agent2kb logo Open source · AGPL-3.0 · pre-1.0 alpha

The knowledge compiler
for agent teams

agent2kb turns everything your agents find into a living, cited Markdown wiki. An LLM compiles it in your own bucket; agents read it over MCP, people read it in Obsidian or any Markdown editor.

Get started → See how it works
Any S3-compatible storage MCP server for agents Runs fully local with Ollama 
# any agent, bot or person adds a source
$ a2kb ingest --url https://example.com/post
✓ raw/20261020-Post-Title-3f9c2a1b.md

# cron compiles it into the wiki
$ a2kb compile
✓ 1 compiled · [[links]] · tags · index

# agents ask, with citations
$ a2kb query "What did we learn?"
→ answer citing wiki/…Post-Title….md

What is agent2kb?

agent2kb is an open-source knowledge compiler for AI agent teams. Agents, bots and people add sources (web pages, PDFs, videos, notes) with one command; an LLM compiles them into an interlinked, cited Markdown wiki stored in your own S3-compatible bucket; agents read it through the CLI or an MCP server, and people read and edit the same pages in Obsidian or any Markdown editor.

  • Type self-hosted CLI, MCP server and Docker backend
  • Licence AGPL-3.0, commercial licence available
  • LLMs Claude Code CLI, Codex CLI, Ollama, Gemini, Groq, OpenRouter
  • Storage AWS S3, Cloudflare R2, MinIO, SeaweedFS, any S3

How it works

How agent2kb works: collect, compile, use

Collection is deliberately simple, understanding is the compile step, and use happens through two interfaces over the same files: the CLI and MCP for agents, Obsidian or any Markdown editor for people.

agent2kb pipeline: a2kb ingest puts sources in raw/, a2kb compile turns them into wiki/ pages with an LLM, agents read over CLI and MCP, people read in Obsidian, and saved answers feed the next compile
1Collect

Any process runs a2kb ingest with a URL, a file or a folder. Web pages, video, audio, PDF, DOCX, images and Web Clipper notes land in raw/.

2Compile

An LLM turns sources into pages with titles, summaries, tags and [[wikilinks]], rewrites only what changed, and labels what it derived.

3Use

Agents query over the CLI or MCP and get cited answers; people read and edit the same Markdown in Obsidian. The index opens with a "Past 7 days" recap, and saved answers feed the next compile.

Why agent2kb

Why agent2kb: understanding, not fragments

Agent teams scatter knowledge across chat logs, scratch notes and browser tabs. A vector store returns fragments; a shared vault mixes projects and drifts. agent2kb keeps one auditable wiki per project, always compiled.

🧠

Compiles, not just retrieves

An LLM writes real pages with titles, summaries, tags and links, so readers get understanding instead of chunks.

🔒

One isolated KB per project

Each KB has its own prefix and, where the provider allows, its own scoped key. An agent on one KB cannot read another.

🤝

Agents and people, one wiki

MCP and CLI for Claude Code, Codex and your bots; plain Markdown with [[wikilinks]] in Obsidian for people.

📈

Knowledge that compounds

Answers you save feed the next compile as derived context, and the index opens with an automatic recap of the past week, so everyone sees what the KB learned.

🪣

Your bucket, no lock-in

AWS S3, Cloudflare R2, B2, MinIO, SeaweedFS, Wasabi. Files you can read, diff, back up and edit by hand.

🛡️

Safe with untrusted content

SSRF-guarded page and file fetches, sandboxed claude and codex with no tools, privacy-first defaults, and a fully local option.

How it compares

Vector store / RAGShared notes vaultAgent session memoryagent2kb
Readable pages, not fragments✕chunks✓hand-written✕transcripts✓LLM-compiled
Kept consistent without manual work✕✕✕✓incremental compile + lint
Isolated knowledge per project◐per index✕one vault◐per agent✓prefix + key per KB
Agents and people share one source✕agents only◐people only✕agents only✓MCP + Markdown editor
Answers cite their sources◐✕✕✓
Plain files in storage you own✕✓✕✓any S3 bucket
Runs with no third party at all◐✓✕✓Ollama + local S3

For people

What your team sees in Obsidian or any Markdown editor

Every compiled page arrives as plain Markdown with its properties, a summary, structured sections and [[wikilinks]] — so graph view, backlinks and tag search work out of the box.

Illustration of a compiled agent2kb page in an Obsidian vault with tags, the article, wikilinks, a graph view and backlinks

Illustration built from the sample note that ships with agent2kb, shown in Obsidian, the reader agent2kb documents and recommends (synced with the Remotely Save plugin). The files are plain Markdown, so other editors can read them too.

Automatic recap

An automatic recap of what is new

Every compile refreshes the top of the index with a recap of the last 7 days: a short summary of the themes the new pages share, written from those pages, and links to each of them.

  • Nobody has to write it. It keeps up with the KB on every scheduled compile.
  • Made for catching up. A teammate back from leave or a learner checking in once a week sees what the KB learned.
  • Your window. 7 days by default, 30 for a monthly recap, or off; an optional weekly digest goes deeper.
wiki/index.mdexample

Past 7 Days (4 articles)

This week the KB focused on safer releases. Two articles compare canary and blue-green deployments and agree that automated rollback matters more than the strategy itself; a third shows how feature flags keep unfinished work out of a release, and an incident review traces an outage to a missing health check, linking it to Rollback and Monitoring.

Anywhere

Every device and every agent, one shared knowledge base

The KB lives in your S3 bucket, not on one machine. Desktop, iPhone, iPad and Android read it in Obsidian; agents on any laptop, server or CI runner read it over the CLI or MCP. Compile runs on a schedule beside the storage, so the wiki keeps improving while your devices are off.

One KB in your S3 bucket, shared by people on desktop, iPhone and Android through Obsidian, and by agents through the CLI and MCP
💻Desktop

Obsidian on macOS, Windows and Linux, synced with the Remotely Save plugin; the vault root is the KB.

📱iPhone, iPad, Android

The same vault on your phone over HTTPS: AWS S3, Cloudflare R2, or the Docker backend's HTTPS profile.

🤖Agents anywhere

Claude Code, Codex and your bots query the same KB with the CLI, or remotely over bearer-token MCP.

Quick start

Quick start: your first page in minutes

Python 3.10+, an S3-compatible bucket (or the Docker backend, which includes storage) and an LLM: a logged-in claude or codex CLI, an API key, or a local Ollama server.

  • Guided setup. a2kb setup asks for storage, LLM, language and privacy, then runs diagnostics.
  • Unattended after that. Cron runs compile, lint, backup and digest; a fallback CLI takes over when one is down.
  • Easiest LLM: the CLI you already use. A logged-in claude or codex CLI needs no API key; switch a KB to a local Ollama model when its content must stay on your machine.
  • Docker in one command. docker compose up -d brings S3 storage with per-KB keys, cron and an MCP server per KB.
# install into a virtual environment
python3 -m venv .venv && source .venv/bin/activate
python -m pip install ".[all]"

# storage, LLM, language, privacy, first KB
a2kb setup
a2kb doctor

# ingest → compile → ask
a2kb ingest --file examples/first-note.md
a2kb compile
a2kb query "What is the codename of the alpha project?"

# give an agent read-only access over MCP
claude mcp add kb-notes -- a2kb -p notes mcp

Built for real teams

Secure and private by default

  • Privacy-first new KBs. Jina Reader and cloud image OCR stay off until you opt in; every run names the provider it uses.
  • SSRF guard on page and file fetches. Private addresses blocked, every redirect checked, connect IP pinned, size capped. Video URLs are fetched by yt-dlp itself, so only ingest them from hosts you trust; over MCP they are limited to YouTube or an allow-list.
  • Sandboxed LLM CLIs. claude and codex run as text transformers: no tools, no MCP, no session.
  • Two layers of isolation. A prefix check in agent2kb plus per-KB credentials scoped by your storage provider.
  • MCP read-only by default. ask and ingest_url are opt-in; HTTP mode needs a bearer token.
  • No telemetry. Data goes only to the storage and the LLM you configure — or nowhere, with Ollama.

Works with

Works with your LLM, storage and sources

LLMs

Claude Code CLICodex CLIGeminiGroqOpenRouterOllama (local)

Storage

AWS S3Cloudflare R2Backblaze B2MinIOSeaweedFSWasabi

Sources

Web pagesYouTube & videoAudioPDF · DOCX · XLSXImages (OCR)MarkdownWeb Clipper

Output language

English by defaultany language you name

FAQ

Frequently asked questions

What is agent2kb used for?

Giving a team of AI agents and the people who work with them one shared, always-up-to-date knowledge base: agents save what they find, agent2kb compiles it into readable wiki pages with citations, and everyone queries or browses the same pages.

How is agent2kb different from RAG or a vector database?

A vector store returns text fragments close to a question. agent2kb compiles sources into readable pages with titles, summaries, tags and links, then searches those pages and answers with citations. Embeddings are optional.

Does agent2kb work with Claude Code and other MCP clients?

Yes. a2kb mcp serves a knowledge base over the Model Context Protocol, read-only by default, over stdio or bearer-token HTTP, so Claude Code, Codex, Claude Desktop and other MCP clients can search and read it.

Is agent2kb free?

Yes. It is open source under the AGPL-3.0; you pay only for the storage and the LLM you choose, or nothing with a local model and local storage. A commercial licence is available for closed-source use.

Is my data used for training?

agent2kb has no telemetry and sends data only to the storage and the LLM you configure. Whether a provider trains on prompts is that provider's policy; a local Ollama server keeps data on your machine.

Can it run fully locally?

Yes. Use Ollama for compile and embeddings, SeaweedFS or MinIO for storage, and enable privacy_mode. Your content then never reaches a third party.

Which LLMs are supported?

A logged-in claude or codex CLI (no API key needed), Gemini, Groq, OpenRouter and Ollama, with separate models for bulk compile and aux jobs.

Does it need a GPU?

No. Only a local Ollama model benefits from one; compiling through a CLI or an API runs remotely.

Can people edit pages?

Yes, in Obsidian or any Markdown editor. A page is regenerated only when its raw source is recompiled; a2kb promote --as wiki-topic keeps hand-written pages untouched.

How do I see what is new?

Every compile refreshes the wiki index, which opens with a "Past 7 days" recap: a short summary of the themes added in that window, written by the LLM, plus links to the new pages. Nobody has to write it. Change the window per KB (for example 30 days for a monthly recap) or turn it off; an optional weekly digest writes a longer recap, a handy reading list for anyone learning from the KB.

Is it production-ready?

It is a pre-1.0 alpha. The CLI, config formats and S3 layout are pinned by contract tests, but expect rough edges and verify on a throwaway stack first.

Licensing

Licence: open source, with a commercial option

AGPL-3.0-only

Community licence

Use, modify and redistribute agent2kb freely. If you run a modified version for others over a network, the AGPL asks you to share that source.

Commercial

Commercial licence

Embedding agent2kb in a closed-source product or service? A separate commercial licence may be available — write to licensing@agent2kb.com.

Give your agents a memory people can read

Ingest once, compile unattended, and let every agent and teammate work from the same cited wiki.

Get started →View on GitHub