Continuity
Optional local semantic retrieval integration that uses Ollama to improve natural-language search and context selection over project sources.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Continuitysearch the project for how handoffs are structured"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Continuity is a local continuity layer. It keeps project identity, memory, provenance and structured handoffs stable while agents and sessions change. Git and your project files stay the source of truth.

The local Dashboard with a demo project. No hosted service, account or model key.
Why Continuity
You switch from Claude Code to Codex, start a fresh session, or hand work to another agent. The agent changes. The project does not.
Without a shared layer, every new session rediscovers the same rules, repeats the same decisions and loses what the last agent was halfway through. Continuity keeps that state with the project, not with a chat:
Claude Code ──┐ ┌──▶ Codex
▼ │
┌────────────────────────────┴───┐
│ Continuity │
│ identity · memory · handoffs │
│ provenance · context audit │
└────────────────────────────────┘
▲ │
CLI / MCP ────┘ └──▶ the next sessionContinuity is | Continuity is not |
Local project continuity across agents and sessions | Another coding agent or an IDE |
Durable project memory with provenance and conflict quarantine | A chat archive or transcript store |
Structured handoffs between agents | A replacement for Git |
Project and workspace isolation | A cloud memory service |
A CLI, MCP server and loopback API | A replacement for native Read, Grep or Git tools |
Related MCP server: @vemdev/mcp-server
What it keeps
Project identity | A local ID bound to the canonical project directory, independent of display name, agent or session. |
Workspaces | Git worktrees share one project identity but keep separate current sources. |
Memory | Source-backed facts and attributed agent lessons, each with its origin and trust level. |
Handoffs | Goal, completed and remaining work, decisions, risks, changed files and the next action. |
Provenance | Source paths, SHA-256 versions and capture times behind every context item. |
Context audit | Which sources and memories were selected for a context bundle, and why others were excluded. |
How it works
continuity initregisters a project.continuity syncbuilds a bounded, hashed index of its current text sources.Agents connect through MCP, the CLI or the local API. The host fixes the project, so an agent can read and write only that project's state.
Agents propose memories and leave handoffs. Core policy decides what becomes durable, at which trust level, and what is quarantined.
Context requests return a byte-budgeted bundle with provenance and an explanation. Changed or deleted sources can no longer support old claims.
Memory model
Source-backed facts are checked against the current file and keep its path and hash.
Attributed agent lessons activate automatically at a lower
agent_observationtrust.Conflicts are quarantined instead of resolved by whichever write came last.
Current sources outrank agent observations. Routine execution noise is rejected.
Human review is optional: an explicit override, not a queue you must work through.
See the memory model. Automatic memory is on main and
not yet part of a release; v0.1.0 uses explicit review for free-form memory.
Quickstart
Requirements: Node.js 24.13 or newer on the Node 24 line, and Git.
Continuity is not published to the npm registry. Do not install an unrelated
package named continuity.
Latest release (v0.1.0). Download continuity-local-0.1.0.tgz from the
release page, then:
npm install --global ./continuity-local-0.1.0.tgzCurrent main (includes the unreleased changes listed under
Status). Requires pnpm 10.30.1:
git clone https://github.com/cubiix3/continuity.git
cd continuity
pnpm install --frozen-lockfile
pnpm check
pnpm link --globalIf pnpm's global bin directory is not configured, run pnpm setup and open a new
terminal. Without linking, pnpm continuity --project <path> init works from the checkout.
Then, inside your project:
continuity init
continuity sync
continuity dashboardOpen http://127.0.0.1:4783. The browser does not open automatically.
On Windows, current main can keep the Dashboard running and sync projects
automatically after sign-in: continuity startup install, then continuity runtime start
to start it now. It runs as the current user, without administrator rights. See the
background runtime for limits and removal.
Leave a handoff and pick it up in another agent's session:
continuity handoff create --file handoff.json
# later, from another agent in the same project:
continuity handoff latestThe runnable example covers memory and handoff inputs.
To give every new agent session this context automatically, install a provider
hook once (current main): continuity integrate claude install or
continuity integrate codex install. See agent bootstrap.
In interactive Claude Code sessions, the same hooks ask the agent, after a turn that
edited files, whether anything is worth keeping or a handoff is finished; it never
reads chats. See agent lifecycle.
Agent integrations
Integration | Status |
Claude Code | Verified on Windows 2.1.278; automatic startup context verified on 2.1.280 · setup |
Codex | Verified on Windows CLI 0.155.1; automatic startup context verified on 0.156.1 · setup and sandbox notes |
MCP stdio | Seven project-bound tools, tested with the official SDK client and real agents |
CLI and TypeScript host API | Implemented and integration-tested |
Local HTTP v1 | Loopback only; token, origin and scope boundaries tested |
RIVET | Experimental draft/shadow integration; RIVET's own state remains authoritative |
Command Code, Grok | Unverified; Grok structured results remain unresolved |
Ollama, OpenViking | Optional local semantic retrieval · tested versions and limits |
A real cross-agent fixture passes structured work between fresh Claude Code and Codex sessions without sharing transcripts. See adapter contracts.
Dashboard
continuity dashboard serves a local, read-mostly view on 127.0.0.1:
Overview, Projects, Workspaces, Handoffs, Memories, Context Audit, Sources and
Diagnostics. It shows project health, recent handoffs, memory origins and
conflicts, source freshness and historical context selection. Writes are limited
to explicit memory actions and workspace sync. See Dashboard operation and security.
Local-first and security
Local by default. State lives in a local SQLite database outside your repository. No cloud service, account, model key or telemetry.
Loopback only. The Dashboard and HTTP API bind to
127.0.0.1. Every HTTP API request needs a bearer token; the Dashboard adds same-origin checks, a restrictive CSP and a per-session capability for writes.Project isolation. MCP and HTTP clients cannot select another project; cross-project retrieval is denied.
Untrusted source text. Repository content is data, never permission to change Continuity policy.
Bounded sources. Symlinks, hard links, nested repositories, secret files and recognizable credentials are excluded; source text is hashed and re-checked.
Not encrypted. Local state is not encrypted at rest. Secret detection is heuristic.
Read the security model before indexing sensitive projects, and SECURITY.md to report a vulnerability.
Large projects
Large repositories can keep one stable project identity while narrowing
Continuity's local source index with continuity sources set --include ... --exclude ....
The filter lives in local state, never in the repository. See source scope.
This is on main and not yet part of a release.
Architecture
CLI · Dashboard · MCP · local HTTP · TypeScript host API
│
project-bound Core client
│
identity · namespace guard · memory policy · handoffs
retrieval · budgeting · provenance · audit
│
storage ports
┌────────┴─────────┐
SQLite + FTS5 filesystem sourcesPolicy lives in the Core. Adapters receive a scoped capability, never database access. FTS5 retrieval works fully offline; semantic retrieval is optional and not consistently better.
Commands
init Register the current directory
status | doctor Identity and last sync | storage and registration health
sync Refresh the source index
sources show | preview Current source scope | read-only selection and limit check (main)
sources set | clear Replace or remove this project's local source filter (main)
search <query> Current source excerpts
context <task> Build a byte-budgeted context bundle
inspect | explain <id> Read a historical bundle | explain its selection
memory list | show <id> Inspect memories
memory remember <text> Record with --key and --source (--agent/--session on main)
memory forget <id> Deactivate; revision history is kept
memory pending Candidates awaiting review and unresolved conflicts
memory approve | reject <id> Human review with --by <reviewer>
handoff create --file <path> Save structured JSON (use - for stdin)
handoff latest | show <id> Retrieve a handoff
handoff close <id> Mark a handoff finished (kept as history)
bootstrap Read-only startup index for agent sessions (main)
integrate claude|codex Provider startup and autosave hooks: install | status | remove (main)
project list | status Inspect local registrations
project rebind <id> Explicit move with --from and --to
retention status Retention classes and eligibility
prune --dry-run Preview only; never deletes
mcp Serve project-bound tools over stdio
serve Local authenticated HTTP API on 127.0.0.1
dashboard Local Dashboard on 127.0.0.1:4783
runtime start | status | stop Background Dashboard and automatic sync (main)
startup install | remove Windows current-user sign-in startup; also status (main)Commands marked (main) are not in v0.1.0. Global flags: --project <directory>, --home <directory>, --json.
CONTINUITY_HOME overrides the default ~/.continuity state directory.
Documentation
Status
Latest release: v0.1.0. Project identity, workspaces, handoffs, reviewed memory, provenance, context audit, MCP, local HTTP and the Dashboard.
Current main (unreleased): automatic-first durable memory, local per-project
source scope, a refreshed and denser Dashboard, the new brand mark, an optional
background runtime with Windows sign-in startup and automatic sync, and agent
auto-bootstrap for Claude Code and Codex. See the changelog.
Next: a model-aware session-end save step and future release packaging. Agent orchestration, cloud sync and a RIVET cutover are out of scope; the RIVET dogfood notes record what was and was not proven.
Development
pnpm check # build, strict typecheck, tests, lint
pnpm exec playwright install chromium
pnpm test:ui # Dashboard browser tests
pnpm test:pack # install the packed tarball into a fresh projectCI runs these on Linux and Windows. Read CONTRIBUTING.md and AGENTS.md before changing Core boundaries. See operations for rebind, review and retention previews.
License
This server cannot be deployed
Maintenance
Related MCP Connectors
Shared project memory that keeps teammates and AI agents aligned across sessions.
- OneLoreOAuthai.onelore
Shared project context for AI agents and teams: docs, tasks, and messages that stay current.
Local-first, governed memory and session continuity for AI coding agents. No cloud, no telemetry.
Project memory, semantic code search, and grounded agent context.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceProvides durable project context for coding agents, including project maps, session history, and explicit memories, all stored locally.11 npm7MIT
- AlicenseNot gradedqualityDmaintenanceGives AI agents durable project memory via the Model Context Protocol, allowing them to read tasks, record decisions, search context, and sync snapshots to the cloud.4 npmMIT
- AlicenseNot gradedqualityCmaintenanceGives AI agents persistent memory, handoffs, and shared context across sessions, enabling seamless continuity and multi-agent collaboration.65 npm68-
- AlicenseAqualityCmaintenanceProvides coding agents with persistent, evidence-backed project memory and knowledge across sessions, using a structured memory tree and local knowledge base.6124MIT