Working fork with pre-registered tools + OAuth runtime control

[SebastianGilPinzon]

Hey all, I've been maintaining a fork that addresses a couple of pain points I've seen come up in discussions here:

1. Tools not visible to MCP clients (#54, #67)

The fork pre-registers all 5 notebook tools at startup instead of relying on notifications/tools/list_changed, so they work with any MCP client (Claude Code, Codex, Kiro, etc.) — no need to wait for a browser connection for tools to appear.

2. Programmatic GPU assignment

Since --enable-runtime was removed from this repo (#41), the fork includes a change_runtime tool that assigns T4/L4/A100 GPUs via OAuth without touching the browser. Token is cached so you only authorize once.

Setup

git clone https://github.com/SebastianGilPinzon/colab-mcp.git

Then point your .mcp.json to the clone with --client-oauth-config for GPU control. Details in the repo README.

Fork: https://github.com/SebastianGilPinzon/colab-mcp

Happy to answer questions if anyone runs into issues with the setup.

[KotVasily]

Thank

[SebastianGilPinzon]

You're welcome

[KotVasily]

One problem is when Agent tries to select a GPU, a session with the GPU is created, but the current session does not change, and as a result, Agent still uses the CPU and cannot change the session.

[SebastianGilPinzon]

Yes, I have to change it mannually everytime, I will investagate if there's a workaround

[henrik253]

Hi,
I published a fork that uses Playwright to identify the specific objects in the DOM tree and automatically select the accelerators. It's not the most robust solution, but it works. I used the same method for other pop-ups, such as the G-Drive and MCP warnings.

https://github.com/henrik253/cool-colab-mcp

[ruben-deice]

Thanks for the fork — pre-registering the cell tools fixed the dynamic-registration gap that broke colab-mcp on Claude Code and Codex. One follow-up.

The post-connection success message advertises seven tools:

Available notebook tools: add_code_cell, add_text_cell, delete_cell, get_cells, move_cell, run_code_cell, update_cell

But only five are pre-registered in _make_injected_tools (session.py). get_cells, delete_cell, move_cell, and run_code_cell still rely on notifications/tools/list_changed — the exact registration path this fork was created to avoid. On Claude Code they don't appear in the tool list even after a successful connection.

Why get_cells matters most: without it, the bridge is effectively write-only. An agent can add_code_cell once, but then has no way to read current notebook state — no cell list, no IDs, no outputs. Iterative work (write → run → read result → adjust) is impossible. Even on a brand-new notebook, the agent loses track after the first cell.

Suggested change — same pattern as existing stubs:

async def get_cells_stub() -> str:
    return NOT_CONNECTED_MSG

async def delete_cell_stub(cellId: str = "", cellIndex: int = 0) -> str:
    return NOT_CONNECTED_MSG

async def move_cell_stub(cellId: str = "", fromIndex: int = 0, toIndex: int = 0) -> str:
    return NOT_CONNECTED_MSG

async def run_code_cell_stub(code: str = "") -> str:
    return NOT_CONNECTED_MSG

Plus four matching Tool.from_function entries in the returned list.

Caveat: I'm guessing argument signatures from tool names — I don't have visibility into the Colab JS side, so the real param schemas (especially for move_cell and delete_cell) may differ. You'd know the correct shapes. Happy to open a PR if you confirm signatures, or do a best-guess PR and let you adjust in review.

Repro on Claude Code:

1. Configure MCP per fork README.
2. Restart client. Five tools appear.
3. Call open_colab_browser_connection. Connection message lists seven tools.
4. Tool list still shows only five. get_cells etc. unreachable.