What it does
Loom is an open-source MCP server, self-hosted on ClickHouse, that lets agents store what they learn in a session so that a later agent or tool can find it. It also provides document search with citations (RAG) over PDF, .docx and .pptx files, and a code index that coding agents query for definitions, callers and change impact. A dashboard brings memories, repositories, tasks and runs into one place. Agents connect through MCP, HTTP, Python or TypeScript.
It suits teams using coding agents such as Claude Code, Codex or Cursor that need context shared across sessions and agents, and developers building an agent application that needs per-user memory. The publisher reports that, in its own tests on the ClickHouse codebase with Claude Code, answer quality was 0.84 with Loom against 0.78 without it, file localization reached 96.4% against 91.4%, and model cost was about 20% to 21% lower; memory answers scored 94% against 90% while using 91% fewer reader prompt tokens. Model cost excludes Loom's server-side model calls, hosting and indexing.
For a personal install, Docker runs an installer that starts Loom and ClickHouse, applies migrations and opens the dashboard on localhost port 7777. The stock installation needs no keys or network, using a small local embedding model, but real AI answers require configuring an LLM and embeddings with API keys. Kubernetes deployment uses a Helm chart, and a Postgres option can be chosen at first boot for live memory state.
Limitations and requirements
Requires Docker, or Kubernetes with the Helm chart for cluster deployment; on Windows it runs in a WSL 2 terminal. Without configured API keys, the LLM gives no meaningful extraction, summaries or answers, and semantic retrieval uses a local model that is weaker than a hosted one. Single-user mode is intended to stay localhost-only; shared access requires the team deployment setup, including TLS, and authenticated accounts. Benchmark results are specific to the tested workloads, and the memory accuracy difference is descriptive rather than statistically established. Loom is not an in-process library for short-lived single-agent memory and not an agent framework.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
ClickHouse Loom
Shared long-term memory, document search and a code index for AI agents, self-hosted on ClickHouse.
v1.2.0As shown on registry.modelcontextprotocol.io when we last read it