🐝 TrembleCode

The self-hosted software house. Hand a PRD to a team of Claude Code agents that plans, builds, reviews and ships — on your machines, on your Claude subscription, with you in the loop only when it matters.

Why "TrembleCode"? When a forager honeybee finds more nectar than the hive can handle, it performs a tremble dance — the signal to recruit more workers. That's the idea: describe the work, and the right team shows up to do it.

What it does

1. New project → paste a PRD, or keep "start with discussion" and interview your way to one with a planning agent.
2. Generate a plan → the team lead explores the repo and proposes milestones, stories, tasks, and the MCP servers it wants.
3. Review & approve the plan in a 6-tab editor; tweak tasks inline.
4. Watch it build. A sandboxed team — team lead, backend/frontend devs, QA — claims tasks, writes code, reviews, and merges through a serialized queue. You only answer Inbox items (questions, destructive ops, milestone gates).
5. See everything live: the board, the comms feed, each agent's terminal, the project wiki, dev-server previews, and per-agent/day/model cost tracking.

Screenshots

PRD discussion — interview a planning agent until you have a spec.	Plan review — milestones, stories, tasks, MCP, and a critical-path Gantt.
Task board — the team claims and moves tasks through the queue.	Comms feed — the agent-to-agent message bus, in real time.
Human-in-the-loop — milestone gates and escalations land in your Inbox.	Project wiki — the agent-maintained, git-versioned knowledge base.
Preview — live dev-server previews per agent, in the dashboard.

Why TrembleCode

• Self-hosted and glass-box. Your code never leaves your infrastructure, and you can watch every agent's terminal in real time. Not a cloud black box.
• Your Claude subscription, no markup. Bring your own Claude login (or an Anthropic API key). You pay Anthropic's price for inference — there's no per-seat AI tax on top.
• A whole SDLC, not a single coding agent. PRD → plan → a role-based team → review → QA → merge queue → project-wiki memory — coordinated, not one bot in a loop.
• Human-in-the-loop only on hot topics. Conservative escalation defaults send destructive ops and ambiguous calls to your Inbox; everything else is autonomous.

Quickstart

Prerequisites: Docker Desktop, Node 22 + pnpm, Python 3.12 + uv, and a Claude subscription (recommended) or Anthropic API key.

docker compose up -d redis      # infra
make image                      # build the sandbox image (once)
make image-flutter              # optional, for Flutter projects

cd server && uv sync && cd ..
cd web && pnpm install && cd ..

make dev                        # server :8400 + web :3000

First run only: create a project, start it, open the team-lead terminal in the dashboard (team card → RW toggle) and run /login. The login persists in ~/.tremblecode/agent-home across all projects and restarts. Prefer API-key mode? Set it in Settings instead.

⚠️ There's no built-in authentication yet — keep the server on localhost or behind a VPN/reverse proxy. See SECURITY.md.

Architecture

• Comms: a message bus (SQLite truth + Redis streams with consumer groups and explicit ACKs). Agents get a one-line tmux notification and pull bodies via MCP check_messages / confirm with ack_message. Humans ride the same bus.
• Workflow: PENDING → ASSIGNED → IN_PROGRESS → IN_REVIEW → APPROVED → MERGING → DONE, with CAS claims, dependency gating, least-loaded QA routing, and a server-serialized merge queue (the lead is the only merger).
• Memory: each project keeps an agent-maintained markdown wiki (Karpathy's like) (repo/.wiki/ — index/log/pages), git-versioned; agents ingest after every task, query index-first, and lint at every milestone.
• Context policy: /clear after every task; standing context (task brief, roster, wiki pointer, pending messages) is re-injected on every SessionStart hook, so context loss is always safe.
• Sandbox: project dirs are identity-mounted (same absolute path inside the container) so git worktrees stay valid; a 10-port block per project exposes dev servers to the host and the dashboard preview tab.

Layout

server/    FastAPI orchestrator (API, docker/git provisioning, merge queue,
           watchdog, terminal bridge, costs)
web/       Next.js dashboard (TUI theme, plan review, board, comms, inbox,
           web terminals, wiki browser, preview)
sandbox/   container image + in-container runtime (relay, MCP server, hooks)
templates/ role prompts, CLAUDE.md/identity/hooks/mcp templates, wiki
           skeleton, MCP catalog

Tests

make test          # server suite (workflow, provisioning, costs, mcp, wiki)
cd web && pnpm build

Sustainability

TrembleCode is AGPL-3.0 and free to self-host, forever — individuals and single teams will never hit a paywall. To fund development, we plan to eventually offer a managed cloud and enterprise support for organizational needs (SSO, multi-user, SLAs). There's nothing to buy today.

The AGPL covers TrembleCode itself, not your work: code your agents write for you is yours, and running TrembleCode as a tool doesn't impose AGPL obligations on the projects it builds.

Community & contributing

• 💬 Discussions — questions, setups, ideas
• 🗺️ Roadmap
• 🤝 Contributing guide · Code of Conduct
• 🔒 Security policy

License

GNU AGPL-3.0-only © TrembleCode contributors.