- Category
- AI Agents
- Rank
- No. 3180Tools index
Previous survey · No. 2989 ·
- Listed in
- #17 Interface with your agents
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Coding
- Interfaces
- CLI · Web · Desktop · API
- GitHub
- 5.7k stars
- Latest release
- desktop-nightly
- Date
About
An open-source agent IDE that can rewrite and automate itself. Work runs in live threads you can watch, steer mid-flight or hand to another agent, and the desktop app, web app, CLI and HTTP API are all first-class ways to drive the same engine, so scripts, cron jobs and bots can start work the way a person does. It orchestrates coding agents you already pay for — Claude Code, Codex, Cursor and anything speaking ACP — on your own provider subscriptions rather than reselling inference.
What it does
bb is a self-hosted workspace for running coding agents. A central server keeps every project, conversation and event in a local SQLite database and pushes changes to clients over WebSocket. A small daemon on each enrolled machine prepares the working directory, often a fresh git worktree on its own branch, launches the agent process, and reports progress back. Each supported agent (Claude Code, Codex, Pi, and ACP agents such as Cursor or Amp) is a plugin that translates its native output into one shared event format. You sign in to those agents yourself; bb drives the CLIs already authenticated on the machine.
Why it's ranked here
The architecture is unusually explicit for a young agent workspace. MIT licensed, with written contracts between server, daemon and clients, a provider plugin interface whose code samples are checked by a test against the real types, and a stated rule that first-party agents use only the public plugin API. CI runs on Ubuntu, macOS and native Windows. The README itself says workflows and surfaces are still evolving, so this is a strong foundation rather than a settled product.
What's good
Worktree handling is specified in detail: a gitignore-style include list copies local secrets into each new worktree, a setup script runs with a 15-minute timeout, a teardown script releases outside resources, and archived work gets a five-minute grace period before cleanup. Ownership between parent and child agent runs is fixed at creation, which rules out cycles. Settings for the Claude Code plugin expose real switches: memory, native subagents, sandboxing and the 1M-token context. Telemetry is documented field by field and one environment variable turns it off.
Tradeoffs
The server API has no authentication and permits command execution and file reads, so the remote-access modes are only safe behind a trusted network such as Tailscale. The desktop app is stable only on Apple Silicon Macs; Linux and Windows builds are alpha. Native add-ons mean npm 12 and later need an explicit allow-scripts flag or startup fails. Production runs send anonymous usage telemetry by default. The phone client is iOS only, with no Android build tested, and in-app updates never roll back.
How to use it well
Fit: a developer or small team already paying for Claude Code or Codex who wants several agents working in parallel, each isolated on its own branch, and wants to script that from a CLI or TypeScript SDK. Commit setup and teardown scripts to the repository before relying on worktrees. Keep it on localhost or behind Tailscale. It does not provide models or inference itself, and provider sign-in still needs a terminal on the host.
Technical notes+
README.md and docs/system-overview.md describe the runtime as server, host daemon, app and CLI, with @bb/server-contract and @bb/host-daemon-contract as the only cross-boundary imports. docs/provider-plugin-api.md defines registration via bb.providers.register, a per-provider bridge exported as experimental_providerBridge, typed provider/recovery kinds (sessionArchived, authRequired, rateLimited and others) instead of error-text matching, and a model/list cache refreshed after 10 minutes. plugins/provider-claude-code/server.ts registers reasoning levels up to max, a fast service tier, and passes CLAUDE_CODE_OAUTH_TOKEN through. apps/server/src/server.ts is a Hono app with @hono/node-ws WebSockets. docs/environment-provisioning.md documents the creating, provisioning, ready, error status walk; docs/worktrees.md the hook contracts. package.json pins Node >=22.19.0 and pnpm 9.15.0 with Turbo and Vitest.
Observed
- License
- MIT (package.json)
- Language
- TypeScript monorepo, pnpm and Turbo (package.json)
- Runtime
- Node.js 22.19 or newer; tested on 22, 24 LTS and 26 (docs/platform-support.md)
- Install
- npx bb-app, or desktop app download (README.md)
- Interfaces
- Desktop app, web UI, bb CLI, HTTP and WebSocket API, TypeScript SDK (README.md, docs/repository-overview.md)
- Platforms
- Desktop app stable on macOS arm64; Linux x64 and Windows x64 alpha; WSL2 supported (README.md, docs/platform-support.md)
- Supported agents
- Claude Code, Codex, Pi, ACP agents (docs/repository-overview.md)
- Storage
- SQLite on the server (docs/system-overview.md)
- Server API auth
- Unauthenticated; remote binding documented as trusted-network only (README.md)
- Telemetry
- Anonymous usage telemetry on production runs, opt out with BB_TELEMETRY=false (README.md)
- Mobile
- iOS client; Android not built or tested (docs/platform-support.md)
Read from README.md, package.json, docs/VISION.md, docs/system-overview.md, docs/repository-overview.md, docs/provider-plugin-api.md, docs/platform-support.md, docs/environment-provisioning.md, docs/worktrees.md, plugins/provider-claude-code/server.ts, apps/server/src/server.ts, docs/configuration.md.
Tags
Tech Stack
Media

Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.