inputs: agent_id (required), endpoint_id (required), auth_config, auth_method, +3 more.
Create or update an agent's goals — targets on quality, speed, volume, outcomes, cost or ROI that Agent Performance tracks. Each goal: {metric, target, comparator?, period?, name?, node?, baseline_value?, notify?, id?}.
Inputs: goals (required), agent_id (required), replace.
Attach existing governance policies to the agent. mode=replace sets the exact list; add/remove adjust it.
Inputs: agent_id (required), policy_ids (required), mode.
Enable, tune, or disable Grounding Trace (claim-to-source verification of regulated report output) on the agent. Assistive, never a certification. Pass disable=true to turn it off.
Inputs: agent_id (required), manual, disable, agent_invoked, +2 more.
Activate, deactivate, or draft a partner binding. Only an ACTIVE binding delivers.
Inputs: status (required), binding_id (required).
Record the AI readiness check: five dimensions, 1–4 each (workflows_documented, data_access, governance, adoption, business_case) → score 0–100 and a level.
Inputs: answers (required).
Set the organization's ROI settings — what one hour of a person's work costs (hourly_cost), plus optional default minutes saved per outcome, a fixed monthly fee and whether platform cost is subtracted. Always in the organization's own currency. Needs Agent Performance manage permission. Only the fields sent change.
Inputs: hourly_cost, fixed_monthly_fee, default_minutes_saved, include_platform_cost.
Create or update (by name) a scheduled run of the agent. Provide EXACTLY ONE cadence form: cadence (frequency-based), every+period (interval), cron_expression, or run_at (one-shot). AGENT mode uses instruction; TOOL mode uses action_tool+action_params (fires one tool deterministically).
Inputs: name (required), agent_id (required), every, period, +11 more.
Replace the agent's trigger list (creates/disables the live webhook rows that let external events invoke it). Send the full desired list; omitted triggers are removed.
Inputs: agent_id (required), used_triggers (required).
Turn 'Use organization context' on or off: when on, the organization profile is added to the agent's instructions at run time. Off by default.
Inputs: enabled (required), agent_id (required).
Bring an archived agent back to draft status so it can be edited and re-published.
Inputs: agent_id (required).
Update a draft's fields (model, tools, system message, knowledge...). status/workflow/organization/user are refused with reasons.
Inputs: fields (required), agent_id (required).
Patch an existing custom API endpoint. Send only the fields to change (same field set as create_api_endpoint).
Inputs: endpoint_id (required), url, name, method, +1 more.
Patch a deployment's name/description/config. config_patch merges key-wise (explicit null deletes a key). Web widget: get_widget_settings lists every setting; texts and behaviour are top-level, ALL colours go inside config_patch.theme (e.g. {"theme": {"primaryColor": "#fadc13"}}); unknown keys and invalid values are refused.
Inputs: deployment_id (required), name, description, config_patch.
Edit an existing policy: name, description, severity, which side it checks (validation_points) or individual configuration keys (configuration_patch, merged into what is stored). Does not activate or deactivate; editing an ACTIVE policy changes live enforcement.
Inputs: policy_id (required), name, severity, description, +2 more.
Patch an existing MCP server. Send only the fields to change (same field set as create_mcp_server).
Inputs: server_id (required), url, name, transport.
Change an opportunity's inputs, Blueprint or status (idea | building | live | archived); the estimate is recomputed.
Inputs: opportunity_id (required), title, status, systems, +7 more.
Change the organization profile (needs the org-settings change permission). Sent fields are written; an empty value clears that field.
Inputs: size, markets, website, industry, +2 more.
Patch a partner binding's fields (send only what changes). Does not change status (use set_partner_binding_status) or the signing secret (use rotate_partner_binding_secret).
Inputs: binding_id (required), name, url_key, static_url, +4 more.
Rename or re-tag an existing vector store.
Inputs: store_id (required), name, metadata.
Pre-flight a draft without running it: checks workflow-graph validity, that every needed tool resolves, and that the model is in the org's allowed set. Returns the problems found (empty = ready) plus advisory warnings that never block (e.g. a workflow node without its own model). Read-only.
Inputs: agent_id (required).
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
MCP gateway: one connection for your AI apps, with tool policies, PII redaction and audit logs.
One MCP endpoint for Claude, GPT & Gemini: 100+ tools + no-code connectors + agent workers.
214 tools for AI agents over one MCP endpoint: image, video, audio, 3D, data, storage, mail, GPUs.
Add one secure layer between your agents and this server.
