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, declarative rules 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 — over 10,000 tasks tracked and counting. Case in point: the entire OpenTelemetry observability stack (tracing, metrics, logging bridge, trace viewer UI) was built autonomously — 10 tasks dispatched across Gemini devs and Claude Opus reviewers, orchestrated by a cron-driven pipeline, completed in ~3 hours with six infrastructure bugs discovered and fixed live. See test-battery.md for the full story.

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

Your AI assistant creates, expands, and closes tasks through MCP tools — no CLI needed:

create_task(title="Add user authentication", task_type="feature")
expand_task(task_id="#42")           → TDD-ordered subtasks
list_ready_tasks()                   → What's unblocked and 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 discovery—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.

🛤️ Rules That Enforce Discipline

Declarative rules that enforce behavior without relying on prompt compliance. The LLM doesn't need to remember constraints—the rule engine evaluates every event and enforces behavior through tool blocks, context injection, and state mutations:

# Block git push - let the parent session handle pushing
no-push:
  event: before_tool
  effect:
    type: block
    tools: [Bash]
    command_pattern: "git\\s+push"
    reason: "Do not push to remote. Let the parent session handle pushing."

# Block file edits without a claimed task
require-task:
  event: before_tool
  when: "not task_claimed and not plan_mode"
  effect:
    type: block
    tools: [Edit, Write, NotebookEdit]
    reason: "Claim a task before editing files."

13 bundled rule groups covering safety, tool hygiene, task enforcement, stop gates, memory lifecycle, and more. Plus on-demand step-based workflows and deterministic pipelines.

🌳 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-agents", "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 instruction sets that teach agents how to perform specific tasks. Skills follow the SKILL.md format and are managed through the database.

• Core skills bundled with Gobby — synced to the database on daemon startup
• Project skills in .gobby/skills/ for team-specific patterns
• Auto-injection — skills with alwaysApply: true inject into every session
• Search and discovery — your AI assistant finds and uses skills through MCP tools

🌐 Web UI

Gobby ships a built-in web interface that auto-starts with the daemon:

• Chat with MCP tool support, voice chat, model switching, slash commands
• Tasks — kanban board, tree view, dependency graph, Gantt chart, detail panel
• Memory — table view, Neo4j 3D knowledge graph
• Sessions — lineage tree, transcript viewer, AI summary generation
• Cron Jobs, Configuration, Skills, Projects, Agent Registry pages
• File browser/editor, terminal panel with xterm.js

Access at http://localhost:60887 when the daemon is running.

🔍 Observability (OpenTelemetry)

Full observability built on OpenTelemetry — no custom metrics frameworks, no vendor lock-in:

• Tracing — @traced decorator, span context propagation, SQLite span storage
• Metrics — instruments for MCP calls, pipeline executions, task lifecycle, hook events
• Logging — OTel logging bridge replaces custom logging
• Exporters — OTLP gRPC, Prometheus
• Trace viewer — built-in UI with waterfall visualization and span detail panel

🧬 Code Indexing

AST-based symbol indexing via the gobby-code MCP server. Search and retrieve code by symbol instead of reading entire files — saves 90%+ tokens on large codebases:

search_symbols("TaskExpander")  → Find symbols by name
get_file_outline("src/foo.py")  → Hierarchical symbol map
get_symbol(symbol_id)           → Just the source you need

Tree-sitter parsing for 15+ languages. Auto-indexes on commit, on init, and on session start.

🚀 Pipelines & Orchestration

Deterministic automation with approval gates — from simple scripts to autonomous multi-agent orchestration:

• Step types: exec, prompt, invoke_pipeline, spawn_session
• Tick-based orchestrator pipeline with cron scheduling
• Clone-based agent isolation — one clone per epic, sequential or parallel dispatch
• Provider fallback rotation — auto-retry across providers on failures
• Approval gates for human-in-the-loop workflows
• Condition evaluation with safe expression engine
• CLI, MCP, and HTTP API access

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

Quick Start

gobby start              # Start the daemon
gobby init               # In your project directory
gobby install            # Installs hooks for detected CLIs

Then add Gobby as an MCP server in your AI CLI (see below) and start coding. Gobby handles sessions, tasks, rules, and context automatically.

Requirements: Python 3.13+ and Claude Code. Additional CLIs (Gemini CLI, Codex CLI, Cursor, Windsurf, Copilot) are supported but Claude Code is the primary development target. Works with your existing subscriptions — or bring your own API keys.

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

* Codex hook limitation: Codex CLI only supports fire-and-forget notify hooks for interactive terminal sessions. Bidirectional hook enforcement (tool blocking, context injection, workflow enforcement) is not possible because Codex lacks a blocking hook protocol — see openai/codex#2109. Gobby includes a Codex app-server adapter that provides full bidirectional control via JSON-RPC, but this runs Codex as a daemon-controlled subprocess rather than an interactive terminal session. Codex agents spawned via pipelines (--full-auto) work fully but bypass hook enforcement. All CLIs connect via MCP for tool access regardless of hook support.

Hook Installation

Run gobby install in your project to auto-detect and set up hooks for all supported CLIs. Hooks handle terminal context capture, session tracking, and rule enforcement. All CLIs also connect via MCP for tool access.

How It Compares

Architecture

AI CLI (Claude/Gemini/Cursor/Windsurf/Copilot)
        │ hooks fire
        ▼
   Hook Dispatcher
        │ HTTP POST
        ▼
  Gobby Daemon (:60887)
        │
   ┌────┼────────┐
   ▼    ▼        ▼
FastAPI WebSocket FastMCP
   │    │         │
   ▼    ▼         ▼
┌──────────────────────┐
│  RuleEngine          │
│  HookManager         │
│  SessionManager      │
│  AgentRunner         │
│  WorkflowEngine      │
│  PipelineExecutor    │
│  MCPClientProxy      │
│  TaskStore           │
│  MemoryStore         │
│  WebUI               │
└──────────────────────┘
        │
        ▼
     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.

Orchestration (gobby-orchestration)
orchestrate_ready_tasks, wait_for_task, poll_agent_status, spawn_review_agent, approve_and_cleanup, 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 and Rules (gobby-workflows)
list_rules, toggle_rule, get_rule_detail, activate_workflow, set_variable, get_variable, get_workflow_status

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

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

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

Pipelines (gobby-pipelines)
list_pipelines, run_pipeline, approve_pipeline, reject_pipeline, get_pipeline_status

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 CLAUDE.md for operating guidance.

Authentication & AI Vendor Policies

Gobby uses the Claude Agent SDK for web chat and workflow agents (subscription mode by default, intended for individual use). For commercial/production use, configure API keys per your AI vendor's terms. Recent Gemini CLI TOS changes prohibit using Gemini CLI OAuth tokens in third-party software.

See AUTH.md for details on authentication modes and AI vendor policies.

Roadmap

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

Shipped: Task system v2, TDD expansion, rule engine (13 bundled rule groups), MCP proxy with progressive discovery, session handoffs, memory v5 (Qdrant + knowledge graph), hooks for all 6 CLIs, orchestration v3 (tick-based pipeline, clone isolation, provider fallback rotation, QA-dev agent), OpenTelemetry observability (tracing, metrics, logging, trace viewer UI), native AST code indexing, autonomous SDK agent execution, session handoff & digest overhaul, stop-gate enforcement, pipeline system with approval gates, web UI (tasks, memory, sessions, chat with voice, cron, config, skills, projects, agents, file browser, traces), skills system, worktree/clone orchestration

In progress: v1 release prep — bug fixing, orchestration battle-hardening, UI polish, documentation

After v1: Ollama support

Future: Pro cloud features, fleet management, plugin ecosystem v2

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

Using Gobby in other projects (from source)

If you're running Gobby from a source checkout, use -C to target another project directory:

uv run --project ~/Projects/gobby gobby init -C /path/to/other/project
uv run --project ~/Projects/gobby gobby install -C /path/to/other/project

The --project flag tells uv to use the Gobby installation from your source repo, and -C tells Gobby which directory to operate on.

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.}