ahub 3.0.0

Русская версия

Let Claude Code hand work to cheap models — and only spend its own tokens on decisions.

agent-hub is a local service between an orchestrator (Claude Code, or any CLI agent) and worker models (opencode: Muse Spark, free models; Gemini via Antigravity). Claude files a task, goes quiet, and is woken by one line when there is something to decide: accept, send back with notes, or reject. Everything in between — running the worker in its own copy of the repo, checking its work, review by other models, retries, budgets — the hub does without Claude.

Why

• Claude's context is the expensive part. A delegated task costs Claude about 2 KB of text for the whole cycle: one event line, a summary capped at 4 KB, a one-word decision. Reports, diffs and logs are there on request, never pushed.
• Cheap models are good enough when the work is checked. A task is not "done" because the model says so: the commit exists, the diff stays inside the allowed files, the acceptance tests pass under a project lock, and one or more reviewer models (in fresh sessions) agree. Typical costs on Muse Spark: recon $0.004–0.06, code with review $0.02–0.13.
• Nothing gets lost. Events are stored until acknowledged, so a restarted Claude session, a restarted hub or a dead terminal does not drop a "done". Worker processes outlive service restarts; orphaned tasks are picked up again; an observer watches the hub itself.

What it does

How it compares

Several projects connect Claude Code to other agents. Most are thin bridges (start a session, poll its status). agent-hub is narrower in scope and deeper in the task lifecycle:

If you want a dashboard for twenty parallel agents with PR automation, use a fleet orchestrator. If you want Claude Code to stop burning its context on work a cheaper model can do — and to trust the result — this is it.

See it work: a real task on Muse Spark ($0.007)

Install

Requires Python 3.11+, git, and at least one provider: opencode (free models work) or Google Antigravity CLI (agy).

pipx install ahub
ahub setup                 # language, project, providers, models, service, Claude skill, Telegram (optional)
ahub doctor                # what is wrong and how to fix it

ahub setup --yes takes all defaults without questions. Telegram is optional: pipx install 'ahub[telegram]'.

Platforms: Linux (systemd), macOS (launchd; tested in CI), Windows via WSL2 only.

Use

In Claude Code (after ahub setup installs the skill):

ahub task new --kind code --title "add retry to the payment client" \
  --spec-file spec.md --paths "app/payments/**,tests/**" --accept "tests/test_payments.py"

Claude keeps a Monitor on ahub watch; when a line like DONE T12 … arrives it runs ahub status T12 and decides: ahub accept T12, ahub rework T12 --notes "…", or ahub reject T12.

You: ahub top to watch (press c for control mode, ? for help), ahub status, ahub history. Plain-language tasks: ahub draft new "what you want, in your words" → preview → ahub draft start N.

Everything else: ahub --help.

Documentation

• docs/ARCHITECTURE.md — code map, start here
• docs/architecture.md — design
• docs/contracts.md — interfaces between the parts

Development

python -m venv .venv && .venv/bin/pip install -e '.[dev]'
.venv/bin/python -m pytest -q          # ~2.5 min, no network, HOME is faked
AHUB_LIVE=1 .venv/bin/python -m pytest -m live   # real providers, costs money

License

MIT