GLM-as-Subagent for Claude Code — plug & play

📦 Canonical source: https://github.com/djerok/glm_mcp_claude — created by @djerok. If you found this via a fork, mirror, or an awesome-list, the original lives here. Please ⭐ / file issues / open PRs at the source.

Add GLM (Zhipu / Z.ai) to Claude Code as a cheap, full-capability subagent (~10× cheaper than Opus), with automatic per-task routing between Opus and GLM. Your main agent stays on Opus; GLM does the well-specified, cost-sensitive work — and can read, write, edit, and run your files directly. One command to install.

✅ Works in the Claude Code app on a subscription-based Claude. Your main agent runs on the Claude you already pay for through the Claude Code app (Pro / Max / Team subscription) — no separate pay-per-token Anthropic API key required. Only GLM needs a (cheap) Z.ai key. Opus orchestrates on your subscription; GLM does the heavy lifting for a fraction of the cost.

↑ The glm subagent (orchestrated by Haiku 4.5, the cheap layer) reading the repo and offloading the heavy work to GLM via the MCP tools — the Opus → Haiku → GLM hybrid in action.

# no clone needed — run straight from GitHub:
npx github:djerok/glm_mcp_claude --key YOUR_ZAI_API_KEY

# or clone and run the installer:
node install.mjs --key YOUR_ZAI_API_KEY

…then restart Claude Code. That's it. (Details below.)

🔑 Your key must be from the Z.ai / Zhipu GLM Coding Plan. Get one at https://z.ai → subscribe to the GLM Coding Plan, then create an API key. A generic / free key without coding-plan access will not work for the coding models used here.

What you get

• glm subagent — a full-tool subagent (read/write/edit/bash) powered by GLM.

• glm_agent tool — GLM as a real file-editing agent with built-in oversight (diff, dry-run, git revert).

• glm_delegate / glm_recommend / glm_status — draft-only delegation, a free routing advisor, and a health check.

• Auto-delegation hook — when you spawn a subagent, it injects a GLM-vs-Opus verdict so cheap work goes to GLM automatically. Zero token cost when you're not spawning subagents.

• If you explicitly name an agent ("use opus", "use the sonnet agent", "use glm"), the hook stays silent and just routes where you asked.

Related MCP server: Ollama MCP Server

Prerequisites

• The Claude Code app (desktop or CLI), signed in with a subscription-based Claude (Pro / Max / Team). Your main agent uses this — no Anthropic API key needed. The claude CLI should be on your PATH (claude --version).

• Node.js ≥ 18 (node -v)

• A Z.ai / Zhipu API key with GLM Coding Plan access — get one at https://z.ai. ⚠️ It must be on the GLM Coding Plan (the coding-plan subscription); a generic or free key won't have access to the coding models this uses. This is the only paid key required, and GLM is ~10× cheaper than Opus.

• Git (optional, but enables glm_agent's one-command revert)

Install (recommended: global, all projects)

# from this folder:
node install.mjs --key YOUR_ZAI_API_KEY

The installer:

1. copies the server to ~/.claude/glm-mcp/ and runs npm install,

2. writes your key into ~/.claude/glm-mcp/.env,

3. installs the glm subagent (~/.claude/agents/glm.md) and the hook (~/.claude/hooks/),

4. wires the hook into ~/.claude/settings.json (backs it up first),

5. adds a short delegation policy to ~/.claude/CLAUDE.md,

6. registers the MCP server with claude mcp add glm -s user.

Then restart Claude Code and run glm_status — you should see "api_key_loaded": true.

Options: --no-register (skip the CLI step), --skip-npm, --claude-dir PATH (custom config dir). Re-running is safe (idempotent). No key on the command line? Run node install.mjs, then edit ~/.claude/glm-mcp/.env and set GLM_API_KEY=....

Per-project instead of global

Don't want it everywhere? Skip the installer. Copy glm-mcp/ into your project, cd glm-mcp && npm install, copy .mcp.json.example → .mcp.json in the project root, set the key, and (optionally) copy agents/glm.md to .claude/agents/ and the hook into .claude/ + .claude/settings.json.

How it works (the short version)

You ask for something
  → Opus orchestrates
  → wants to delegate a chunk → spawns a subagent
       → [hook fires] "[GLM router] GLM-suitable repo task → use glm_agent (dry_run first)"
              (or "keep on Opus" for hard/sensitive work)
       → Opus runs glm_agent (GLM edits the files, runs tests) — or keeps it on Opus
       → you get a diff + action log + a one-command revert

The routing rules live in glm-mcp/src/router.js and the hook — not in always-on context — so they cost nothing until a subagent is actually spawned.

Routing in one line: GLM is the default (it's ~10× cheaper); Opus is the exception for work where being wrong is expensive — subtle debugging, architecture, large refactors, security, tool-heavy dependent loops, huge context, vision, or anything you mark sensitive.

The tools

Example: directly calling the GLM agent

A real run — asking GLM (via glm_agent) to write a file end-to-end on disk:

Prompt: "Using the GLM agent glm_agent, write a 2000-word essay in Shakespearean format about the usefulness of an umbrella, into my Desktop."

GLM did it itself — created the file directly, no round-tripping the content through the main agent:

• Output: Umbrella-Essay-Shakespeare.md — ~2,260 words of Early Modern English (thee/thou/thy, doth/hath) with two blank-verse interludes

• Work: 18 tool-loop iterations; one file created, nothing existing touched

• Cost: ~$0.064 — a fraction of running the same task on Opus

That's the point: the orchestrator stays on Opus while glm_agent does the heavy, file-touching work for cents.

Oversight (how you stay in control of glm_agent)

• Entry: you/Opus choose when to call it and with which workdir.

• dry_run: true: GLM proposes a full diff and writes nothing — approve, then apply.

• After a real run: you get the unified diff, an action log, and a one-command git revert (git checkout <baseline> -- .).

Note: file/bash ops inside glm_agent run in the MCP server process (not gated per-edit) and are scoped to the workdir you pass. That's intentional (max autonomy) — point it only at repos you're fine letting it modify.

Configuration (~/.claude/glm-mcp/.env)

Full list with comments: glm-mcp/.env.example.

Uninstall

node uninstall.mjs          # remove agent, hook, settings entry, MCP registration
node uninstall.mjs --purge  # also delete ~/.claude/glm-mcp (and its .env)

Security

• Never commit/share your .env or a .mcp.json containing the key. .gitignore excludes them.

• GLM routes through servers in China — don't send secrets/regulated code you wouldn't send to a third-party API. (Routing keeps sensitive-flagged work on Opus, but you decide what to delegate.)

Troubleshooting

More background and the research behind the routing rules: see docs/.

Contributing

PRs and issues welcome — see CONTRIBUTING.md. Good first areas: routing rules (glm-mcp/src/router.js + the hook), provider adapters, and docs. Please never commit secrets/.env.

License

MIT © djerok

Original / canonical repository: https://github.com/djerok/glm_mcp_claude. If you fork, mirror, or redistribute this project, please keep a link back to the source so others can find updates, file issues, and contribute. Built by @djerok.