Your AI dev team. Plugs into your tools. Ships code while you sleep.
Floor Agents gives any software team an autonomous AI engineering team. Connect your GitHub, your Linear, your preferred AI models — and get a team of specialized agents that write code, review PRs, and iterate on feedback.
Linear issue (labeled "agent")
→ Backend Dev (Gemma, local) writes code
→ CTO (Claude Code Opus) reviews the PR
→ If approved → done
→ If changes requested → dev revises, CTO reviews again
- Create an issue in Linear with the
agentlabel - The orchestrator picks it up and dispatches to the right agent
- The dev agent reads the codebase, writes code, creates a PR
- The CTO agent reviews the PR — approves or requests changes
- The issue is updated with the PR link
No proprietary UI. All work happens in your Linear and GitHub.
Each agent is a different AI model chosen for its strengths:
| Agent | Default Provider | Role |
|---|---|---|
| Backend Dev | LM Studio (Gemma) | Writes code, creates PRs |
| Frontend Dev | LM Studio (Gemma) | Writes code, creates PRs |
| CTO | Claude Code (Opus 4.6) | Reviews PRs, approves or rejects |
| PM | Anthropic API (Sonnet) | Decomposes tasks |
| QA | Anthropic API (Sonnet) | Writes tests |
Every agent is configurable — swap models, providers, or add custom agents. Use local models for free coding, cloud models for quality-critical reviews.
# Install
bun install
# Configure
cp .env.example .env
# Edit .env with your GitHub token, Linear key, etc.# Start LM Studio with Gemma loaded, then:
bun run src/main.tsSee Getting Started for the full setup guide.
A monorepo with 9 packages, each with a single responsibility:
packages/
├── core/ Types, config loader, validation
├── anthropic/ Claude API adapter
├── claude-code/ Claude Code CLI adapter (native worktree execution)
├── lmstudio/ LM Studio adapter (local models)
├── openai/ OpenAI-compatible adapter (OpenAI, Ollama, etc.)
├── gemini/ Google Gemini adapter
├── github/ GitHub REST API (branches, commits, PRs)
├── task/ Task managers (Linear, Things 3, GitHub Issues)
├── context-builder/ File selection (v2 import tracing), prompt rendering, token budgets
└── orchestrator/ Pipeline, native runner, review, guardrails, cost tracking
- Two execution modes — API agents (LM Studio, Gemini) use tool-use calls; Claude Code agents work directly on git worktrees with full file access
- Vendor-agnostic AI — customers choose their LLM providers per agent. Mix local and cloud models in the same team
- Config-driven — a single YAML file defines the entire team: agents, models, guardrails, cost limits
- Crash-recoverable — 10-step execution state machine with file-based persistence. Restart safely at any point
- Multi-agent review loop — dev writes code, CTO reviews, dev revises (up to 3 cycles)
Everything is in config/templates/default.yaml:
agents:
- id: backendllm:
provider: lmstudio # or: anthropic, claude-code, openai, ollamamodel: google/gemma-4-e2b # any model the provider supportscapabilities: [read_code, write_code, create_pr, write_tests]
- id: ctollm:
provider: claude-codemodel: opuscapabilities: [read_code, review_pr, approve, reject]guardrails:
maxFilesPerTask: 20blockedPaths: [".env*", "*.pem", ".github/workflows/*"]costs:
maxCostPerTask: 5.00maxCostPerDay: 50.00See Configuration Reference for the full spec.
| Provider | Package | Use Case |
|---|---|---|
| Claude Code | @floor-agents/claude-code | Full codebase access. Best for dev + reviews. Max plan. |
| LM Studio | @floor-agents/lmstudio | Local models (Gemma, Llama, Qwen). Free. |
| Gemini | @floor-agents/gemini | Google Gemini 2.5 Pro/Flash. |
| Anthropic | @floor-agents/anthropic | Claude API (direct). Pay-per-token. |
| OpenAI | @floor-agents/openai | GPT-4o, o3. Also works with Ollama, Together, Groq. |
| Provider | Status |
|---|---|
| Linear | Supported |
| Things 3 (macOS) | Supported |
| GitHub Issues | Supported |
| Jira | Planned |
| Provider | Status |
|---|---|
| GitHub | Supported |
| GitLab | Planned |
| Bitbucket | Planned |
- Guardrails — file count/size limits, blocked paths (
.env,.pem, CI configs), blocked extensions, path traversal detection - Cost controls — per-task and per-day spending limits. Local models report $0
- Review loop — CTO agent reviews every PR before it's marked ready. Max 3 revision cycles, then
needs-human - Crash recovery — execution state persisted to disk between every step. Idempotent operations (branch creation, PR creation)
bun test# 116 tests across 22 files
bun run typecheck # type check all packages
bun run src/main.ts # start the orchestrator- Getting Started
- Configuration
- First Run Guide
- Architecture
- Package Docs
- Adding an LLM Provider
- Adding a Task Manager
- Testing
- Runtime: Bun
- Language: TypeScript (strict)
- Structure: Monorepo with Bun workspaces
- Config: YAML
- State: File-based JSON (Phase 1), PostgreSQL (Phase 3)
- Tests: bun:test
Proprietary. Floor IO SA.