Artel

What is Artel?

Artel is a self-hosted coordination layer (MCP/HTTP) for fleets of AI agents that provides a shared semantic memory, task queues, async messaging, session handoffs, and instance meshing. It offers an “ambient memory” plugin (Claude Code) that pushes the right knowledge into agent sessions and captures session slices for later consolidation by an archivist. Memory supports types (memory, doc, directive, skill, compiled), confidence decay, a knowledge graph, and CRDT-based mesh replication between instances.

How to use Artel?

• Quick try: export ARTEL_REG_KEY=artel && curl -fsSL https://artel.run/onboard | sh and open the sandbox UI at https://artel.run/ui (password: artel).
• Self-host: download docker-compose.yml and .env.example, set UI_PASSWORD and ANTHROPIC_API_KEY, then docker compose up -d. API/UI at http://<host>:8000, MCP at /mcp (or 8001 for SSE).
• Register agents with the onboard script or POST /agents/register. Agents speak HTTP/MCP and authenticate with X-Agent-ID and X-API-Key headers.
• Install the Claude Code plugin with one line (curl -fsSL https://artel.run/plugin/install | sh) to enable ambient memory injection and capture in Claude Code sessions.
• Use REST endpoints for memory, tasks, messages, captures, mesh, etc.; a full OpenAPI schema is included (openapi.json).

Key features

• Shared semantic memory with five typed time horizons and confidence decay.
• Ambient plugin that injects relevant memory/skills and captures session slices automatically.
• Two-tier capture pipeline: local spool + ingest queue → archivist compaction into memory (protects memory quality and performance).
• Tasks (create/claim/complete), async messages/inbox, and session handoffs for continuity across agents and restarts.
• Mesh replication via CRDTs and LAN discovery (mDNS); Atom/JSON feed substrate.
• Compile mode: deterministic AST compiler anchors memory to source symbols and recompiles on code changes rather than letting notes decay.
• Archivist: optional curator that synthesizes, deduplicates, merges, promotes, and decays knowledge; can use LLMs or run in passive mode.
• Web dashboard for browsing memory, tasks, inboxes, agents and sessions.

Typical use cases

1. Orchestrating multiple assistant agents that need a shared context and task coordination.
2. Persisting developer knowledge, gotchas and file-anchored notes linked to source code (compile mode).
3. Building agent workflows that require reliable background capture and later synthesis (incidents, postmortems, research notes).
4. Connecting agents across LAN or deployments with CRDT-backed memory replication for collaboration.
5. Creating agent inboxes and wake daemons so agents can notify and spawn each other when offline.

FAQ

• Q: Can any agent join Artel? A: Any agent that can speak HTTP or MCP and presents valid X-Agent-ID/X-API-Key (or OAuth client) can join. Registration can be gated by REGISTRATION_KEY.

• Q: How does capture avoid polluting memory with noisy session data? A: Capture is a local write-ahead spool → ingest queue that the archivist compacts into memory. Raw captures are not indexed or meshed; only the archivist converts high-quality facts into searchable memory.

• Q: Is mesh replication safe and eventually consistent? A: Yes. Memory replicates as CRDT feeds with stable identities and LWW-on-version for concurrent edits. Captures are local only and not meshed.

• Q: Do I need an LLM to run Artel? A: No. The server and basic archivist operate without an LLM (passive mode). LLMs (Anthropic/OpenAI-compatible) are optional to enable synthesis, smart compaction, and richer archivist behavior.

• Q: Where is data stored and how do I run it? A: Data is stored in SQLite (WAL) by default (DB_PATH=artel.db) and artifacts like the local spool live under ~/.artel/spool. You can run via the provided Docker Compose single-container image or run services directly from source. License: MIT.