Busabase: An Approval-First Database and Knowledge Base for AI Agents

Uh oh!

There was an error while loading. Please reload this page.

Notifications You must be signed in to change notification settings

Fork 0

Star 9

BranchesTags

Open more actions menu

Folders and files

NameName

Last commit message

Last commit date

Latest commit

History

28 Commits

28 Commits

.github/workflows

.github/workflows

apps

apps

packages

packages

.gitignore

.gitignore

LICENSE

LICENSE

README.md

README.md

biome.json

biome.json

package.json

package.json

pnpm-lock.yaml

pnpm-lock.yaml

pnpm-workspace.yaml

pnpm-workspace.yaml

Repository files navigation

Local-first review database for AI-generated content, business data, datasets & multimodal knowledge.

AI can generate endless data — Busabase is where you review, approve, and merge what's good enough to trust.

中文  · 日本語  · 한국어

Quick Start  · Screenshots  · Use Cases  · API  · Compare  · Roadmap

Busabase is a free and open-source (MIT) app for one simple problem:

AI can generate endless content and data, but someone still needs to decide what is good enough to trust.

Busabase gives that approval process a home — an approval-first database and knowledge base for AI agents. It is a private CMS, project database, and structured source of truth with built-in Change Requests, Operations, comments, audit trails, and a simple API for apps and AI agents.

Free & open source. Local-first. Review-first. Agent-ready. Run it yourself — no SaaS, no account, no vendor. Your data never has to leave your machine.

Quick Start

Pick whichever way you like — all of them give you the same review-first database.

⚡ Run it now — one command, zero setup

npx busabase server

Open http://localhost:15419/dashboard/inbox. That's the whole setup: a full local instance with no database to run and nothing to configure. Busabase seeds example Bases, records, and Change Requests on first request, so you can inspect the review workflow immediately.

npm i -g busabase # install once, then just: busabase server npx busabase-cli --help # the API client on its own (talks to any busabase server)

🐳 Docker

docker run --rm -p 15419:15419 busabase/busabase

Open http://localhost:15419/dashboard/inbox. Stores everything locally — no external services. Images are published to Docker Hub (busabase/busabase) and GHCR (ghcr.io/busabase/busabase).

🖥️ Desktop app

Prefer a native app? Download Busabase for macOS, Windows, and Linux at busabase.com/download. Fully native and fully offline — all your data stays on your computer, never online.

🔧 From source

pnpm install cp apps/busabase/.env.example apps/busabase/.env pnpm --filter busabase dev

Open http://localhost:15419/dashboard/inbox. A local-start check runs first: if dependencies, PG_DATABASE_URL, or STORAGE_URL are missing, it fails with a setup message instead of a blank dashboard. The default .env.example uses PGlite under .data/busabase and local file storage under .data/busabase-storage.

What you get after launch:

an Inbox for reviewing Change Requests

example Bases and records

record-level history and audit trails

local PGlite persistence under .data/busabase

REST API endpoints for apps, workflows, and AI Agents

Connect your AI agent

Busabase has no built-in model — you point your own agent (Claude Code, Cursor, Codex, Gemini CLI, OpenClaw, Hermes…) at it. Paste one prompt and it onboards itself, then proposes changes as Change Requests you approve.

Copy the onboarding prompt

Read and follow the Busabase Agent Skill — it is the single source of truth: http://localhost:15419/SETUP_SKILL.md

Follow its onboarding to set me up, and never merge a ChangeRequest without my approval. Reply to me in English.

→ Bring Your Own Agent — the full flow, then npx skills add busabase/skills installs a permanent skill so you never re-paste it.

Where your data lives

npx busabase server and the desktop app share one local data root so you see the same database whichever way you launch:

~/.busabase/data/ ├── pgdata/ # embedded PGlite database └── storage/ # uploaded attachments

Point them somewhere else with BUSABASE_DATA_DIR (or busabase server --db ), or set PG_DATABASE_URL / STORAGE_URL directly to use external Postgres / S3.

Docker writes to /data inside the container — bind-mount the same host folder to share that one database with the CLI and desktop app:

docker run --rm -p 15419:15419 -v ~/.busabase/data:/data busabase/busabase

(Running from source with pnpm dev instead uses the repo's .env, i.e. .data/busabase.) Only one process can hold the PGlite database at a time, so run one launcher at a time.

Screenshots

Inbox with pending Change Requests, reviewer status, and approval actions Agent-proposed changes before merge, including field diffs and reviewer actions

Record detail page with fields, comments, review history, and lineage Base table showing structured records and rich fields

Records inside a Base — typed fields, rich values, and approval status at a glance Graph view showing relationships between seeded records across Bases

On mobile

Review and approve agent Change Requests from your phone — the same inbox, proposal preview, and trusted records, in the Busabase mobile app.

Why This Exists

Most databases are good at storing data. Most CMS tools are good at publishing content. Most code platforms are good at reviewing files.

Busabase is for the middle layer that AI-heavy teams now need:

Need Busabase gives you

AI drafts a blog post Review it before it becomes a published CMS record

Humans clean QA data Approve high-quality examples before training or evaluation

Agents label videos Check multimodal metadata before it enters the dataset

Agents update project or ERP data Human reviewers approve changes before the system of record changes

A local AI tool needs memory Expose a private, audited API over approved knowledge

Data changes should trigger work Fire webhooks, automations, or external agents after approved merges

Someone changes a record Track who proposed, reviewed, merged, viewed, or deleted it

It is approval-first by default, agent-friendly by design, and still small enough to run locally.

Concepts

Core concepts:

Concept Meaning

Base A table-like collection of records

Field A typed property on a Base

Record An approved row of data

Change Request A reviewable proposal to change data

Operation A create, update, delete, or variant action inside a Change Request

Commit Immutable data snapshot behind an Operation

Comment Discussion attached to records, Change Requests, operations, or commits

Audit Event A trail of important reads, writes, reviews, merges, and deletes

Use Cases

Any base an agent can write — reviewed first. A few of the things people build on Busabase:

Use case What you review

Blog CMS for Next.js AI-drafted posts, reviewed before they publish

SEO Landing Pages AI-generated landing pages, approved before they go live

Configuration Management Service YAML/JSON reviewed before it ships

Finance & Invoice Review Automated finance entries a human signs off

Data Stewardship & CRM Hygiene A review queue that keeps business data clean

Compliance & Audit Checklists Recurring checks with evidence and an audit trail

QA & Training Datasets Curated, approved examples for training / eval / RAG

Multimodal Content Review Image / audio / video metadata reviewed first

Market Intelligence A human-reviewed research & monitoring feed

Content Factory Pipeline Idea → draft → approved asset, end to end

Dataset Labeling Pipeline Agent-first labeling with human review

Project Management & ERP Operational data changes behind an approval gate

Canonical System of Record The single source of approved, canonical data

Local Personal Knowledge Base A private, audited database for you and your AI tools

Verified Routine Work Recurring work that's done, reviewed, and recorded

Field Type Lab Every field type and review op in one local scenario

→ See all 16 use cases, with screenshots

Automation and ACP Agents

Busabase can become an event source for data workflows.

During review, a human can ask an ACP-compatible agent to improve the Change Request before it is merged:

clean fields

enrich missing metadata

normalize categories

rewrite a draft

generate summaries or tags

check policy, quality, or consistency

After merge, approved data can trigger downstream automation:

send a webhook

update an external system

notify a reviewer or channel

refresh a Next.js site

kick off an ETL or dataset export

call an external ACP Agent to continue the workflow

That makes Busabase more than a place to store data. It becomes a controlled handoff point between humans, applications, and agents.

Local Agents Operate Your Knowledge Base

Busabase is built to be driven by the agents already running on your own computer.

Because the API is local and trusted, you can point coding and automation agents — OpenClaw, Codex, Claude Code, Hermes, and similar local skills — directly at your Busabase instance. They can read approved knowledge, run skills against it, and propose changes back as Change Requests.

What a local agent can do with Busabase:

read your private, approved knowledge as grounded context

run a local skill that queries or summarizes your bases

propose new records or edits as reviewable Change Requests

enrich, clean, or label data without uncontrolled write access

wait for human approval before anything becomes trusted

The pattern is simple:

Local agent reads approved knowledge -> proposes a Change Request -> you review on your own machine -> approve -> merged into your local source of truth

This keeps the loop entirely on your hardware. The agent gets a real, structured memory to work against, and you keep authority over what becomes trusted — no private data has to leave your computer for any of it to work.

→ Step-by-step: Bring Your Own Agent — paste one prompt to onboard, then npx skills add busabase/skills to install a permanent skill your agent uses every session.

If OpenClaw is the revolution for agents on your local computer, then BusaBase is the revolution for the database and knowledge base on your local computer.

What Busabase Cares About

Busabase is not just asking "what is the latest value in this row?"

It also asks:

Who proposed this data?

Why should it change?

Which fields changed?

Is this a create, update, delete, or variant operation?

What did the AI Agent produce before it was accepted?

Who reviewed the agent output?

Did a human ask the agent to revise?

Was the proposal merged or rejected?

What automation ran after merge?

Can we trace the decision later?

That makes Busabase especially useful for AI Agent work. Agents can produce drafts, labels, summaries, reconciliations, or operational updates, but Busabase gives humans a preview layer before those outputs become trusted data.

How Busabase Compares

Busabase overlaps with familiar tools, but it's optimized for a different job: AI agents writing data, and humans approving it before it's trusted.

Tool Great for What Busabase adds

Airtable Flexible cloud tables for human teams Local-first ownership + an approval gate: agents propose, humans approve, with diff preview, history, and audit trails

APITable Open-source, API-first Airtable alternative API-first plus a review layer between proposal and trusted record

NocoDB Spreadsheet UI on top of your SQL database Every write is a reviewable Change Request, not a direct row edit

Baserow Self-hosted no-code database Change Requests, audit trails, and agent hooks

Notion Cloud docs, databases, and team knowledge A pure, local, structured knowledge base with a built-in review flow — no vendor cloud

Confluence / Lark Vendor-hosted team wikis Runs on your machine first; data never has to leave it

Obsidian Local-first Markdown notes for a human Local too — but a structured

[truncated for AI cost control]