🚀 tpu-skill-claude — TPU Management Skill & MCP Agent

This repository packages the tpu-management Claude Code skill and the tpu-devops MCP server: an AI DevOps/SRE agent for operating Google Cloud TPU capacity and Gemma 4 vLLM serving on TPU VMs. It finds and provisions TPU capacity (flex-start VMs, queued resources), starts and debugs vLLM, verifies model health, runs benchmarks, analyzes logs with the self-hosted Gemma 4 model, and tears everything down safely.

GitHub: https://github.com/xbill9/tpu-skill-claude

project-setup.sh installs the tpu-management skill and registers the tpu-devops MCP server for any project (idempotent — re-run to refresh):

./project-setup.sh /path/to/project --project <gcp-project-id>   # one project (.mcp.json + .claude/skills)
./project-setup.sh --global                                      # all projects (~/.claude/skills + user-scope MCP)
make init TARGET=/path/to/project ARGS='--project <id>' # same, refreshing skill snapshots first

It uses the system python3 (warning with a pip install -r requirements.txt hint if server dependencies are missing — it never creates a venv), merges the server entry into the project's .mcp.json without touching other servers, and prints the remaining manual steps (restart Claude Code, gcloud auth, HF token). See ./project-setup.sh --help for all options. The installer is also bundled inside the skill itself (mcp/project-setup.sh), so an unzipped dist/tpu-management-skill.zip is self-installing.

This repo's own .mcp.json is gitignored (it embeds your GCP project id) and is generated automatically: ./init.sh registers the server on first run (leaving an existing entry untouched), or regenerate it any time with ./project-setup.sh . --project <gcp-project-id> [--model ... --accelerator ... --tp ...].

Claude Code auto-discovers any skill folder containing a SKILL.md in two places:

• Project-level: <project>/.claude/skills/tpu-management/ — available only in that project (this repo ships its own copy, so working inside this repo needs no install).
• User-level: ~/.claude/skills/tpu-management/ — available in every project on the machine.

Pick the install path that fits:

Option A — Claude Code plugin marketplace (recommended):

/plugin marketplace add xbill9/tpu-skill-claude
/plugin install tpu-management@tpu-skill-claude

This installs the tpu-management skill and registers the tpu-devops MCP server in one step, with updates managed by Claude Code (/plugin → manage/update). Configure the server through environment variables (e.g. GOOGLE_CLOUD_PROJECT, MODEL_NAME, ACCELERATOR_TYPE) — see SKILL.md or the get_help tool for the full list.

Option B — clone and install (all projects on this machine):

git clone https://github.com/xbill9/tpu-skill-claude
cd tpu-skill-claude
make skill-install                                   # skill only
./project-setup.sh --global                          # skill + user-scope tpu-devops MCP server

Option C — zip install, no clone (straight from the packaged zip):

curl -L -o /tmp/tpu-management-skill.zip \
  https://github.com/xbill9/tpu-skill-claude/raw/main/dist/tpu-management-skill.zip
mkdir -p ~/.claude/skills && unzip -o /tmp/tpu-management-skill.zip -d ~/.claude/skills/
~/.claude/skills/tpu-management/mcp/project-setup.sh --global   # optional: register the MCP server

All of these first run make skill (refresh_skill.py), which regenerates the bundled snapshots from the repo-root sources: server.py, project-setup.sh, and requirements.txt are copied into the skill's mcp/ folder, and references/tpu-guide.md is rebuilt from tpu.md with the embedded screenshots stripped. SKILL.md and mcp/startup_script_template.sh are hand-maintained and never overwritten.

After installing (or updating), restart Claude Code or start a new session so it picks up the skill. Verify with /skills — tpu-management should be listed.

Because installs are refresh-and-copy (not symlinks), an installed copy goes stale when server.py, tpu.md, or SKILL.md changes — rerun make skill-install (or make init ...) after editing those files.

The tpu-devops MCP server covers the full TPU serving lifecycle (catalog with usage guidance in SKILL.md, live listing via the get_help tool):

• Capacity discovery & provisioning: sweep zones for available capacity (find_tpu_vm for flex-start VMs, find_tpu for queued resources), check quotas (get_zones_with_available_quota), estimate cost, create flex-start TPU VMs (v6e/v5p) or legacy queued resources (v5e) with an auto-serving startup script, then wait_for_vllm_ready until the model is up.
• Serving stack control: manage the vLLM Docker container (manage_vllm_docker — works on both flex-start VMs and queued-resource nodes), fetch endpoints and the gcloud deployment one-liner, store the HF token in Secret Manager.
• Health, logs & diagnostics: system status dashboard covering both serving paths, model health verification, vLLM/docker/system/serial logs, Cloud Logging retrieval, and Gemma-4-powered log triage (analyze_cloud_logging).
• Inference & benchmarking: query the deployed Gemma 4 endpoint (optional TTFT/throughput stats), run vllm bench serve for benchmark metrics.
• Universal SRE help: a standardized get_help tool describing the active configuration and all exposed tools.

make skill         # Refresh skill snapshots from server.py / tpu.md (also syncs the plugin copy in skills/)
make skill-install # Refresh + copy the skill to ~/.claude/skills (all projects)
make skill-package # Refresh + build dist/tpu-management-skill.zip
make init TARGET=/path/to/project [ARGS='--project my-gcp-id']
                   # Refresh + install skill AND register the tpu-devops MCP server

Edit the repo-root sources (server.py, tpu.md, project-setup.sh), then run the appropriate target — never edit the snapshot copies directly.

The server reads its configuration from environment variables: GOOGLE_CLOUD_PROJECT (falls back to the active gcloud config), GOOGLE_CLOUD_ZONE (default europe-west4-a), GOOGLE_CLOUD_REGION, MODEL_NAME, ACCELERATOR_TYPE, TENSOR_PARALLEL_SIZE. Prerequisites: pip install -r requirements.txt, an authenticated gcloud CLI with alpha components, the TPU API enabled, and a Hugging Face token stored as Secret Manager secret hf-token (the save_hf_token tool does this for you).

When deploying to Google Cloud or Hugging Face, secure credentials using:

• Hugging Face Access Token: Saved locally or to Google Secret Manager.
• Application Default Credentials (ADC): Set up using GCP credentials helper scripts (set_adc.sh).

• SKILL.md — the skill itself: lifecycle, tool catalog, required vLLM flags, field notes, cautions
• GEMINI.md — Gemini CLI integration via a LiteLLM proxy pointed at the self-hosted Gemma 4 TPU endpoint
• references/tpu-guide.md — TPU getting started guide: flex-start zones, quotas, troubleshooting

Google Cloud credits are provided for this project.

#AgenticArchitect #GoogleAntigravity