busabase/busabase

Open-source database & workspace for AI agents — structured data, durable knowledge, reusable skills, runnable apps, and human review on the writes that matter. Local-first and self-hostable.

Busabase – a shared, version‑controlled workspace for AI agents

What it is – Busabase is an open‑source database + UI that lets large‑language‑model agents (Claude Code, Codex, Cursor, Gemini, etc.) and humans read and write the same structured records, documents, skills, and tiny apps. Every change an agent makes is stored as a Change Request with a diff, author, and history, so the work is durable, auditable, and can be reviewed before it is merged.

Why it matters – Today most LLM‑powered tools keep their output in a temporary chat or a personal note. When you switch tools you lose context and have to re‑explain the project. Busabase gives agents a permanent “system of record” so they can start from the same knowledge base, propose updates, and let you (or a policy) decide whether to accept them.

Quick start (5 minutes)

# Run a local server (embedded PGlite + file storage)
npx busabase server

Open http://localhost:15419/dashboard/local. No DB, account, or config is needed.

Point any LLM‑agent to the server by feeding it the URL of the Agent Skill file, e.g.:

http://localhost:15419/SETUP_SKILL.md

The agent will follow the instructions, read the workspace, and write its output back as a Change Request.

You can also run Busabase via Docker, install the CLI (npm i -g busabase), or download the desktop app (macOS/Windows/Linux) for an offline, local‑first experience.

Core features (as described in the README)

Unified workspace – Typed records (Bases), docs, files, skills, AirApps, whiteboards, and workflows are all first‑class nodes in a single tree.
Shared skills – SKILL.md playbooks and custom prompts live in the workspace; any connected agent can look them up before acting.
Change Requests – Every write is a diff with author, message, and timestamp. Permissions decide whether it merges automatically or waits for review.
Live apps on data – Small “AirApps” run inside Busabase, reading workspace data to provide dashboards, CRMs, content desks, etc.
Agent‑agnostic – No built‑in model. Connect Claude Code, Codex, Cursor, Gemini CLI, n8n, or any custom process via Agent Skill, MCP, OpenAPI, or a simple CLI.
Templates – One‑command installation of a full workspace (schema, sample data, apps, and an agent manual).
Local‑first & open source – MIT‑licensed, runs on an embedded PGlite database or external Postgres/S3, works offline, and powers a hosted SaaS version (Busabase Cloud).

Typical use cases

Domain	What agents do	What stays in Busabase
Software delivery	Turn feedback into tasks, write/test code, ship releases	Feedback, tasks, specs, conventions, release notes
SEO & content	Track trends, draft articles, publish	Keyword research, briefs, drafts, published pages
CRM	Research prospects, log visits, follow‑ups	Companies, contacts, notes, follow‑up history
Team memory	Capture decisions, sources, context	A searchable knowledge base for future agents
Datasets	Label examples, attach evidence, score quality	Curated, reviewed training/evaluation data

Start from a template, e.g. a ready‑made CRM workspace:

busabase-cli install https://github.com/busabase/templates/tree/main/templates/busa-crm

Architecture snapshot

Frontend – apps/busabase – a Next.js UI that runs locally (desktop) or in the cloud.
Engine – packages/busabase-core – implements nodes, relations, file trees, review primitives, search, agent integration, and the public API.
API surface – exposed via MCP, OpenAPI, and the busabase-cli client.
Storage – default embedded PGlite (SQLite‑style Postgres) with file storage; can be swapped for external Postgres and S3‑compatible object storage.
Security model – scoped credentials (changeRequest level, autoMerge flag, etc.) and a review inbox; the open‑source server is intended for trusted local or private‑network use.

Getting involved

Run locally – pnpm install && pnpm --filter busabase dev
Contribute – lint, type‑check, and submit PRs via the GitHub repo.
Community – website, docs, forum, YouTube channel, X, LinkedIn.

License

MIT – free for personal, commercial, or cloud use.

TL;DR – Busabase provides a durable, queryable, and auditable workspace that both humans and LLM agents can read and write. It lets you keep agent output as structured, version‑controlled data rather than fleeting chat logs, making multi‑agent workflows, hand‑offs, and long‑term knowledge retention practical.