MiniMax is Agent-Ready to agents.
Discry independently scored how well an AI agent can discover and understand the MiniMax API from what’s public — not whether it’s usable. Below: every signal we checked, what’s costing the score, and what to change.
SCORED UNDER RUBRIC 1.2 · A full re-launch under Discry Score 2.5 — a new behavioral instrument, not comparable to these scores — is in progress.
Discovery
45% of score · 95/100Comprehension
55% of score · 96/100What we found
- An agent discovering MiniMax would land on a Mintlify-hosted docs site with a best-practice llms.txt at /docs/llms.txt that carries a title, structured sections (Docs / OpenAPI Specs / AsyncAPI Specs), per-link descriptions, and clean .md mirror URLs for every page — near-ideal machine discovery.
- The OpenAPI contract is not hidden: llms.txt directly links 15 publicly reachable OpenAPI 3.1.0 specs (e.g. openapi-chat-anthropic.json) complete with request schemas and worked examples, so an agent can generate a typed client without scraping HTML.
- Endpoint docs are task-oriented and copy-pasteable: every capability page opens with what it accomplishes, and code examples use realistic values across curl / Python / Node.js — an agent can lift and run them immediately.
- Capability boundaries are unusually explicit for an AI vendor: per-model context windows, RPM/TPM limits, file-size caps, character limits, voice-clone expiry (168h), and audio-URL TTL (9h) are all stated, so an agent learns limits from docs rather than by failing.
- The main gap is error-recovery depth: the Error Codes table pairs codes with solutions, but many resolutions are generic ('Please retry your requests later') rather than actionable self-correction steps, so an agent hitting a transient error gets weak guidance on what to change.
What to change
Prioritized by impact on discoverability. You (or your docs platform) deploy these — Discry never touches your API.
- 01Deepen error-recovery guidance: replace generic 'retry later' solutions with actionable steps (which Retry-After/backoff to apply on 1002 rate limit, which parameter is invalid on 2013, how to remedy 1042 invisible-character-ratio failures) so agents can self-correct instead of blindly retrying.
- 02Publish a .well-known/mcp.json (currently 404) so MCP-aware agents can auto-discover MiniMax's official Python/JS MCP servers directly from the platform domain rather than only via third-party registries.
- 03Document rate-limit response headers (whether Retry-After or remaining-quota headers are returned) so agents can pace requests programmatically instead of inferring from the static RPM/TPM tables.
- 04Add idempotency-key support (or explicitly document its absence) for the asynchronous task-creation endpoints (video, long-text TTS) so agents can safely retry create calls without duplicate billing.
Execution coverage · INFORMATIONAL, UNSCORED
Whether an agent can actually complete a call and recover from errors is the deeper Audit layer — documented here, but not part of the Discry Score.
Auth is a single Bearer API key model (pay-as-you-go key or Token Plan subscription key), documented clearly with env-var setup. A dedicated Error Codes page lists numeric codes with messages and per-code solutions, and rate limits are published as explicit RPM/TPM tables per model. Pagination and idempotency are not documented, though most core surfaces are synchronous single-shot calls or task-id polling flows where pagination is less relevant.