openfugu

An experimental project that makes heavy use of LLM-generated code. It was inspired by Sakana AI's Fugu Ultra and the related papers. The official Sakana subscription is quite expensive, so I wanted to build an open version that you can use as long as you have subscriptions to Claude Code, Codex, and Antigravity (you need all three)—and that's why I'm sharing it here. Issues and pull requests are very welcome!

Boundaries

• openfugu is architecture-inspired by Sakana AI's Fugu Ultra and related research papers; it is not a reproduction of trained coordinators, private routers, or benchmark claims from those papers.
• openfugu is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Google, or Sakana AI.
• openfugu does not call vendor model APIs directly.
• openfugu does not read, store, or reuse API keys, OAuth tokens, cookies, or keychain secrets.
• The default policy is subscription-only. API-key, unauthenticated, and unknown auth are rejected unless a future explicit policy says otherwise.
• Vendor CLI use follows the user's plan, limits, organization policy, and provider-side overage settings. Additional provider-side charges can still happen when the provider account allows them.
• Git worktrees isolate candidate changes, but they are not a security sandbox.

Build

zig build
zig build test
zig fmt --check .

Zig 0.16.0 is required.

Install

zig build install -p ~/.local

Make sure ~/.local/bin is on PATH, then run:

openfugu --help
openfugu doctor

Coding-agent skill

skills/openfugu/SKILL.md is a bundled skill for coding agents (opencode, Claude Code, Codex). Install it into your agent's skills directory (e.g. ~/.agents/skills/openfugu/SKILL.md) and the agent can route tasks through openfugu using the recommended doctor -> route -> run flow. The skill documents the subcommands, planners, exit codes, and task-kind vocabulary so the agent never has to read --help to discover the interface.

MCP server mode

openfugu also speaks MCP (Model Context Protocol) over stdio. Coding agents that support local MCP servers can call openfugu tools without spawning the CLI directly:

// opencode.jsonc / opencode.json
"mcp": {
  "openfugu": {
    "type": "local",
    "command": ["/home/USER/.local/bin/openfugu", "mcp"],
    "enabled": true
  }
}

Exposed tools: route, run, submit, poll, wait, doctor, list_agents, status. The server uses the same code path as the CLI, so tool calls and CLI calls behave identically.

Real CLI smoke checks are opt-in because they can consume provider quota:

zig build smoke -Dreal-cli-tests=true

Model-invoking task smoke checks require a second opt-in. This command asks the provider CLIs to run a real task and can consume subscription quota:

zig build smoke -Dreal-cli-tests=true -Dreal-cli-task-tests=true

The normal test suite uses fake agents and temporary git repositories only.

Task execution uses the subscription-agent planner by default. It asks an available subscription CLI to route the task first, then falls back to local scoring if the router output is invalid. Pass --planner heuristic to skip the subscription router, or --planner capability to use a deterministic capability-aware planner that gates agents on edit_files / run_commands / structured_output. Add --explain-routing to print the route, preferred agent, score, and selected agent. Local scoring also uses prior accepted/failed outcomes from the local ledger.

Usage

openfugu --help
openfugu

# Inspect routing without running (no quota consumed)
openfugu route "your task"

# Route and execute
openfugu run "your task"

# Fire a task and return immediately; poll or wait for the result
openfugu submit "your task"
openfugu poll <job-id>
openfugu wait <job-id>
openfugu jobs

# Dry run (candidate worktree only, no apply)
openfugu run --no-apply "your task"

# Setup and dependency diagnostics
openfugu doctor

# Agent and ledger summary
openfugu status
openfugu list-agents

# View routing decisions and agent selection details
openfugu --explain-routing "your task"

Both --flag value and --flag=value are accepted. The explicit subcommands route and run are the recommended entry points for coding agents because they avoid ambiguity about what is a flag and what is the task text. The positional form (openfugu "task") still works for backward compatibility.

NDJSON streaming

openfugu run --json "task" and openfugu route --json "task" emit line-delimited JSON events so a parent process can consume routing and execution progress programmatically:

{"event":"route_start","task":"fix the flaky test"}
{"event":"route_decision","router":"heuristic","route":"test_fix","selected":"claude","score":125}
{"event":"agent_start","agent":"claude"}
{"event":"agent_done","agent":"claude","accepted":true,"applied":true,"reverified":true}
{"event":"result","code":0,"agent":"claude","accepted":true,"applied":true,"reverified":true}

Event types: route_start, route_decision, agent_start, agent_done, result, error.

Use the --no-apply flag to perform dry runs.

Running openfugu without arguments starts the ZigZag-based interactive TUI. It uses raw-key input on an interactive terminal, so cursor movement and common line-editing shortcuts work inside the fullscreen dashboard. Type a task to route and execute it, press Tab to accept command suggestions, use Up/Down for input history, press PageUp/PageDown to page output, use Home/End to jump to the output top or bottom, or type :quit to exit. Input history is stored in .openfugu/tui-history. The TUI starts in agent=auto, mode=auto, planner=subscription-agent; use :planner heuristic only when you want local scoring without the subscription router. The TUI also accepts :status, :doctor, :agents, :usage, :ledger, :where, :pwd, :worktrees, :git, :changed, :remote, :branch, :branches, :tags, :describe, :log, :diff, :staged, :staged-patch, :patch, :ci, :watch-ci, :pr, :pr <number-or-url>, :pr-checkout <number-or-url>, :issues, :issue <number-or-url>, :verify, :build, :test, :fmt, :check, :fetch, :pull, :push, :push-force-with-lease, :stash, :stashes, :stash-show, :stash-patch, :stash-pop, :cancel, :rerun, :save <file>, :stage <path>, :unstage <path>, :commit <message>, :commit-amend <message>, :commit-amend-no-edit, :switch <branch>, :new-branch <branch>, :show <rev>, :run <command>, :rg <pattern>, :todo, :ls [path], :files [path], :cd <path>, :cwd <path>, :load <file>, :open <file[:line]> with line numbers, :head <file>, :tail <file>, :dry-run, :no-apply, :apply, :agent <name>, :mode <name>, :planner <name>, :reset-routing, :plan [task], :route [task], :last, :replay <run-id>, :clear, :history, and :clear-history. Agent names accept claude, claudecode, claude-code, codex, agy, and antigravity.