Gobby

The control plane for AI coding tools.
One daemon. All your agents. No more context window roulette.

---

Gobby is a local-first daemon that unifies your AI coding assistants—Claude Code, Gemini CLI, Cursor, Windsurf, Copilot, and Codex—under one persistent, extensible platform. It handles the stuff these tools forget: sessions that survive restarts, context that carries across compactions, workflows that keep agents from going off the rails, and an MCP proxy that doesn't eat half your context window just loading tool definitions.

Gobby is built with Gobby. Most of this codebase was written by AI agents running through Gobby's own task system and workflows. Dogfooding isn't a buzzword here—it's the development process.

Note: Gobby is currently in alpha. Expect rough edges and breaking changes until the first stable release.

Why Gobby?

🎯 A Task System That Actually Works

If you've tried Beads or TaskMaster, you know the pain: databases that corrupt, agents that can't figure out the schema, worktrees that fall out of sync. Gobby's task system was designed by someone who got fed up with all of them.

• Dependency graphs that agents actually understand
• TDD expansion — describe a feature, get red/green/blue subtasks with test-first ordering
• Validation gates — tasks can't close without passing criteria (with git diff context)
• Git-native sync — .gobby/tasks.jsonl lives in your repo, works with worktrees
• Commit linking — [task-id] feat: thing auto-links commits to tasks

# Create a task
gobby tasks create "Add user authentication" --type feature

# Let the AI break it down with TDD ordering
gobby tasks expand <task-id>

# See what's ready to work on
gobby tasks list --ready

🔌 MCP Proxy Without the Token Tax

Connect 5 MCP servers and watch 50K+ tokens vanish before you write a single line of code. Gobby's proxy uses progressive disclosure—tools stay as lightweight metadata until you actually need them:

list_tools()           → Just names and descriptions (~200 tokens)
get_tool_schema(name)  → Full inputSchema when you need it
call_tool(name, args)  → Execute

Add servers dynamically. Import from GitHub repos. Search semantically. Your context window stays yours.

🔄 Session Handoffs That Don't Lose the Plot

When you /compact in Claude Code, Gobby captures what matters: the goal, what you changed, git status, recent tool calls. Next session, it injects that context automatically. No more "wait, what were we doing?"

Works across CLIs too. Start in Claude Code, pick up in Gemini. Gobby remembers.

🛤️ Workflows That Enforce Discipline

YAML-defined workflows with state machines, tool restrictions, and exit conditions:

# auto-task workflow: autonomous execution until task tree is complete
name: auto-task
steps:
  - name: work
    description: "Work on assigned task until complete"
    allowed_tools: all
    transitions:
      - to: complete
        when: "task_tree_complete(variables.session_task)"

  - name: complete
    description: "Task work finished - terminal step"

exit_condition: "task_tree_complete(variables.session_task)"

on_premature_stop:
  action: guide_continuation
  message: "Task has incomplete subtasks. Use suggest_next_task() and continue."

Built-in workflows: auto-task, plan-execute, test-driven. Or write your own.

🌳 Worktree Orchestration

Spawn agents in isolated git worktrees. Run tasks in parallel without stepping on each other. Gobby tracks which agent is where and what they're doing.

call_tool("gobby", "spawn_agent", {
    "prompt": "Implement OAuth flow",
    "task_id": "#123",
    "isolation": "worktree",
    "branch_name": "feature/oauth"
})

🔗 Claude Code Task Integration

Gobby transparently intercepts Claude Code's built-in task system (TaskCreate, TaskUpdate, etc.) and syncs operations to Gobby's persistent task store. Benefits:

• Tasks persist across sessions (unlike CC's session-scoped tasks)
• Commit linking — tasks auto-link to git commits
• Validation gates — define criteria for task completion
• LLM expansion — break complex tasks into subtasks

No configuration needed — just use Claude Code's native task tools and Gobby handles the rest.

📚 Skills System

Reusable instructions that teach agents how to perform specific tasks. Compatible with the Agent Skills specification and SkillPort.

• Core skills bundled with Gobby for tasks, sessions, memory, workflows
• Project skills in .gobby/skills/ for team-specific patterns
• Install from anywhere — GitHub repos, local paths, ZIP archives
• Search and discovery — TF-IDF and semantic search across your skill library

# Install a skill from GitHub
gobby skills install github:user/repo/skills/my-skill

# Search for relevant skills
gobby skills search "testing coverage"

Installation

Try it instantly

uvx gobby --help

Install globally

# With uv (recommended)
uv tool install gobby

# With pipx
pipx install gobby

# With pip
pip install gobby

Requirements: Python 3.13+

Quick Start

# Start the daemon
gobby start

# In your project directory
gobby init
gobby install  # Installs hooks for detected CLIs

Requirements: At least one AI CLI (Claude Code, Gemini CLI, or Codex CLI)

Works with your Claude, Gemini, or Codex subscriptions—or bring your own API keys. Local model support coming soon.

Configure Your AI CLI

Add Gobby as an MCP server. Choose the command and args that match your installation:

• pip/pipx install: "command": "gobby", "args": ["mcp-server"]
• uv tool install: "command": "uv", "args": ["run", "gobby", "mcp-server"]

Claude Code (.mcp.json or ~/.claude.json):

{
  "mcpServers": {
    "gobby": {
      "command": "gobby",
      "args": ["mcp-server"]
    }
  }
}

Or with uv:

{
  "mcpServers": {
    "gobby": {
      "command": "uv",
      "args": ["run", "gobby", "mcp-server"]
    }
  }
}

Gemini CLI (.gemini/settings.json):

{
  "mcpServers": {
    "gobby": {
      "command": "gobby",
      "args": ["mcp-server"]
    }
  }
}

Codex CLI (~/.codex/config.toml):

[mcp_servers.gobby]
command = "gobby"
args = ["mcp-server"]

Gemini Antigravity (~/.gemini/antigravity/mcp_config.json):

{
  "mcpServers": {
    "gobby": {
      "command": "/path/to/uv",
      "args": ["run", "--directory", "/path/to/gobby", "gobby", "mcp-server"],
      "disabled": false
    }
  }
}

CLI Support

CLI	Hooks	Status
Claude Code	✅ Full support	Native adapter, 12 hook types
Gemini CLI	✅ Full support	Native adapter, all hook types
Codex CLI	✅ Full support	Native adapter via app-server
Cursor	✅ Full support	Native adapter, 17 hook types
Windsurf	✅ Full support	Native adapter, 11 hook types
Copilot	✅ Full support	Native adapter, 6 hook types

Hook Installation

Gobby uses Python hook dispatchers that capture terminal context and communicate with the daemon. Run gobby install in your project to set up hooks:

gobby install           # Auto-detect and install hooks for all CLIs
gobby install --claude  # Install for specific CLI
gobby install --gemini
gobby install --codex
gobby install --cursor
gobby install --windsurf
gobby install --copilot

The dispatchers handle:

• Terminal context capture (TTY, parent PID, session IDs)
• Proper JSON serialization and HTTP communication
• Exit code handling for blocking actions

All CLIs can also connect via MCP for tool access (see configuration examples above).

How It Compares

	Gobby	TaskMaster	Beads	mcp-agent
Task dependencies	✅	✅	✅	❌
TDD expansion	✅	❌	❌	❌
Validation gates	✅	❌	❌	❌
Progressive MCP discovery	✅	Partial	❌	❌
Multi-CLI orchestration	✅	❌	❌	❌
Session handoffs	✅	❌	❌	❌
YAML workflows	✅	❌	❌	✅
Worktree orchestration	✅	❌	❌	❌
Zero external deps	✅	❌	✅	❌
Local-first	✅	✅	✅	✅

Architecture

AI CLI (Claude/Gemini/Codex)
        │ hooks fire
        ▼
   Hook Dispatcher
        │ HTTP POST
        ▼
  Gobby Daemon (:60887)
        │
   ┌────┴────┐
   ▼         ▼
FastAPI   FastMCP
   │         │
   ▼         ▼
┌─────────────────────┐
│  HookManager        │
│  SessionManager     │
│  WorkflowEngine     │
│  MCPClientProxy     │
│  TaskStore          │
│  MemoryStore        │
└─────────────────────┘
        │
        ▼
     SQLite
  (~/.gobby/gobby-hub.db)

Everything runs locally. No cloud. No API keys required (beyond what your AI CLIs need). Works offline.

MCP Tools

Gobby exposes tools via MCP that your AI coding assistant can use:

Task Management (gobby-tasks) create_task, expand_task, validate_task, close_task, claim_task, list_ready_tasks, suggest_next_task, link_commit, and more.

Session Management (gobby-sessions) get_current_session, pickup (restore context), get_handoff_context, list_sessions, send_message

Memory (gobby-memory) remember, recall, forget — persistent facts across sessions

Workflows (gobby-workflows) activate, advance, set_variable, get_status, end_workflow

Agents (gobby-agents) spawn_agent (unified API with isolation: current/worktree/clone), list_agents, get_agent, kill_agent

Worktrees (gobby-worktrees) create_worktree, list_worktrees, delete_worktree, merge_worktree

Clones (gobby-clones) create_clone, list_clones, delete_clone, merge_clone_to_target

Artifacts (gobby-artifacts) save_artifact, get_artifact, list_artifacts — capture and retrieve session artifacts

MCP Proxy list_mcp_servers, add_mcp_server, import_mcp_server, list_tools, get_tool_schema, call_tool, recommend_tools, search_tools

Skills (gobby-skills) list_skills, get_skill, search_skills, install_skill, remove_skill, update_skill

For AI Agents

If you're an AI agent reading this: Check out docs/AGENT_INSTRUCTIONS.md for your operating procedures, and CLAUDE.md for Claude Code-specific guidance.

Roadmap

See ROADMAP.md for the full plan, but highlights:

Shipped: Task system v2 (commit linking, validation gates, Claude Code interop), TDD expansion v2 (red/green/blue generation), workflow engine (state machines, tool restrictions, exit conditions), MCP proxy with progressive discovery, session tracking and handoffs, memory v3 (backend abstraction), hooks integration for all CLIs, unified agent spawning, worktree and clone orchestration, skills system, artifacts capture

Beta: Autonomous orchestration (conductor daemon, inter-agent messaging, token budget tracking, review gates)

Next: Security posture for MCP (allow/deny lists, audit logging), Observability (tool call tracing, session timelines), Web UI (read-only dashboard)

Vision: Always local first, but Pro cloud features to keep the lights on: Fleet management, Plugin ecosystem, Team workflows, Enterprise hardening

Changelog

See CHANGELOG.md for release history and detailed changes.

Development

uv sync                    # Install deps
uv run gobby start -v      # Run daemon (verbose)
uv run pytest              # Tests
uv run ruff check src/     # Lint
uv run mypy src/           # Type check

Coverage threshold: 80%. We're serious about it.

Contributing

We'd love your help. Gobby is built by developers who got frustrated with the state of AI coding tool orchestration. If that's you too, jump in:

• Found a bug? Open an issue
• Have a feature idea? Open a discussion first
• Want to contribute code? PRs welcome — check the roadmap for what's in flight
• UI/UX skills? We really need you. The maintainer is colorblind and Photoshop makes him itch.

See CONTRIBUTING.md for details.

License

Apache 2.0 — Use it, fork it, build on it.

---

_{Built with 🤖 by humans and AI, working together.}