Continuum MCP Server

Shared memory + orchestration for your coding agents.

One MCP server. Tell any client to read the plot — it becomes the orchestrator. Dispatches fresh agents in isolated worktrees, shares lessons across every Claude Code / Codex / Cline chat on the project.

Quick Start • Memory & Vectors • Workflow • MCP Tools • Wire It In • Configuration

Your agents forget. Continuum remembers.

Point any MCP client at Continuum and tell it to read the plot. The client becomes an orchestrator — it researches the codebase, verifies the plan, persists a dispatch record, and hands you a ready-to-paste prompt for a fresh agent in an isolated git worktree. Every other Claude Code / Codex / Cline chat on the project sees what's reserved, shares what's been learned, and never re-explains the codebase. One local server: vector knowledge base, orchestration protocol, agent registry — shared across every MCP-speaking app on your machine.

flowchart TD
    CC["Claude Code"]
    CX["Codex"]
    CL["Cline"]
    YOU["YOU"]

    subgraph CONTINUUM["CONTINUUM ORCHESTRATOR"]
        direction LR
        PLOT["PLOT.md<br/>protocol"]
        KNOW["KNOWLEDGE<br/>(vectorized)"]
        AGENTS["AGENTS<br/>registry"]
        PANEL["WEB PANEL<br/>(humans)"]
    end

    DB[("SQLite + sqlite-vec<br/>one file, your disk")]

    CC -- MCP --> CONTINUUM
    CX -- MCP --> CONTINUUM
    CL -- MCP --> CONTINUUM
    YOU -- HTTP --> CONTINUUM
    CONTINUUM --> DB

No SaaS. No telemetry. No keys to manage. Boots in seconds. Survives reboots. Scales with you.

What you get

Related MCP server: trw-mcp

Why Continuum exists

Quick start

🐳 The easy path — one Docker command

docker compose -f docker-compose.dev.yml up

That's it. The whole stack comes up with hot reload, a SQLite browser, and a single shared volume — no Node, no pnpm, no native-build dance on your host machine.

What you get:

The API embeds in-process by default (Snowflake/snowflake-arctic-embed-m-v1.5 via Transformers.js, 768-dim, ONNX q8). No external service required for semantic search. Override via EMBEDDER_URL to point at any HTTP-based embedder (Ollama, TEI, OpenAI-compatible). Turn the bundled TEI container on with the embedder profile:

HF_TOKEN=hf_xxx docker compose -f docker-compose.dev.yml --profile embedder up

Bind mounts keep the source tree on your host — edits are instant. Named volumes hold node_modules so installs survive restarts. Database lives at ./.local-data/dev-orchestrator.db on your host, gitignored, safe to wipe.

Stop everything:

docker compose -f docker-compose.dev.yml down            # keep data
docker compose -f docker-compose.dev.yml down -v         # nuke volumes too

🛠️ Or run on bare metal

pnpm install
pnpm dev

Dev SQLite at ./.local-data/orchestrator.db. Gitignored. Wipe to reset — next boot recreates and migrates forward.

Requires Node >=22.12.0 and pnpm 10.33+. The compose path skips both.

Memory & vectors

This is the headline. Continuum's knowledge layer is what makes your agents stop being amnesiac.

How a lesson becomes memory

agent finishes a dispatch
        │
        ▼
knowledge_create({
  project, agentSlug, slug,
  content: "When wiring rate-limit guards in Nest,
            register them in app.module.ts AFTER
            the auth guard, not before. Order matters.",
  kind: 'situational'
})
        │
        ▼
content → embedder (in-process by default, or Ollama / TEI / OpenAI-compatible)
        │       Snowflake Arctic Embed v1.5 → 768-dim vector
        ▼
SQLite row + sqlite-vec index entry
        │       freshness profile = (URL, model, dim)
        ▼
forever queryable by ANY future agent on this project

How recall works

Phase 1 of every dispatch (the orchestrator does this automatically):

knowledge_list({ project, kind: 'fundamental' })
  └─ returns metadata for every binding lesson — slug + agentSlug + timestamps.
knowledge_get({ project, slug })  // for each fundamental slug
  └─ full body — base branch, worktree convention, no-AI-attribution, etc.

knowledge_search({ project, q: "rate limit middleware nest interceptor order" })
  └─ vector similarity over your prose
  └─ returns top-K ranked metadata (no content)
  └─ does NOT need exact wording — natural language works
knowledge_get({ project, slug })  // for each relevant hit
  └─ pulls the full lesson body

Search and list return metadata only — slug + kind + agentSlug + timestamps. Call knowledge_get for the body of any lesson worth reading. Keeps top-K context cheap; the orchestrator triages first, fetches second.

The result: the agent that's about to plan your task already knows what your last five agents learned the hard way. No prompt engineering. No "here's our convention" copy-paste.

Two kinds of knowledge

The freshness profile

Vectors are pinned to a (EMBEDDER_URL, EMBEDDER_MODEL, EMBEDDER_DIM) triple. Change any one and Continuum marks stale vectors for rebuild on next boot. You stay in control of which embedder owns your memory — local Ollama, self-hosted TEI, or anything OpenAI-shaped.

Pluggable embedders

The embedder is required and runs in-process by default (Snowflake/snowflake-arctic-embed-m-v1.5, 768-dim, ONNX q8). The model weights are bundled in the API image — no extra container, no extra config, semantic search works out of the box.

Override via EMBEDDER_URL to delegate embedding to an HTTP service:

# Option A — Ollama (simplest, fully local)
ollama pull <embedding-model>
# EMBEDDER_URL=http://127.0.0.1:11434/api/embed
# EMBEDDER_MODEL=<embedding-model>   EMBEDDER_DIM=768

# Option B — Hugging Face TEI (production-grade, batched)
HF_TOKEN=hf_xxx docker compose -f docker-compose.dev.yml --profile embedder up embedder
# EMBEDDER_URL=http://127.0.0.1:8080/v1/embeddings
# EMBEDDER_MODEL=<repo-id>   EMBEDDER_DIM=768

# Option C — anything OpenAI-compatible /v1/embeddings
# Works out of the box. Match the dim.

Built on sqlite-vec

The vector index lives in the same SQLite file as your projects, agents, and plot. One file. One backup. One restore. No separate vector DB to provision, no Pinecone bill, no Chroma daemon to babysit. WAL mode keeps reads concurrent with writes; foreign keys keep your registry consistent; busy_timeout = 5000 keeps things calm under contention.

Wire it into your AI client

Claude Code

// ~/.claude.json  →  mcpServers
{
  "mcpServers": {
    "continuum": {
      "transport": "streamable-http",
      "url": "http://127.0.0.1:7776/mcp"
    }
  }
}

Restart Claude Code. Tools appear under mcp__continuum__*.

Codex

# ~/.codex/config.toml
[mcp_servers.continuum]
transport = "streamable-http"
url = "http://127.0.0.1:7776/mcp"

Anything else (Cline, Cursor, custom)

Point any MCP-compatible client at http://127.0.0.1:7776/mcp. Streamable HTTP, stateful sessions (UUID-keyed). Server name continuum.

Your first project

Inside any wired client:

Create a Continuum project for /absolute/path/to/my/repo,
then read the PLOT.md and act as orchestrator
for the following task: <describe what you want done>.

What happens:

1. Client calls project_create({project: "/absolute/path/..."})

2. Server seeds PLOT.md (the orchestration protocol) and an empty knowledge base

3. Client calls plot({project}), reads the protocol, enters orchestrator mode

4. Client follows the 4-phase workflow — researches, verifies, persists a dispatch record, hands you a ready-to-paste prompt for a fresh agent in a git worktree

You paste the prompt into a new Claude Code / Codex session. That agent codes inside an isolated worktree against its reserved paths. Other parallel agents see the new agent's reservations via the coordination brief.

When work merges, the orchestrator records what was learned via knowledge_create — and every future dispatch on this project benefits from it forever.

The workflow

1. Intake

registry_list({project, status:'active'})     // paginated metadata: slug + reservedPaths + timestamps
agent_get({project, slug})                    // bodies (request/plan/implPrompt/coordinationBrief) per active agent
knowledge_list({project, kind:'fundamental'}) // binding-rule metadata
knowledge_get({project, slug})                // body per fundamental slug — read every one
knowledge_search({project, q: "<task intent>"}) // RAG → ranked metadata
knowledge_get({project, slug})                // body for each relevant hit

2. Research

Spawn an Explore / Plan subagent. It reads the codebase against the loaded knowledge. Produces a file-by-file plan: create / modify / delete. Plan only — no code.

3. Verify

Spawn a critique agent. It challenges every file touched, looks for orphans and broken call sites, cross-checks against active agents' reserved paths. Returns annotated deltas.

4. Handoff

agent_create({
  project, slug, branch, worktree,
  reservedPaths,        // paths this agent owns
  request,              // verbatim human task
  plan,                 // final agreed plan
  implPrompt,           // self-contained prompt for the dispatched agent
  coordinationBrief,    // short block for every other active agent
})

Agent lands as draft. Promote with agent_update({status:'active'}) when you dispatch. On merge, record the SHA and bank the lesson:

agent_update({ project, slug, status:'merged', mergedCommit, postMergeNotes })
knowledge_create({ project, agentSlug, slug, content, kind:'situational' })

The orchestrator never edits source, runs regen, merges branches, or pushes commits. It researches, verifies, coordinates, and persists. Code lives in fresh agents inside isolated worktrees.

MCP tools

Every tool takes project (canonical absolute path) as its first argument.

Project

Plot — the orchestration protocol

plot_update validates --- a/PLOT.md / +++ b/PLOT.md / @@ headers strictly. Context mismatch throws hunk_mismatch with the failing hunk header. Re-read with plot, regenerate the diff, retry. No silent overwrites.

Knowledge — the vector memory

knowledge_search runs vector lookup over the sqlite-vec index. Pass natural prose — questions, task statements, descriptions all work. SQL wildcard syntax does not. Response is ranked metadata; call knowledge_get({project, slug}) for the body of any hit worth reading. Same pattern for knowledge_list (use kind: 'fundamental' to enumerate the binding set, then get each).

Agents

Allowed transitions: draft → active → merged | abandoned. merged requires a 7–40 char commit SHA. abandoned requires a reason. Only active agents reserve paths.

registry_list returns the collision matrix metadata in one paginated envelope. To inspect a dispatch's plan or implPrompt, follow with agent_get({project, slug}). Page through with page until hasNextPage is false.

Example session

human: orchestrate adding rate-limit middleware to /api routes

claude: → registry_list({project, status:'active'})  // 0 active
        → knowledge_list({project, kind:'fundamental'})
                                              // metadata for binding rules
        → knowledge_get({project, slug:'base-branch'})       // → main
        → knowledge_get({project, slug:'worktree-convention'}) // → ../<slug>/
        → knowledge_search({project, q:'rate limit middleware nest interceptor'})
                                              // ✨ 2 ranked hits on guard ordering (metadata)
        → knowledge_get({project, slug:'rate-limit-guard-ordering'}) // bodies fetched
        → spawn Plan subagent → plan returned
        → spawn critique subagent → 1 issue: missed `app.module.ts` registration → fixed
        → agent_create({
            slug: 'rate-limit-middleware',
            branch: 'feature/rate-limit',
            worktree: '/abs/path/../rate-limit-middleware',
            reservedPaths: ['apps/api/src/common/rate-limit/**', 'apps/api/src/app.module.ts'],
            request: 'orchestrate adding rate-limit middleware to /api routes',
            plan: '...',
            implPrompt: '...',
            coordinationBrief: 'reserves common/rate-limit/**, edits app.module.ts (additive)'
          })
        → here's the impl prompt to paste into a fresh session: ...

[ task ships ]

claude: → agent_update({slug:'rate-limit-middleware', status:'merged', mergedCommit:'a1b2c3d'})
        → knowledge_create({
            agentSlug: 'rate-limit-middleware',
            slug: 'rate-limit-guard-ordering',
            kind: 'situational',
            content: 'Rate-limit guards must register AFTER auth guard in app.module.ts.
                      Order in providers array decides execution order. Reverse → 401s
                      bypass rate limiting.'
          })
        // ✅ lesson embedded → vector indexed → permanent memory

Next time anyone — you, Claude, Codex — touches anything rate-limit-shaped on this project, that lesson surfaces automatically.

Web panel

Astro 6 + React 19 islands. Tailwind v4. shadcn/ui.

Settings stored in app_settings (SQLite) override env vars. Headless deploys still work via env alone.

Configuration

Both env vars and panel settings supported. Panel takes precedence.

Stack

Scripts

pnpm dev          # turbo run dev — api + web in watch
pnpm build        # turbo run build
pnpm lint         # turbo run lint
pnpm check-types  # turbo run check-types
pnpm format       # prettier across ts/tsx/js/astro/md/json

pnpm -F @continuum/api dev          # api only — nest start --watch
pnpm -F @continuum/api mcp:dev      # api with local SQLite at .local-data/
pnpm -F @continuum/api test         # jest unit tests
pnpm -F @continuum/api test:e2e     # jest e2e

pnpm -F @continuum/web dev          # web only — astro dev
pnpm -F @continuum/web build        # astro build → dist/

Conventions

• project is always a canonical absolute path. Pass the project root, never a worktree subdirectory. Worktree paths throw project_not_found.

• Plot mutations flow through unified diffs. No full-text writes. Bad context → server rejects with the failing @@ header.

• Knowledge updates are whole-content replace. Pass the full new content. Re-embedded on save.

• Reserved paths are JSON arrays of globs. Only active agents reserve.

• Migrations are forward-only on boot. Production blocks destructive migrations unless explicitly opted in.

Troubleshooting

Roadmap

• Cross-project knowledge sharing (kind: 'global')

• Embedding re-rank with a small reranker model

• One-shot import of existing CLAUDE.md / AGENTS.md / Cursor rules → fundamentals

• Read-only public knowledge mirror (publish a project's lessons as a static site)

• Built-in observation hooks for SessionStart / SessionEnd events

Contributing

Contributions welcome. Fork, branch, test, PR. Match Prettier defaults (100-col, double quotes, semicolons, ES5 trailing commas). Keep MCP tool changes Zod-typed.

Support

Stop letting your agents forget.

pnpm install && pnpm dev — and they never will again.