Busabase
Open-source database & workspace for AI agents: structured data, durable knowledge, reusable skills, runnable apps, and human-reviewed changes — local-first and self-hostable.
The trust layer (Change Request → diff → review → merge → audit) is genuinely exercised in integration tests including autoMerge and restore paths, so user_confirmation earns 2. Least privilege, data-flow transparency, sensitive data handling, dependency security, and external effects (webhooks, Cloud Connect tunnel, remote Sandock engine) rest mostly on README claims; implementation details of permissions/audit are not in the supplied files, so 1 each. Rollback evidence is limited to a single view-restore test. Attribution is a valid MIT copyright line but the publisher is unverified; the deduction is that many safety claims cannot be statically confirmed.
Tests are coherent with the README: a full PGlite CR→review→merge flow, plus CI lint/typecheck/test/e2e/docker jobs that deliberately need no external services (2 each). Failure messages are only asserted ('local-start check reports missing dependencies') without code shown — deducted to 1.
Audience and scenarios are well documented (six use-case classes, four connection methods, 2), and environment fit is strong across npm/Docker/Desktop/source with swappable Postgres/S3 (2). Capability boundaries and trigger precision lack concrete documentation of agent permission limits, prohibited actions, or input validation — each deducted to 1.
Information architecture is clear (monorepo layout, node-type doc pointers, architecture diagram, 2); install notes cover npx/npm/Docker/Desktop/source (2); naming is stable across packages. Examples/FAQ and known-limitations point to files outside the evidence set or a single Security caveat (1 each); version 0.52.1 exists but no CHANGELOG is shown (1); maintenance responsibility is only community links plus a bot account (1). MIT license is complete and consistent — 3.
Output usability is supported by typed APIs, three integration channels (MCP/OpenAPI/CLI), and a tested CR preview route with security headers (2). Marginal value is argued concretely via the category comparison table (2). Cost/benefit is favorable on paper: embedded storage, no login, external-service-free CI (2). The deduction is that runtime costs and benefits were not executed-verified.
Claims are traceable: the CR→review→merge loop is asserted step-by-step and CI workflow comments are detailed (2). Cross-source corroboration is weak — everything is self-description and self-tests from one repo, and the coverage badge is a self-committed artifact (1). Fact/inference separation is adequate: external products are linked rather than asserted as fact (2).
- This is a static, low-confidence review: the Change Request approval, audit, and permission machinery is validated only by the repo's own tests, not by independent execution.
- The README warns against exposing write endpoints publicly, yet the local server is login-free by default; deploy on a trusted network with your own reverse proxy and credentials.
- No dependency-audit or lockfile vulnerability scanning is evidenced, and local patches to third-party packages (@scelar/nodepod etc.) exist — review them before adopting.
- Publisher identity is unverified; evaluate exfiltration risk before enabling Cloud Connect tunnels or webhooks.
- Examples, known limitations, and maintenance commitments rely heavily on files outside the reviewed evidence; verify the referenced docs and commit history before adoption.
What does this agent do, and when should you use it?
Busabase is an MIT-licensed open-source workspace that gives agents like Claude Code, Codex, and Cursor one place for structured databases, durable knowledge, reusable skills, runnable AirApps, and a human-review trust layer. Every building block is a first-class node in the same workspace, addressable by humans through the UI and by agents through Agent Skills, MCP, OpenAPI, or the CLI. Agent writes arrive as Change Requests with field-level diffs, comments, and approvals, merging into canonical data only after review. The local edition runs with an embedded PGlite database and local file storage, no login required; a hosted multi-user Busabase Cloud is also available. It fits teams that want agents to persist useful work while keeping humans in control of what becomes canonical.
After npx busabase server, open http://localhost:15419/dashboard/local. The workspace contains Bases (typed records, fields, relations, views, forms), Docs (versioned Markdown knowledge), Files/Drives, Skills (reusable instructions and scripts), AirApps (runnable apps on workspace data), Whiteboards, Workflows, and an Inbox/Activity stream. Agents connect four ways: following the Agent Skill at http://localhost:15419/SETUP_SKILL.md, calling the MCP endpoint, consuming the OpenAPI spec at http://localhost:15419/api/v1/doc, or using busabase-cli. The agent reads workspace context, proposes data/doc/skill/app changes, a Change Request shows the exact diff and source, and a human approves, requests changes, or rejects; merged work becomes canonical knowledge with a complete audit trail.
- A team has agents collect notes, sources, and decisions into a durable knowledge base future agents can reuse
- Sales/research: agents enrich companies and deduplicate CRM records, producing verified intelligence after human review
- Content teams draft posts, docs, and pages via agents with an editorial approval trail, acting as a headless CMS
- Product operations teams maintain projects, tasks, vendors, and policies in an auditable operational database
- Data teams label examples, attach evidence, and score quality for human-reviewed training/evaluation datasets
- Internal tooling: agents build AirApps and workflows over workspace data that inherit the same source of truth
What are this agent's strengths and limitations?
- Writes go through a Change Request loop with field-level diffs and audit trails, making agent output reviewable and traceable
- Database, knowledge, skills, and apps live in one workspace exposed via MCP, OpenAPI, CLI, and Agent Skills — model- and provider-agnostic
- Local edition embeds PGlite with no external database or login, while supporting external Postgres, S3 storage, and Docker self-hosting
- The open-source server is designed for a trusted local machine or private network; public exposure requires your own auth and reverse proxy setup
- Only one process can hold the same PGlite database at a time, limiting local concurrency
- Collaboration, roles, and permissions require the hosted Busabase Cloud or a Cloud Connect tunnel rather than the free local edition alone
- No built-in model — you must supply and configure your own agent (Claude Code, Codex, etc.) to get value
How do you install or deploy this agent?
Three options:
- Run now:
npx busabase server, then open http://localhost:15419/dashboard/local; or install once withnpm i -g busabaseand runbusabase server. - Docker:
docker run --rm -p 15419:15419 -v ~/.busabase/data:/data busabase/busabase(also published to ghcr.io/busabase/busabase). - Desktop: download the native macOS/Windows/Linux app at busabase.com/download.
From source: pnpm install, cp apps/busabase/.env.example apps/busabase/.env, pnpm --filter busabase dev. Data defaults to ~/.busabase/data/ (pgdata/ and storage/); use BUSABASE_DATA_DIR, PG_DATABASE_URL, or STORAGE_URL for external Postgres or S3-compatible storage.
How do you use this agent?
Start the server and use the dashboard UI directly. To connect an agent, have it read the Agent Skill at http://localhost:15419/SETUP_SKILL.md, or open 'Agent Skills' in the sidebar for the current MCP endpoint and OpenAPI spec. API client: npx busabase-cli --help. Agent writes land in the Inbox as Change Requests; review diffs, approve or reject in the web UI or mobile app. Machine-readable API docs: http://localhost:15419/api/v1/doc.
How does this agent compare with similar options?
The README contrasts Busabase with human databases (Airtable, Baserow, NocoDB), human knowledge tools (Notion, Confluence, Obsidian), plain databases (Postgres), and agent runtimes/chat tools: those lack agent context, reusable skills/apps, and a native proposal boundary, whereas Busabase treats the agent as the primary operator with a built-in propose-review-merge loop. It positions itself as a new category between these.
FAQ
Does it cost money?
Does it include an AI model?
Where is my data stored?
~/.busabase/data/; you can switch to external Postgres or S3-compatible storage via environment variables. Personal Desktop works fully offline.