cafitac-hermit-agent 0.3.20

Hermit keeps Claude Code or Codex as the orchestrator while a local MCP executor handles cost-aware coding work.

HermitAgent

Run Claude Code or Codex cheaper — Hermit is an MCP executor that handles the high-token mechanical work while your orchestrator stays in charge.

┌──────────────┐
│  Claude Code │──┐
│  (planner)   │  │    ┌──────────────┐   any OpenAI-compatible   ┌───────┐
└──────────────┘  ├───▶│  HermitAgent │ ────────────────────────▶ │  LLM  │
                  │    │  (executor)  │                           └───────┘
┌──────────────┐  │    └──────────────┘
│    Codex     │──┘         ~$0 / flat-rate
│  (planner)   │
└──────────────┘

Claude Code or Codex stays the orchestrator — planning, interviewing, code review. Hermit takes the rest: file edits, test runs, commits, refactors, on a cheap local or flat-rate model via MCP. The switch is one word in a slash command: /foo → /foo-hermit.

Install

npm install -g @cafitac/hermit-agent
hermit install    # guided setup (Claude Code + Codex)

Requires Node.js 20+ and Python 3.11+. The npm package bootstraps a managed Python runtime under ~/.hermit/ on first run — no repo checkout needed.

To upgrade: hermit update

Quick start

hermit-mcp-server   # starts the gateway + MCP stdio server

Then in Claude Code:

/feature-develop-hermit <task>

Claude interviews, writes the plan, and delegates implementation to Hermit over MCP. Executor tokens never hit your orchestrator bill.

Reference skills

Four example skills ship under .claude/commands/. Fork these into your own workflow:

Command	Claude does	Hermit does
/feature-develop-hermit	interview + plan	implement + test
/code-apply-hermit	read PR review	apply every change
/code-polish-hermit	pick what to polish	lint/test loop
/code-push-hermit	write PR description	commit + push

See docs/hermit-variants.md to add your own.

Executor LLM

ollama (local, free):

brew install ollama && ollama pull qwen3-coder:30b

z.ai (flat-rate subscription) — add to ~/.hermit/settings.json:

{
  "providers": {
    "z.ai": {
      "base_url": "https://api.z.ai/api/coding/paas/v4",
      "api_key": "<your key>",
      "anthropic_base_url": "https://api.z.ai/api/anthropic"
    }
  }
}

Configuration

~/.hermit/settings.json (created by hermit install):

{
  "gateway_url": "http://localhost:8765",
  "gateway_api_key": "hermit-mcp-…",
  "model": "glm-5.1",
  "routing": {
    "priority_models": [
      {"model": "gpt-5.4", "reasoning_effort": "medium"},
      {"model": "glm-5.1"},
      {"model": "qwen3-coder:30b"}
    ]
  }
}

Providers not configured or installed are skipped automatically. Explicit model names are strict — a missing provider returns a clear error instead of silently falling back.

Architecture

• AgentLoop — LLM turn → tool call → result → compact on context fill
• Gateway — FastAPI relay in front of the executor (routing, 429 failover, dashboard at :8765)
• MCP server — run_task / reply_task / check_task / cancel_task
• TUI — optional React+Ink terminal UI for standalone interactive sessions (hermit)

Tests

.venv/bin/pytest tests/

Status

Early, working, MIT. No release cadence guarantees.

License

MIT — see LICENSE.

See also

• docs/cc-setup.md — Claude Code MCP registration details
• docs/hermit-variants.md — the -hermit skill family
• docs/measure-savings.md — cost-savings measurement protocol
• benchmarks/ — reproducible task specs and community datapoints