Loading ToolBench
Windows only. Generic stdio MCP package for Cursor Bridge: CCE search and bounded Cursor Agent execution. Evidence-backed Cursor Context Engine search and bounded, explicitly continuous Cursor Agent execution over CDP.
Mixed quality across 11 tools. Strengths: all tools have descriptions (194-400 chars, within baseline); input schemas present with types and descriptions; tool annotations (readOnlyHint, destructiveHint) correctly applied. Weaknesses: parameter descriptions are often generic or lack actionable constraints (e.g., 'Optional request provenance' repeated 8 times); no documented output schemas; error handling guidance absent; confirmation parameters (enum-gated) are security-positive but add friction; some tools conflate multiple concerns (grok_session_inspect accepts 8 optional params with interdependencies, undocumented).
Cursor Context Engine (CCE) read-only, evidence-driven project-understanding engine. Resolve natural-language intent into verifiable code context using indexed semantic retrieval, exact text search, symbol/reference tracing, and targeted source reading. Returns compact evidence anchors ordered by strength.
Cancel an active delegation prompt in a Cursor Agent session. Requires explicit confirmation.
Create a new bounded Cursor Agent session in the exact workspace with explicit delegation mode control. Returns session ID and initial state.
Send a delegation prompt to an existing Cursor Agent session and optionally wait for completion. Requires explicit user authorization via confirmation parameter.
Query the state of an existing Cursor Agent session, including agent execution status, reply content, and completion signals. Supports polling with optional wait for terminal state.
Explicitly initialize or reinitialize the persistent local HTTP proxy used by Grok Build Supervisor. With no URL, discovers and CONNECT-verifies a local listener; with proxyUrl, verifies that exact loopback endpoint. Refuses while the Supervisor has active work and never opens a TUI or sends a prompt.
grok_session_inspect has 8 optional parameters with undocumented interdependencies (view, afterSequence, limit, runId, waitMs, sequences, sessionId, sessionQuery). No guidance on which combinations are valid or what each view returns.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does cce_search return? Is it a list of code snippets, file paths, or structured evidence objects?).
Parameter descriptions are boilerplate and lack actionable constraints. 'Optional request provenance metadata with sender (user|model|unknown) and source (user|model|mixed|unknown)' appears 8 times verbatim; no guidance on when to include it or how it affects behavior.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Cancel the active prompt, disconnect ACP, stop the Supervisor-owned Leader, or durably acknowledge an exact unknown-after-restart record after all related process boundaries are inactive. Acknowledgment preserves the outcome as unknown and does not cancel, resume, prompt, kill, or claim completion.
Read cursor-coalesced state from the persistent Grok Supervisor daemon, including exact workspace-trust requests, optionally waiting up to 25 seconds for completion, failure, permission, or clarification. Returns interaction, status, summary, delta, or evidence views with bounded metadata.
Transactionally create or resume one daemon-owned Grok session in the exact workspace. Different workspaces have independent connections and writer leases. The default presentation is a visible Windows Terminal PowerShell TUI. Headless ACP-only presentation allowed only after explicit user request and separate confirmation.
Start one asynchronous prompt turn in the exact attached Grok session. The prompt may cause Grok actions; call only after user authorization and pass SEND_TO_GROK confirmation.
Answer one exact pending Grok ACP permission or form-elicitation request. Never guesses a permission option or clarification value.
Confirmation parameters (enum-gated strings like 'DELEGATE_TO_CURSOR', 'CANCEL_CURSOR_SESSION') are security-positive but lack explanation of why they exist or what happens if omitted. LLMs may not understand the intent.
No error handling guidance. Tools do not document what errors are retryable, user-fixable, or fatal. LLMs cannot recover from failures without explicit recovery hints.