lobe-forge MCP server
Commit Nov 3, 2026Provides a shared Git-backed Markdown wiki for AI tools to read, write, and cite long-term memory with human review capabilities.
Also known as MinimalOne/lobe-forge
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 3, 2026
- License
- MIT
What is lobe-forge MCP?
lobe-forge provides a shared Git-backed Markdown wiki for AI tools to read, write, and cite long-term memory with human review capabilities.
Summary generated from the project's README
About lobe-forge MCP server
Building a shared knowledge base that AI assistants can read, write, and cite is what lobe-forge enables. It allows users to create organized wikis called lobes, which are stored as plain Markdown files within Git repositories. This structure ensures that your data remains portable and under your control, rather than being locked inside a specific application. You can run it locally on your computer or host it on a remote server using lobe-server and Docker to access your notes from multiple devices.
Developers can connect various AI tools through the Model Context Protocol, including Claude Code, Cursor, VS Code, and Gemini CLI. The system distinguishes between AI-generated content and human-verified information by using specific stamps that only a person can provide. It also manages knowledge decay through page types like Runbook or Decision, which have different shelf lives. To maintain quality, the lint tool identifies broken links, orphan pages, and outdated content that requires human review.
What you can do with it
- 1Create a cited knowledge baseAsk an AI to transform raw PDFs or web pages into organized, linked wiki pages with proper citations.
- 2Maintain long-term information accuracyUse the linting feature to find stale pages or broken links that need human verification.
- 3Synchronize memory across different AI toolsShare a single set of facts between a coding assistant and a mobile chat application.
- 4Ensure human oversight of AI notesVerify that specific wiki pages have been checked by a person to ensure they are still correct.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableminimalone/lobe-forge is public and reachablegithub.com/minimalone/lobe-forge ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-03github.com/minimalone/lobe-forge ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHublobe-forge
A shared memory for your AI tools: wikis that your AI writes, you own, and every AI tool you use can read.
You collect things worth keeping: a web page, a PDF, meeting notes, a chat. Your AI assistant reads them and writes tidy, linked wiki pages that say where each fact came from. Later you ask a question, and the assistant answers from those pages and cites them.
Each topic gets its own wiki, called a lobe: homelab, work, family, german. A lobe is an ordinary folder of text files with full version history (Git), so nothing is locked inside an app and every change can be undone.
- Works with the AI you already use, through the Model Context Protocol (MCP):
- AI coding tools: Claude Code, Codex CLI, Gemini CLI, Antigravity, Cursor, VS Code
- AI chat apps: Claude.ai, Claude Desktop, Claude mobile
- Personal AI agents: OpenClaw, Hermes Agent
- One memory, many tools. A fact one assistant saves today, another can answer from tomorrow.
- On your computer or your own server. Keep lobes on your computer, or run
lobe-serveron a home server and reach them from every device, including your phone. - Yours. Plain Markdown files in Git. Push them to any Git host, or keep them entirely offline.

flowchart LR
S["Sources<br/>web page, PDF, notes"] -->|add-raw| R["raw/"]
R -->|ingest| W["wiki/ pages<br/>cited and linked"]
W -->|ask| A["Answer with citations"]
subgraph Lobe["A lobe: one Git repository"]
R
W
end
T1["Claude Code"] -->|MCP| Lobe
T2["Codex, Gemini CLI, Cursor"] -->|MCP| Lobe
T3["Claude.ai, mobile"] -->|MCP| Lobe
Words you'll see
| Word | What it means |
|---|---|
| Lobe | One wiki about one topic. It is a folder on disk that is also a Git repository. |
| Source | Something you add, such as a web page, PDF or note. It is kept unchanged in the lobe's raw/ folder. |
| Page | A wiki page the AI writes from your sources, in Markdown (plain text with simple formatting). |
| Ingest | Asking the AI to read new sources and turn them into pages. |
| Lint | A health check of the wiki: broken links, outdated pages, pages no person has reviewed. |
| Verify | A person confirming that a page is correct. Only a person can do this, never the AI. |
| MCP | Model Context Protocol, the standard way AI tools connect to outside data. |
| Git | Version history for files. Every change is saved, can be undone, and can be copied to GitHub or another host. |
| lobe-server | An optional program that hosts your lobes on a server, so all your devices and AI tools reach the same lobes. |
| Token | A long secret password that an AI tool sends to lobe-server to prove it may connect. |
Why it matters
Letting an AI keep your notes is easy. Trusting those notes a year later is not. An AI-written wiki goes wrong in quiet ways, and lobe-forge is built around the three most common:
- You can't tell which pages a person checked. lobe-forge records on every page who last wrote it and whether a person reviewed it. An AI cannot mark its own work as reviewed, and any later change shows the page as "changed since review".
- A page that was right in spring is wrong by autumn. Each page has a type with its own shelf life (a how-to goes stale sooner than a decision record), and lint lists what is due for a check.
- The knowledge ends up locked in one app. Pages are plain Markdown in Git, and a whole lobe exports to the open Open Knowledge Format.
What you get
- Answers with sources. Every claim on a page cites the source it came from, and answers cite the pages they used.
- Search that finds the right page. Matches in a page's title, aliases and one-line facts ("atoms": commands, fixes, decisions) rank above matches in the body. Assistants read only the section they need and edit pages in place instead of rewriting them.
- Provenance and human review.
generatedandverifiedstamps that only the server writes, and a review that only a person can give. How trust works - Shelf life by page type. Note, Runbook, Reference, Decision, Session and Todo, each with a default staleness period you can override per page. Page types
- A health check. Lint finds broken links, orphan pages, index drift, likely duplicates, stale pages and pages changed since review, and ignores anything inside code blocks. Lint
- Safe when several tools write at once. Every change is a Git commit. Writes take turns, and a write based on an outdated copy of a page is refused instead of overwriting newer work.
- Open export.
export-okfturns a lobe into an OKF v0.2 bundle that any tool can read. Export - Nothing extra to install. The wiki tools and lobe-server use only the Python standard library, so they run anywhere Python 3.10 and Git do, including a Raspberry Pi. The optional
lobectlcommand-line tool is a single Go binary.
Choose how to run it
| On your computer (local lobes) | On your own server (remote lobes) | |
|---|---|---|
| Where lobes live | ~/lobes/<name> on your computer |
on your server, with lobe-server in Docker |
| How AI tools connect | the Claude Code plugin, or any MCP tool on the same computer | any MCP tool, over the internet (HTTPS) |
| Best for | one computer, everything offline | several devices, phone access, many AI tools |
| Guide | Local lobes | Self-hosted lobe-server |
Not sure? Start on your computer. Both use the same lobe format, so a lobe can move to a server later.
Quick start
On your computer, with Claude Code
You need Claude Code, Git, and Python 3.10 or newer.
-
Install the plugin:
claude plugin marketplace add MinimalOne/lobe-forge claude plugin install lobe-forge@lobe-forge -
In Claude Code, run
/lobe-forge. Give the lobe a name (for examplehomelab) and say what it is for. -
Restart Claude Code.
Now talk to it in plain words: "save this to homelab", "ask homelab how the reverse proxy is set up". Each lobe also gets four slash commands, for example /homelab-add-raw, /homelab-ingest, /homelab-ask and /homelab-lint.
To use the same lobes from Codex, Gemini CLI, Antigravity or another MCP tool, see Local lobes without a server.
On your own server, with lobe-server
You need Docker. Replace change-me with a long random secret (for example from openssl rand -hex 32): it is the token your AI tools use to connect.
docker run -d --name lobe-server -p 127.0.0.1:8080:8080 \
-e LOBE_API_TOKEN=change-me -v ~/lobes:/data/lobes \
ghcr.io/minimalone/lobe-forge:latest
claude mcp add --transport http --scope user lobe-server http://127.0.0.1:8080/mcp \
--header "Authorization: Bearer change-me"
Then ask your assistant: "Create a homelab lobe on lobe-server for my server notes."
This only listens on the same machine. To reach it from other devices, with HTTPS, Git backup and sign-in for Claude.ai and mobile, follow docs/server.md. To connect other AI tools, see docs/clients.md.
Documentation
| Guide | Read it when you want to |
|---|---|
| Connect AI tools | Add your lobes to Claude, Codex, Gemini CLI, Antigravity, Cursor, VS Code, OpenClaw or Hermes Agent |
| Local lobes | Create, adopt, move, upgrade or remove lobes on your computer |
| Self-hosted lobe-server | Host lobes on a server with Docker, Git push, sign-in (OAuth) and human review |
| How it works | Understand what is inside a lobe, local versus server, trust stamps, page types, lint, export and Git sync |
| Reference | Look up settings, MCP tools, REST endpoints and lobectl commands |
| OpenAPI spec | Generate a REST client |
| Contributing | Run the tests and understand releases |
License
lobe-forge MCP questions
Can I run this locally or on a server?
Yes, you can keep lobes on your local computer or run lobe-server on a home server to access them from multiple devices.
How does the system ensure information is trustworthy?
It uses generated and verified stamps to track authorship and human review, and includes a linting tool to find stale or unreviewed pages.
What happens if multiple tools try to write at the same time?
The system is safe for concurrent writes because every change is a Git commit and writes take turns to prevent overwriting newer work.