Vibeleaderboard
Index / tool
Visit github.com
Category
Developer Tools
Rank
No. 1222Tools index
Pricing
Open Source
Type
TOOL
GitHub
188 stars
Latest release
v1.3.0
Date

About

Loom is a self-hosted MCP server that gives AI agents shared long-term memory, document search and a code index. Memories are linked to the files and symbols they concern and surface when an agent edits them, so what one agent learns in a session is available to the next. It runs on ClickHouse, with Claude Code and Codex plugins, Python and TypeScript SDKs, and per-agent tokens scoped to namespaces.

What it does

A server you run yourself, backed by the ClickHouse database, that holds what coding bots and custom AI applications should remember between conversations. It stores facts, ingests documents such as PDFs and slide decks for cited retrieval, and parses repositories so a bot can ask where a function is defined and who calls it. Notes can be pinned to particular pieces of code and get flagged when that code changes. A web dashboard shows stored facts, repositories, tasks and runs; clients connect over the Model Context Protocol, plain HTTP, or a command-line tool.

Why it's ranked here

Apache 2.0, published by ClickHouse, Inc., with a documentation tree that covers deployment, upgrades, backup, monitoring, security architecture and a page on when the tool is the wrong choice. The installer boots a working stack with no API keys and no network, then checks that a saved memory reads back. The project publishes its own benchmark table: on the ClickHouse codebase with Claude Code, it reports a 20% lower model bill and higher answer quality, with the caveat that results are specific to the tested workloads. Those numbers come from the vendor, not an outside test.

What's good

Team access is built in rather than bolted on: each agent gets a token scoped to namespaces with reader or writer roles, and tokens can be revoked by ID or by issue time. Maintenance commands that rewrite stored rows default to a dry run and need an explicit apply flag, and every write is stamped with the command that made it, so the audit log shows which operation touched a row. Single-user mode is the same feature set as team mode, bound to localhost. The default embedding model runs on the CPU inside the image, so search works before any provider is configured.

Tradeoffs

This is infrastructure, not a library. The server is shipped as a Docker image and the project explicitly says not to run it from a pip install in production, so you take on a ClickHouse instance, migrations and upgrades. Without a configured language model, the stock install answers with a fake echo model: storage and search work, but extraction, summaries and answers do not. The bundled local embedder is described as weaker than a hosted one, and switching embedding models later means running a re-embed job by hand. The docs also note that offering a tool to a model does not guarantee the model uses it.

How to use it well

Fits a team running several coding agents against the same codebase that wants lessons, decisions and code lookups to persist across sessions, or an agent product that needs per-user memory with audit and erasure. Start in personal mode, add a real language model and hosted embeddings before judging answer quality, and run the vendor's evaluate-on-your-workload guide rather than trusting the headline benchmark. Pick team mode from first boot if more than one person will connect. The project says it is not an in-process library for one agent's short-lived memory and not an agent framework; your own orchestration still runs the agents.

Technical notes+

pyproject.toml declares the server package as Python 3.12+, Apache-2.0, built with hatchling, and marks it Private :: Do Not Upload; the PyPI SDK is a separate package. Core dependencies are FastAPI, uvicorn, clickhouse-connect, asyncpg (for the optional Postgres storage profile), pyjwt, argon2-cffi for password hashing, google-re2 for linear-time caller-supplied regexes, snowballstemmer for verb canonicalisation, and OpenTelemetry exporters. The code index is an optional extra pulling tree-sitter grammars for C++, Python, TypeScript, JavaScript, Rust, Go, Java and Bash plus numpy for a weekly PageRank. server/main.py is the FastAPI entrypoint: its lifespan refuses to boot without an instance label, checks for a real JWT signing secret and safe bind settings, runs fenced schema migrations under a ClickHouse-held lock, and elects one process to run background workers via a lease. docs/reference/cli.md documents the dry-run-by-default convention and per-command query attribution. server/memory/__init__.py is a one-line package docstring.

Observed

License
Apache-2.0 (declared in pyproject.toml and the README badge)
Server language
Python, requires 3.12 or later; FastAPI on uvicorn
Storage
ClickHouse, with Postgres selectable at first boot for live memory state
Distribution
Server as a Docker image; Python SDK on PyPI and TypeScript SDK on npm; Helm chart for Kubernetes
Interfaces
MCP server, HTTP API, Python and TypeScript SDKs, a loom CLI and a web dashboard
Platform support
Docker on Linux or macOS; Windows through WSL 2
Default models
Local CPU embedder and a fake echo LLM until provider keys are configured
Code index languages
C++, Python, TypeScript, JavaScript, Go, Rust, Java (tree-sitter extras)

Read from README.md, pyproject.toml, docs/index.md, docs/reference/cli.md, docs/use-cases/index.md, server/main.py, server/memory/__init__.py.

Tags

mcpagent-memoryclickhousecode-indexclaude-codeself-hosted

Tech Stack

PythonDocker

Media

ClickHouse Loom

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.