Loading ToolBench
MCP server for xAI Grok models, providing chat completion, image generation, image editing, image description, and video generation capabilities
Strong naming and descriptions across all 7 tools. Tool names follow verb_noun pattern (grok_chat, grok_generate_image, etc.). Descriptions are detailed (150-300 chars) and explain WHEN to use each variant. All parameters have type definitions and descriptions. However, output schemas are not explicitly documented, responses are described as 'text only' or 'inline content' without structured field definitions. Error handling is present but minimal (validation of reasoning_effort and agents params only). No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Image/video tools accept file paths without documented path traversal validation.
Send a chat completion to xAI Grok's flagship model (grok-4.6: 500k context, vision, function calling). This model always reasons before answering and cannot be told not to — for a fast non-reasoning answer call grok_chat_fast instead, and for a hard problem worth a team of agents call grok_chat_multi_agent. Stateless by default. Pass session_id to enable an in-memory thread that persists for the lifetime of this server process; the same session_id on subsequent calls appends to and replays prior turns. Returns the assistant text only.
Send a chat completion to a dedicated non-reasoning Grok model (grok-4.20-0309-non-reasoning: 1M context). The model answers immediately without thinking first, which makes this the fast, cheap path for lookups, classification, extraction, summarising and reformatting. There is deliberately no reasoning option — if the answer needs actual thought, call grok_chat. Sessions work exactly as in grok_chat. Returns the assistant text only.
Send a hard problem to Grok's multi-agent model (grok-4.20-multi-agent-0309), where a team of agents works the question in parallel and reconciles their answers into one. Markedly slower and far more token-hungry than grok_chat — even a trivial question burns thousands of tokens on agent overhead — so reserve it for problems where independent cross-checking is actually worth it. The model tends to append a '\confidence{N}' marker to its answer; that is the model's own output, not a wrapper artifact. Sessions work exactly as in grok_chat. Returns the assistant text only.
Output schemas not documented. Tools return 'text only' or 'inline content' without specifying field structure, data types, or pagination. LLMs cannot plan downstream calls or extract structured data reliably.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications. Image/video generation and editing tools are marked WRITE but lack destructiveHint annotation. Chat tools are READ_ONLY but lack readOnlyHint.
File path parameters (output_path in image/video tools) lack validation guidance. No documented constraints on path traversal, absolute vs relative paths, or filesystem permissions. Tools accept 'absolute path' but do not validate or sanitize.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Describe one or more images using Grok's vision capabilities. Returns the assistant's description text only.
Edit one or more images with Grok and save the results to disk. Saves the bytes to the caller-supplied output_path (must be absolute) and ALSO returns inline image content so the calling agent can see them. When n>1, output_path is treated as a template: '-1', '-2', ... is inserted before the extension.
Generate one or more images with Grok and save them to disk. Saves the bytes to the caller-supplied output_path (must be absolute) and ALSO returns inline image content so the calling agent can see them. When n>1, output_path is treated as a template: '-1', '-2', ... is inserted before the extension.
Generate a video with Grok and save it to disk. Saves the bytes to the caller-supplied output_path (must be absolute) and ALSO returns inline video content so the calling agent can see it.
Error handling is minimal. Only grok_chat validates reasoning_effort enum and grok_chat_multi_agent validates agents count. No guidance on retryability, user-fixable vs fatal errors, or recovery steps. Image input resolution and file writing errors are not documented.
Session management (session_id, reset_session) is stateful and process-scoped. Descriptions note 'persists for the lifetime of this server process' but do not warn about data loss on restart or multi-instance deployment risks. No guidance on session cleanup or TTL.