AnythingLLM

AnythingLLM is a Node.js application: an Express server handles chats, vector database access and LLM calls, a separate collector service parses uploaded documents, and a React frontend provides the UI. It ships as a desktop app for macOS, Windows and Linux and as a Docker image; storage defaults to a local folder with LanceDB as the vector store, and a pg image variant uses PostgreSQL with PGVector.

The agent runtime (AIbitat in the server code) runs a tool-calling loop for one agent per chat session. Tools come from built-in skills, Agent Flows, custom Node.js skills and MCP servers, and every agent shares the same enabled tools across workspaces while working inside the workspace it was invoked from. The runtime's source defines agents and group channels, but the documentation describes only a single agent per session and no multi-agent orchestration, handoff or supervisor feature, so multi-agent designs are out of scope for this product.

Workspaces hold embedded documents, chat history and settings. The chat API can partition history by an external sessionId. Opt-in memories persist small facts per workspace or globally and are added to the system prompt. Scheduled jobs keep a trace of every run, which can be reopened as a normal workspace thread.

Agent skills that are long-running, expensive or otherwise risky ask the user to approve the tool call before it runs; an administrator can pre-approve specific skills or all of them with the AGENT_AUTO_APPROVED_SKILLS environment variable. A user leaves an agent session by typing exit, and each scheduled job is limited to the tools chosen for it, with every run's thoughts, tool calls and output recorded for later review and follow-up in a thread.

docker pull mintplexlabs/anythingllm:latest

# Start the Docker image (CPU must support AVX2; 2 GB RAM minimum)
export STORAGE_LOCATION=$HOME/anythingllm && mkdir -p $STORAGE_LOCATION && touch "$STORAGE_LOCATION/.env"
docker run -d -p 3001:3001 --cap-add SYS_ADMIN \
  -v ${STORAGE_LOCATION}:/app/server/storage \
  -v ${STORAGE_LOCATION}/.env:/app/server/.env \
  -e STORAGE_DIR="/app/server/storage" \
  mintplexlabs/anythingllm:latest
# In the UI (http://localhost:3001): choose an LLM, create a workspace named
# "research", enable agent skills, and create an API key. MCP servers go in
# $STORAGE_LOCATION/plugins/anythingllm_mcp_servers.json.

# "automatic" mode lets the agent call tools when the model supports tool calling
curl -X POST http://localhost:3001/api/v1/workspace/research/chat \
  -H "Authorization: Bearer $ANYTHINGLLM_API_KEY" -H "Content-Type: application/json" \
  -d '{"message": "Search the web for recent MCP news and summarize it", "mode": "automatic", "sessionId": "demo-1"}'

• The CPU must support AVX2 or the container exits when LanceDB loads; the documented minimum is 2 GB RAM, a 2-core CPU and 5 GB storage.
• Mount a storage folder and .env as shown, or data is lost when the container is rebuilt; UID/GID mismatches between host and .env cause permission errors.
• The Docker image does not start MCP servers when the container boots; they start when the agent needs them, and many servers slow startup. Only MCP tools are supported, not resources, prompts or sampling.
• On the hosted cloud, MCP and custom agent skills are disabled and there is no built-in LLM.
• Scheduled Jobs need v1.13.0 or later and only appear in single-user mode.
• Anonymous telemetry is on by default; set DISABLE_TELEMETRY=true to turn it off.

Yes, as a client. Agents can use tools from MCP servers listed in anythingllm_mcp_servers.json over stdio, SSE or streamable HTTP. Resources, prompts and sampling are not supported, and MCP is disabled on the hosted cloud.

Not as a documented feature. Each chat runs one agent that can use skills, Agent Flows and MCP tools, and the model may chain several flows in one task, but there is no supervisor or handoff between agents.

The code is MIT-licensed and the desktop app and Docker image are free. Mintplex Labs sells managed cloud instances: Basic at $50 per month, Pro at $99 per month, and Enterprise on request.

JavaScript: a Node.js Express server and document collector with a React frontend. Custom agent skills are also written for Node.js.

No official support was found in the README, the docs source or repository code on 30 September 2026.