
A coordination backbone for multi-agent systems that you run on your own infrastructure. Agents connect via HTTP or MCP and get shared memory with semantic search, task queues, async messaging, and session handoffs. Memory entries carry confidence scores that decay over time unless reinforced. An optional archivist agent synthesizes findings, merges conflicts, and promotes stable scratch notes into documentation. Instances mesh together over feeds and mDNS, replicating memory as a CRDT with no central coordinator. Ships as a single Docker container with SQLite, embeddings, a web dashboard, and both REST and MCP interfaces. Useful when you need Claude, AutoGen, or custom agents to coordinate without hardcoding orchestration logic.
Your fleet's smart notepad — one that learns.
One pad that you and every agent you run write into. Whatever any of you figures out is written down once and handed back the moment it matters: the gotcha about this file right before you edit it, where you stopped on Friday, the thing another agent already learned the hard way. Nothing to file, nothing to tag, nothing to look up — a normal notepad waits to be opened, and this one speaks up.
It also doesn't just accumulate. A background archivist works the pile while you're gone, so the pad gets sharper the more the fleet uses it. What one session learns at 3am, the rest know by morning; nobody solves the same thing twice.
You run it on your own machine. None of it goes to anyone's cloud.
| An agent is about to… | Artel says | who wrote it |
|---|---|---|
edit auth.py | "the token refresh silently no-ops when the clock skews" | a different agent, last month |
| start work Monday | "Friday you stopped mid-migration; here's where" | you, before the weekend |
| debug a flaky test | "seen in March — it was the shared fixture, not the test" | an agent on another machine |
| ask a question | the three notes that answer it, before it finishes typing | whoever hit it first |
Nobody opened a file to find any of that, and nobody had to know who to ask.
There is no public instance to point at — this is your notepad, so you run it. One container, one port:
curl -O https://raw.githubusercontent.com/NicolasPrimeau/artel/master/docker-compose.yml
curl -O https://raw.githubusercontent.com/NicolasPrimeau/artel/master/.env.example
cp .env.example .env
# edit .env: set UI_PASSWORD, and a key for the archivist if you want one
# (ANTHROPIC_API_KEY, or OPENROUTER_API_KEY with ARCHIVIST_PROVIDER=openrouter)
docker compose up -d
API + UI at http://<host>:8000, MCP at http://<host>:8000/mcp. Images at ghcr.io/nicolasprimeau/artel:edge.
Once running, register an agent:
curl -fsSL http://<host>:8000/onboard | sh
mDNS note: the
mdnsservice usesnetwork_mode: hostand only works on Linux. Remove it on Mac/Windows Docker Desktop.
A server, a database, and a librarian. Notes go in over HTTP or MCP, embeddings make them findable by meaning rather than keyword, and everything below the queue happens without an agent asking for it.
you · Claude Code · opencode · Claude API · AutoGen
│ push: notes/skills/gotchas in ┄ capture: sessions out
▼
REST / MCP ──► Artel Server ──► SQLite (WAL) + embeddings
├── notes — semantic search · confidence decay · knowledge graph
├── captures queue ──► archivist compaction ──► notes
├── tasks · messages · events · session handoffs
└── archivist — capture · synthesis · merge · decay · promote
│
mesh (CRDT feeds + mDNS) ◄──► your other machines
Each of these has a page in the docs; this is the map.
| The plugin | The half that speaks up — injects the right note at session start, on each prompt, and before you edit a file. |
| Capture | Sessions become notes on their own, spooled in ~10 ms so writing never slows an agent down. |
| Archivist | The part that learns: merges duplicates, resolves contradictions, decays what stopped being true, promotes what held up. |
| Compile mode | Notes about code pinned to the code, so they re-derive instead of rotting. |
| Blueprints | A procedure compiled into a self-expanding task DAG, with contracts the server checks before a run advances. |
| Decisions | Append-only record of what you chose and why — never merged, never decayed. |
| Mesh and feeds | Several machines converging as CRDTs, plus RSS/Atom subscriptions from the outside world. |
| Dashboard | Browse, search, and watch the fleet from a browser. |
Five kinds of note, with different lifespans: memory (fades if it stops being true), doc (settled reference), directive (standing instruction, never fades), skill (how to do a thing), compiled (pinned to source).
Any agent that speaks HTTP or MCP joins — Claude Code, OpenCode, Zed, a raw httpx script. See connecting clients.
All requests require X-Agent-ID and X-API-Key headers (except /agents/self-register and /onboard).
Full REST reference → — every endpoint, generated from the OpenAPI schema. MCP tool reference → — all 47 tools an agent can call.
A running server also serves interactive docs at /docs and the raw schema at openapi.json.
Configured entirely through environment variables (or a .env file). The essentials:
| Variable | Description |
|---|---|
AGENT_KEYS | agent-id:api-key pairs, comma-separated. Optional :proj1;proj2 suffix scopes an agent to projects. |
UI_PASSWORD | Password for the dashboard. |
ANTHROPIC_API_KEY | Enables the archivist. Without it, Artel runs in passive mode. |
REGISTRATION_KEY | Required by /agents/self-register. Unset disables open registration. |
PUBLIC_URL | Externally reachable base URL, used in OAuth metadata and onboarding. |
Full configuration reference → — all 56 settings across the server, MCP adapter, and archivist, generated from the settings classes.
uv sync --dev
uv run pytest tests/ -v
MIT. See LICENSE.md.