Go SDK for the Mistral AI API, providing types and client methods for chat completions, agents, audio transcription/speech, and related services
This is a Mistral Go SDK exposing 13 agent management tools via MCP. Tool naming follows verb_noun convention (CreateAgent, ListAgents, DeleteAgent) which is good. All tools have descriptions (10-50 chars, mostly adequate). Input schemas are present with typed parameters and descriptions. However, output schemas are completely undocumented, no tool specifies what fields it returns, forcing LLMs to guess. Parameter descriptions are minimal (10-30 chars) and lack constraints, ranges, or format guidance. Error handling is absent from all tool definitions. No tool declares idempotency, permissions, or whether it modifies state (though Risk field hints at this). Composition is reasonable, tools are single-purpose and chainable via IDs. Overall: solid foundation but incomplete for production LLM agents.
Sends an agents completion request
Sends an agents request and returns a stream of chunks
Creates a new agent
Deletes an agent
Deletes an agent version alias
Retrieves an agent by ID
Retrieves a specific version of an agent
Lists all version aliases for an agent
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., agent_id from CreateAgent response). This violates the response-shaper pattern and forces agents to guess field names.
Parameter descriptions are minimal (10-30 chars) and lack constraints. E.g., 'page' has no min/max; 'temperature' has no range (0-2?); 'model' has no enum of valid models. LLMs will pass invalid values without guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
Lists all versions of an agent
Lists agents with optional filters and pagination
Creates or updates an agent version alias
Updates an agent, creating a new version
Switches an agent to a specific version
No error handling guidance in any tool definition. LLMs do not know what to do if CreateAgent fails (retry? ask user? fatal?). No recovery hints like 'If agent_id not found, call ListAgents first.'
ListAgents and ListAgentVersions lack pagination documentation. No mention of result limits, total count, or next_cursor. Returning thousands of agents will blow context windows.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) in definitions. Risk field exists in metadata but is not exposed to LLM. DeleteAgent and DeleteAgentAlias should be marked destructive; GetAgent should be marked read-only.