A shared brain for your team's AI agents
Brainfold is an open source, self-hosted MCP server that turns a folder into your team's shared memory: files with permissions that Claude Code, Codex or any other harness reads and writes. One server per team; each person connects their own agent.
What each agent knows stays with the team
Every person on a team runs their own harness, and what each agent learns stays in that person's session. Brainfold gives them one place to read and write, with the same permissions the people have.
Permissions per folder
Each folder says who reads and who writes it, by person or group. Nothing is open until a folder grants it.
Any harness
Fifteen MCP tools over plain files. Claude Code, Codex, Cursor or any MCP client connects with one command and a personal token.
Skills and learnings
Procedures the team wrote once reach every agent allowed to read them, and agents note what they solve next to the work.
History and activity
Every change is versioned and can be restored. Every call is logged with who made it and whether it was allowed.
Same server, different view
An agent has exactly the access of the person whose token it uses. In this synthetic example, Bruno is on the costs team and Carla is not.
- common
- projects
- costs@costs
- process.md
- 2026-q4-cloud-review.md
- LEARNINGS.md
- skills
read /projects/costs/2026-q4-cloud-review.md✓ 305 bytes · version 09c484c9
@costs. His agent lists the folder and reads the review stored there.- common
- projects
- contracts@contracts
- skills
read /projects/costs/2026-q4-cloud-review.md✗ unavailable: The path is not available.
Try it on your machine
It needs uv. The example brain is synthetic. The README continues with the first account and the token.
# Start the server over the example brain git clone https://github.com/jackspolski1103/brainfold && cd brainfold uv sync cp -R examples/brain /tmp/brainfold-brain uv run brainfold serve --root /tmp/brainfold-brain # Connect a harness with the token you create claude mcp add --transport http brainfold http://127.0.0.1:8020/mcp \ --header "Authorization: Bearer <token>"
Questions
What is Brainfold?
Brainfold is an open source MCP server that turns a folder into a team's shared brain: folders and files with permissions, served over the Model Context Protocol. Each person connects their own harness and their agent reads and writes the same brain, with exactly the access that person has.
Is Brainfold an agent or a model?
No. Brainfold is not an agent and does not run models. It is the memory your agents share. Your harness (Claude Code, Codex or another) stays in charge of the model, the tools and running skills.
Which agent harnesses work with Brainfold?
Any MCP client that speaks HTTP with a bearer token. The panel gives setup commands for Claude Code, Codex and Cursor. The connection was checked with Claude Code; the Codex and Cursor settings follow their documentation.
How do permissions work?
Access is decided per folder, for two actions: read and write, granted to people or groups. Reading is denied unless a folder grants it, and what is denied looks the same as what does not exist. An agent has exactly the access of the person whose token it uses.
How is it different from CLAUDE.md or a memory per agent?
A memory file or a per-agent memory stays with one person and one harness. Brainfold is one server per team: what one agent writes, every teammate's agent that is allowed to read it finds in its next session, whichever harness they use.
Does Brainfold use embeddings or RAG?
No. The brain is plain files with a git history. Agents find what matters through text search and through the description and start-here list that every folder carries.
Is there a hosted version?
No. Brainfold is self-hosted: each team runs its own server, with Docker Compose or on a hosting platform that builds a Dockerfile and provides a volume.
What license is Brainfold under?
Apache-2.0. It is a personal, independent open source project in early development.
Where it stands
Brainfold is young. The server, permissions, history, skills, accounts and the web panel work and are tested, and it runs on the author's own deployment. It has not been used by many teams yet, so expect rough edges and open an issue with what you find.