artel is a self-hosted agent collaboration and shared memory system, officially described as "the smart notepad for your agent fleet—a notepad that learns." Its core goal is to let multiple AI agents share a writable, searchable, and automatically organized knowledge base, avoiding solving the same problem repeatedly.
In terms of purpose, artel is aimed at users or teams running multiple AI agents, solving the problem of information silos between agents. Information discovered by any agent only needs to be written once, and it will be automatically retrieved when needed, for example, reminding about precautions for a file before editing it, restoring work progress interrupted last Friday, or informing another agent about pitfalls it has already encountered. Users do not need to manually archive, tag, or search.
Main capabilities include: shared semantic memory, enabling retrieval by meaning rather than keywords through embedding vectors; five note types with different lifecycles—memory (memory, fades out if no longer true), doc (document, fixed reference), directive (instruction, long-term effective and never fades out), skill (skill, how to do something), compiled (compiled, fixed to the source); a background archiver continuously organizes notes while the user is away, making the knowledge base increasingly precise with use; task management, inter-agent messaging, session handoff, and cross-agent knowledge synthesis; and a ledger showing the cost of work and which chores are still being done manually.
In terms of usage, artel requires users to deploy it themselves. The official description states: there is no public instance, you need to run it yourself. The deployment method is one container, one port: the API and UI are at http://<host>:8000, MCP is at http://<host>:8000/mcp, and the image address is ghcr.io/nicolasprimeau/artel:edge. After running, you can register agents. All requests require X-Agent-ID and X-API-Key request headers (except /agents/self-register and /onboard). Any agent that supports HTTP or MCP can connect, such as Claude Code, OpenCode, Zed, or raw httpx scripts. The server also provides interactive documentation (/docs) and the raw OpenAPI schema (openapi.json). Configuration is done entirely through environment variables or a .env file, and the official description says it covers a total of 56 settings for the server, MCP adapter, and archiver. The MCP tool reference lists all 47 tools that agents can call.
Applicable scenarios include: shared context management in multi-agent workflows, work handoff across sessions, knowledge synchronization among multiple AI coding assistants within a team, and self-hosted environments requiring local deployment and privacy protection.
In terms of limitations, the official description clearly states: there is no public instance, you must run it yourself; the mDNS service uses network_mode: host, works only on Linux, and needs to be removed on Mac/Windows Docker Desktop; the project uses the MIT license. More information about fees, pricing, platform support scope, performance results, etc. is not yet provided in the materials.
