v1.3.0 — colab Claude Code plugin

Ship the repo as the colab Claude Code plugin.

Added

• Claude Code plugin support: .claude-plugin/plugin.json, .mcp.json registering the bundled colab MCP server, and a uv-based launcher (hooks/scripts/launch-mcp.sh) that resolves the project virtualenv from pyproject.toml on first launch.
• /colab:setup — guided first-run setup (uv check, optional Playwright extra, Google sign-in, verification).
• /colab:help — grouped overview of every tool the server exposes.

Install

/plugin marketplace add ringo380/robworks-claude-code-plugins
/plugin install colab@robworks-claude-code-plugins

v1.2.0

Round-trip helper tools for the Unsloth -> merged -> HF -> local MLX fine-tune workflow.

Added

• Colab-side: save_and_push_merged (reliable save_pretrained_merged with safetensors verification, README seeding, and stale-file clearing via delete_patterns) and ensure_repo_readme.
• Local MLX (host-side): normalize_tokenizer_config (fixes Colab's TokenizersBackend tokenizer_class), download_from_hf, and convert_to_mlx (clears the target dir, auto-normalizes a local source, surfaces a sentencepiece hint, preserves the prior output on failure).
• --mlx-python CLI arg / COLAB_MCP_MLX_PYTHON env var to select the interpreter the local MLX tools invoke (defaults to the server's own interpreter).

v1.1.1

Security

• Bump fastmcp 2.14.5 -> 3.2.0, resolving three advisories in unused fastmcp features (OpenAPI provider SSRF/path traversal, OAuth proxy confused-deputy, Gemini CLI command injection). None were reachable in this codebase.

Changed

• Use the non-deprecated fastmcp.server.providers.proxy import path for FastMCPProxy.

Added

• Installation section in the README with clone and dev-install instructions.
• .gitignore covering Python build artifacts, virtual environments, and macOS files.

Full Changelog: v1.1.0...v1.1.1

v1.1.0 - Expanded Colab tool coverage

First tagged release. Adds 22 tools for broad Google Colab coverage on top of the multi-session, headless-capable baseline.

New tools

• Runtime: get_resource_usage (live GPU util/mem, RAM, disk)
• Drive: mount_drive, unmount_drive, list_drive_files
• Secrets: inject_secret_to_env (sets an env var without returning the value), get_secret
• VM filesystem: list_vm_files, read_vm_file, write_vm_file, delete_vm_file
• Inspection: list_variables, inspect_variable (type, shape/dtype, DataFrame head)
• Async execution: run_async, poll_execution (incremental output via cursor), stop_async (cooperative), list_jobs - background in-kernel thread for hours-long jobs
• Browser/UI (Playwright sessions): change_runtime_type, connect_runtime, factory_reset_runtime, save_notebook, complete_drive_mount_consent
• Rich output: optional capture_plots on execute_code returns matplotlib figures as image blocks

Notes

• Tools reach Colab three ways: frontend-proxied (execute_code, cell ops), execute_code-composed Python (Drive/secrets/VM/inspection), and Playwright DOM control for actions the kernel cannot perform. Browser-only tools return a clear error on the non-Playwright backend.
• inject_secret_to_env refuses loader-sensitive variables (PATH, LD_PRELOAD, etc.) and never echoes the secret value.
• Includes a pytest suite (37 tests); the Playwright-dependent tests skip cleanly when the headless extra is not installed.

Install: pip install . or pip install '.[headless]' for browser-driven tools.