grok-build-x-search-mcp

data-ai MCP Server

Grok Build plugin for Claude Code, Grok, Codex, Cursor and others. Provides /grok:search, /grok:review, /grok:rescue commands + grok_search MCP tool for live X/web search.

VerifiedInstall Ready
data-aidata-ai
2 views25 stars4 forksv0.1.4Apache-2.0

Why This Matters

Discovered via github-topic:mcp and last synced 1mo ago.

VerifiedInstall Ready
Source
github-topic:mcp
Stars
25
Last synced
1mo ago
Install
Instructions detected

Install

1. Install the package

npx grok-build-x-search-mcp

2. Add to claude_desktop_config.json

{
  "mcpServers": {
    "grok-build-x-search-mcp": {
      "command": "npx",
      "args": [
        "grok-build-x-search-mcp"
      ]
    }
  }
}

Config file location: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%\Claude\claude_desktop_config.json (Windows)

6
Tools
0
Resources
0
Prompts
Standard I/O
Transport

Available Tools (6)

GROK_BIN

Path to the `grok` binary if it isn't named `grok` on `PATH`.

Agent

How to add

medium

high>` are optional runtime controls. You can also just ask in natural language, e.g. *"Ask Grok to redesign the retry logic to be more resilient,"* and Claude will route it to the subagent. ### `/grok:status`, `/grok:result`, `/grok:cancel` ```bash /grok:status # list running and recent jobs for this repo /grok:result # final output of the most recent completed job /grok:result --job <id> # a specific job /grok:cancel # cancel the most recent running job ``` `/grok:result` includes the Grok **session id** so you can reopen that run directly with `grok -r <session-id>` and continue it inside Grok. ## How it works The plugin wraps the local `grok` CLI's **headless mode**: ```bash grok -p "<prompt>" --output-format json # → { "text": ..., "stopReason": ..., "sessionId": ..., "requestId": ... } ``` There is no separate server or daemon. Grok manages its own session store (under `~/.grok/sessions`), so: - it uses the same Grok install and login you use directly, - sessions started here can be resumed in the Grok TUI with `grok -r <id>` (and vice versa), - read-only commands (search/review) use a **denylist** of write/shell/MCP/media tools plus `--sandbox read-only` and `--no-subagents`, so Grok cannot edit your repo. They also disable vendor MCP imports in the child so a search turn cannot re-enter `grok_search` (fork-bomb defence; see issue #1). > **CLI note (`--tools` bug):** on grok-cli 0.2.x, `grok -p ... --tools web_search,web_fetch` > often fails session creation (`Requirements unsatisfied` for `run_terminal_cmd` / > `auto_background_on_timeout`). This plugin therefore **never** uses `--tools` for > search/review; it uses `--disallowed-tools` instead. Set `GROK_CC_FORCE_TOOLS_ALLOWLIST=1` > only if you know your CLI build has fixed allowlists. Background jobs run the headless turn in a detached process, streaming output to a per-repository log under `~/.grok/cc-plugin/jobs/`. `/grok:status` and `/grok:result` read from there. ### Environment variables

GROK_CC_STATE_DIR

Override the background-job state directory (used by tests).

GROK_SEARCH_MCP_MAX_INFLIGHT

Max concurrent `grok_search` calls per MCP server process (default `2`).

GROK_CC_FORCE_TOOLS_ALLOWLIST

Set to `1` to re-enable `--tools` allowlists (not recommended on CLI 0.2.x).