Gemini CLI MCP Extension

Bring Gemini CLI automations to GitHub Copilot inside VS Code. This extension exposes a local MCP stdio server that Copilot (or any MCP-aware client) can call to run long-running Gemini tasks such as linting, formatting, repo analysis, or end-to-end test runs. Every task is tracked, logged, and surfaced in the editor so you always know when Copilot is busy on your behalf.

Requirements

Node.js 18+ (MCP server runtime).
Gemini CLI installed locally and accessible through gemini or a fully qualified path.
A VS Code workspace folder or an explicit absolute geminiMcp.taskCwd path; the server refuses to run against the extension’s installation directory.
(Optional) Install @google/gemini-cli-core either globally (npm install -g @google/gemini-cli-core) or in the workspace so the MCP server can reuse the core SDK. The server already falls back to the CLI binary when the core package is absent.

How the Extension Runs

Activation: the MCP server is registered as soon as VS Code finishes loading and a workspace (or configured taskCwd) exists.
Triggering: Copilot invokes Gemini tools automatically when it needs help. No extra confirmation prompts are shown because all calls originate from your local Copilot session.
Visibility: a status-bar item and the Gemini CLI MCP Tasks view show running/queued counts, last log lines, and duration. Selecting a task opens tasks://{id}/log.
Safety: if the CLI fails validation (missing binary, quota exhausted, etc.), Copilot calls are rejected until you fix the issue and reload VS Code.

Key Features

Local MCP stdio server (server/dist/index.js) with concurrency control via a worker pool.
Persistent task tracking under .vscode/gemini-mcp/ (metadata, live status snapshots, rotating logs).
Quick commands: Gemini CLI MCP: Show Worker Status and Gemini CLI MCP: Open Task Log.
All worker limits, queue settings, and default priorities are configurable via VS Code settings or environment variables.

Configuration in VS Code (`geminiMcp.*`)

Setting	Description
`maxWorkers`	Maximum concurrent Gemini processes (default `3`).
`taskCwd`	Working directory for spawned tasks (supports `${workspaceFolder}`); must be absolute if no workspace is open.
`maxQueue`	Queue size before new tasks are rejected (default `200`).
`modelBridge.*`	Settings that control the optional local automation bridge (stdio/socket).

Changing any of these settings reloads the MCP provider and refreshes the worker status view.

Model Automation Bridge

geminiMcp.modelBridge.allowedTools: whitelist of MCP tools that automation callers may invoke; leave empty to allow all tools discovered from the workspace mcp.json manifest.
geminiMcp.modelBridge.allowOrchestrator (default true): when true, the bridge also allows orchestrator requests so the orchestrated review workflow can be triggered programmatically.
geminiMcp.modelBridge.requestTimeoutMs (default 120000): per-request timeout (ms) for proxied MCP tool calls; requests exceeding this will be canceled and return a timeout response.

The bridge gives Copilot-style agents the same model/tool surface that the extension already exposes via geminiMcp.orchestrateReview and geminiMcp.invokeTool. The current implementation is a local stdio/socket bridge rather than an HTTP server, and it is intended for use by trusted local agents.

Environment Variables (Server Side)

Use these when launching VS Code or wrapping code in a script (e.g., .envrc, shell profile):

GEMINI_CLI, GEMINI_MAX_WORKERS, GEMINI_TASK_CWD, GEMINI_MAX_QUEUE
GEMINI_TIMEOUT_MANIFEST

The server validates that GEMINI_TASK_CWD (or the resolved workspace root) lives inside a workspace folder, your home directory, or the OS temp directory.

MCP Surface

Developer helper tools (code intelligence and transformation):

dev.summarizeCode — summarize a source file or snippet (3-5 sentences / bullet points).
dev.explainSnippet — provide a step-by-step / line-by-line explanation of a code snippet.
dev.generateComments — generate detailed inline and header comments for provided code.
dev.refactorCode — refactor code according to a specified goal, returning refactored code or a diff.
dev.extractInterface — extract/infer an interface or type definition from implementation code.
dev.generateTests — generate unit tests targeting a specified framework.
dev.translateCode — translate code between languages while preserving functionality.

Web lookup tools (search external references and examples):

web.findCodeExample — search the public web for code examples matching a feature or API query (returns snippets and links).
web.findLibraryUsage — search for documentation, usage examples and API references for a given library/component/plugin.
Resources: tasks://{id}/log, tasks://{id}/summary

All responses include logUri pointers so Copilot (or you) can stream output inside VS Code.

Logs & Retention

Task metadata: .vscode/gemini-mcp/tasks.json
Live status snapshot: .vscode/gemini-mcp/status.json
Logs: .vscode/gemini-mcp/logs/{taskId}.log
Cleanup: completed tasks older than 7 days are pruned automatically.

Notes

Allowed working directories: workspace folders, your home directory, or the OS temp directory. Anything else is rejected.
If Gemini CLI health transitions to missing, quota_exhausted, or error, reload VS Code (which restarts the MCP server) after fixing the underlying issue; the server does not automatically retry those states.

Gemini CLI MCP

lpyedge