Beever Atlas

by beever-ai

Not rated

GitHub

About

Open-source LLM knowledge base that turns team chat (Slack, Discord, Teams, Mattermost) into a typed knowledge graph and auto-generated wiki, exposed via a 28-tool MCP server.

Details

Author: beever-ai
Categories: AI, Knowledge Base, Communication

Option 1 — One-line install (recommended)

Theatlasinstaller walks you through a guided 5-step checklist:
- Embedding model— pick a provider (Jina / OpenAI / Cohere / Voyage / Gemini / Mistral / Ollama), then its API key.
- Agent LLM provider— pick a provider for the 16 ADK agents (Google Gemini / OpenAI / Anthropic / Mistral / DeepSeek / Groq / MiniMax / Ollama / Custom); optional second provider for hybrid setups.
- Graph backend— Neo4j (default) or skip.
- Optional integrations— Tavily web search, MCP server for Claude Code / Cursor.
- Auth tokens— keep dev defaults or rotate now.

Under the hood it verifiesdocker+docker compose, copies.env.example→.env(preserves your values on re-run,chmod 600), auto-generatesCREDENTIAL_MASTER_KEY(64 hex) andWEAVIATE_API_KEY(32 hex), runs a port-conflict preflight, launches the stack viadocker compose up -d --build --force-recreate --remove-orphans, and polls/api/healthbefore printing the ready card.

When you see"Beever Atlas is ready", open](https://github.com/Beever-AI/beever-atlas/blob/HEAD/demo/README.md)http://localhost:3000— thenSettings → AI Setupto manage providers, assign LLMs per-agent, run Test Connection, or discover models. For CI / Docker / GitOps, configure declaratively:BEEVER_LLM_API_KEY=...(single-provider shortcut),BEEVER_ENDPOINTS='[...]'+BEEVER_PRESET=..., or commit anatlas.yamland runatlas apply— seedocs/runbooks/ai-setup.mdand docs/runbooks/atlas-yaml.md.

For CI or unattended installs — skip prompts, pre-seed keys from shell env:

GOOGLE_API_KEY=... JINA_API_KEY=... ./atlas --non-interactive

Re-running./atlason an existing stack is idempotent.

Open.envand fill in the two required keys:

GOOGLE_API_KEY=your_gemini_key JINA_API_KEY=your_jina_key

Generate two required secrets and paste them into.env:

# CREDENTIAL_MASTER_KEY — AES-256-GCM key for stored platform credentials (64 hex chars) python -c "import secrets; print(secrets.token_hex(32))" # WEAVIATE_API_KEY — auth between backend and Weaviate (required by docker-compose) python -c "import secrets; print(secrets.token_hex(16))"

First run takes 2–3 minutes while images build and databases initialize. Subsequent runs start in seconds.

Databases in Docker, app services native for hot-reload.

Prerequisites:Python 3.12+ with uv, Node.js 20+

cp .env.example .env # Fill in GOOGLE_API_KEY, JINA_API_KEY, CREDENTIAL_MASTER_KEY, WEAVIATE_API_KEY (same as Option 2) # Start just the databases docker compose up -d weaviate neo4j mongodb redis # Backend (terminal 1) uv sync uv run uvicorn beever_atlas.server.app:app --reload --port 8000 # Bot (terminal 2) cd bot && npm install && npm run dev # Web (terminal 3) — Vite dev server with HMR cd web && npm install && npm run dev

Openhttp://localhost:5173(the Vite dev port —not:3000).

The Vite dev server proxies/api/tohttp://localhost:8000(configured viaVITE_API_URL).

.env.exampledefaults are tuned for local testing. Before any real deploy, rotate the secrets that ship with placeholder values and flip the environment flag:

Option 1 (./atlas) handles all of this through the"Rotate auth tokens"prompt in step 4 of the checklist — answerYand the installer generates random tokens and mirrors the VITE_ values for you. If you used Option 2 or 3, you can re-run./atlason the existing.env, skip every other prompt with Enter, and only accept the rotation prompt.

Navigate to the URL for your chosen option:

- Options 1 & 2→http://localhost:3000
- Option 3→http://localhost:5173

- Real mode(default,ADAPTER_MOCK=false): connect a workspace inSettings → Connections— Slack / Discord / Teams tokens are entered through the UI, not.env.
- Mock mode(ADAPTER_MOCK=true): uses fixture data — opt in for local UI iteration without platform credentials.

From the dashboard:Connections → Add Workspace → Select channels → Sync.

Or via API (auto-extracts your bearer token from.env):

curl -X POST http://localhost:8000/api/channels/C12345/sync \ -H "Authorization: Bearer $(grep -E '^BEEVER_API_KEYS=' .env | cut -d= -f2 | cut -d, -f1)"

Media shared in synced channels (images, PDFs, video) is persisted durably so it keeps rendering after the platform CDN link expires. It defaults to in-database storage with zero extra infra, and can use MinIO/S3 at scale. See docs/media-persistence.mdfor the mechanism,CHANNEL_MEDIA_configuration, the MinIO/S3 backend, and backfilling existing channels.

Beever Atlas exposes a curated MCP (Model Context Protocol) server at/mcpfor AI agents like Claude Code and Cursor. This allows external code assistants to query your team's knowledge base without using the dashboard.

- Tool catalog— 28 tools for discovery, retrieval, wiki reading, graph traversal, and long-running operations
- Auth setup— generating and managingBEEVER_MCP_API_KEYS
- Client configuration— ready-to-use.mcp.jsontemplates for Claude Code and Cursor
- Rate limits— principal-keyed limits to prevent one agent from throttling others
It also ships astandalone stdio mode(python -m beever_atlas.api.mcp_server/beever-atlas-mcp) that exposes the same tool catalog with no HTTP server or backing stores — handy for MCP registries (Glama.ai) and local introspection. See docs/mcp-server.md.

{ "mcpServers": { "beever-atlas": { "url": "https://atlas.example.com/mcp", "transport": "streamable-http", "headers": { "Authorization": "Bearer ${BEEVER_MCP_KEY}" } } } }

docker compose up -d # Start in background docker compose logs -f beever-atlas # Tail backend logs docker compose down # Stop (keeps data) docker compose down -v # Stop and DELETE all indexed data make demo # Full stack + seeded demo corpus make docker-up # Shortcut for docker compose up -d`
Beever Atlas collects no telemetry. No usage data, error reports, or analytics are sent anywhere by default. All LLM calls go through API keys you configure in your own.env, and all data stays in the databases you control.

All/api/endpoints areUNSTABLEin 0.1.0. v0.2.0 will introduce a/api/v1/*prefix; clients pinning current paths will break. SeeSECURITY.md.


- Discord:discord.gg/VshBCUUX— get help, share what you're building, talk to the team

- X / Twitter:@Beever_AI— release notes, posts, announcements

- Website:beever.ai— about the company and other projects

- GitHub Discussions:github.com/Beever-AI/beever-atlas/discussions— longer-form questions and ideas

Commercial support, partnerships, or press:tech@beever.ai`.

Security policy:SECURITY.md| Community standards:CODE_OF_CONDUCT.md

Local stdio MCP server for coding agents: shared troubleshooting memory, agent-to-agent help posts, replies, and a unified inbox. Quick start: npx -y agent-only-mcp

One shared context every AI tool your team uses reads and writes over MCP, so Claude Code, Cursor and ChatGPT stay current together.

Your company's brain, connected to Claude, ChatGPT, Gemini, Cursor, and VS Code. Turn your team's email history into shared memory that any AI assistant can query with natural language.

Source-grounded MCP, A2A, and x402 metadata for verified knowledge commerce.

Local-first deterministic project memory for AI agents: context packs, decisions, gates, risks, claims ledger and explicit checkpoints in project-owned files.

Open-source team memory layer for AI coding agents. Markdown files in git, a user→team→org hierarchy, one cross-vendor MCP server. Apache-2.0.

Access control, conflict resolution, and audit for shared agent memory.

One memory. Every AI. Nothing to install. Memxus works across Claude, ChatGPT, Cursor and Slack — and any app your team uses that needs to remember (Notion, Linear/Jira, Gmail, Salesforce/HubSpot, Intercom, Zendesk, Google Drive, Confluence, GitHub). No extension, no local setup, no repeated context. Save once. Remember everywhere.

Capture AI coding sessions (Claude Code / Codex / Cursor) and distill them into weekly reports, ADRs and a knowledge graph — self-hosted.

MCP server for Obsidian vaults — search, memory, link graph, 23 tools, OAuth-protected.

Turn your team's Slack, Discord, Teams & Mattermost chats

into a self-maintaining wiki — automatically.

Beever Atlas pulls the conversations your team already has on Slack, Discord, Microsoft Teams, and Mattermost, extracts atomic facts, deduplicates them, and clusters them into topic pages with citations. A graph store links the people, decisions, and projects mentioned across channels. Ask questions in natural language and get answers cited back to the source messages — through the dashboard, or through MCP into Claude Code and Cursor.

If you want a knowledge base that grows on its own from the chats your team already has, this is it.

Six short clips — connect a workspace, sync history, watch memory build, browse the auto-generated wiki, ask questions, plug external AI agents in via MCP.

Conversations from any supported platform flow into a unified ingestion pipeline that produces two complementary memory systems — a3-tier semantic store(channel / topic / atomic fact) for fast hybrid search, and agraph storethat extracts entities and their relationships. Those memories fuel two consumer surfaces: theLLM Wiki(distilled, auto-maintained) andQA Agents(served through the dashboard directly, or throughMCPinto Claude Code / Cursor).

From chat platforms to MCP agents — one ingestion path, two memory systems, two delivery surfaces.

Under the hood, three services (backend, bot, frontend) are backed by four data stores (Weaviate, Neo4j, MongoDB, Redis). See thearchitecture overviewon the documentation site for the full design — component responsibilities, dual-memory internals, and the smart query router.

Most RAG systems answer questions by retrieving raw message snippets and feeding them straight to an LLM. Beever Atlas takes a different approach: it continuously distils conversations into a structured, auto-maintained wiki — with topic pages, entity graphs, decisions, and citations — before any query is issued. When you ask a question, the retrieval layer works against clean, deduplicated knowledge rather than noisy chat history. This means answers are more consistent, citations are traceable to source messages, and the wiki itself becomes a useful artifact your team can browse independently of the Q&A interface. The dual-memory architecture (semantic + graph) lets the query router pick the right retrieval strategy per question, keeping latency low and context precise.

A live auto-generated channel wiki: overview, concept map, topics, FAQ, glossary — distilled from 246 Slack messages, not hand-written.

The inspiration: LLMs read wikis, not chat logs

The per-channel wiki concept is directly inspired by Andrej Karpathy's observationthat LLMs are far better at reasoning over curated, encyclopedic content (books, docs, wikis) than over raw conversational transcripts. Chat history is noisy, redundant, temporally scattered, and full of implicit context that only humans resolve. A wiki, by contrast, is thealready-distilledform of that knowledge — deduplicated, structured, citation-bearing, and organised by topic rather than by timestamp.

Beever Atlas operationalises this insight: every synced channel gets its ownauto-generated, continuously-updated wiki— sections for topics, entities, decisions, open questions, and timelines — rebuilt incrementally as new messages arrive. The QA agent retrieves against this wiki first, falling back to raw messages only when a fact hasn't been distilled yet.

- Better answers, fewer hallucinations— retrieval operates on fact-dense prose with explicit entity relationships, not on fragmented turn-by-turn chat.
- Traceable citations— every wiki claim links back to the source messages that produced it, so answers are auditable all the way down to the original Slack/Discord/Teams thread.
- A browsable artifact, not just a Q&A box— the wiki is usefulon its own. New teammates onboarding to a channel can read the distilled wiki instead of scrolling three months of history.
- Cheaper inference at query time— the expensive distillation work happens once, at ingestion. Queries hit compact, pre-digested context instead of re-summarising raw logs on every request.
- Graph-aware reasoning— the entity graph built alongside the wiki lets the query router answer relational questions ("who worked on X with Y?") that pure vector RAG struggles with.

For a detailed comparison with other LLM knowledge tools, see the comparison pageon the documentation site.

Beever Atlas ships as a Docker Compose stack (backend + bot + web + 4 datastores). You can try a seeded demo in 30 seconds with zero keys, then pick one ofthree deployment optionsto install it for real.

git clone https://github.com/beever-ai/beever-atlas.git cd beever-atlas

2. Try the demo first (optional, no keys needed for seeding)

make demobrings up the full stack pre-loaded with a public Wikipedia corpus (Ada Lovelace + Python history). Seeding uses pre-computed fixtures — no API keys required. Asking questions via/api/askneeds a free-tierGOOGLE_API_KEYbecause the QA agent calls Gemini. See demo/README.mdfor curl examples.

Skip this step if you're ready to install for real.

Two free keys are required before installing. Both offer generous free tiers — enough to sync a small team's channels for testing.

Optional (skip unless you know you need them):

Tip:Keep the two required keys handy before you start. Option 1 prompts for them interactively; Options 2 and 3 need them pasted into.env.