Kinde is Good to agents.
Discry independently scored how well an AI agent can discover and understand the Kinde 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 · 79/100Comprehension
55% of score · 96/100What we found
- An agent looking to operate Kinde programmatically would find a complete, publicly hosted OpenAPI 3.0 spec (425KB, ~30 tagged resource groups with operationIds, summaries and descriptions) linked directly from llms.txt — no signup wall, no scavenger hunt.
- Kinde ships a first-party MCP server with setup docs for Cursor, Claude Desktop, Claude Code and VS Code, so an agent can manage users/orgs/roles in natural language — a stronger agent signal than most auth competitors offer, though it is not yet indexed in the major third-party MCP registries.
- The discovery surface is purpose-built for LLMs: llms.txt (concise, API-focused index), llms-abridged.txt, a 2.6MB llms-full.txt, and per-section .txt bundles — an agent can choose its altitude instead of scraping HTML.
- Documentation is genuinely task-oriented with realistic, copy-pasteable multi-language examples (curl, Node, Python using values like john.snow@example.com and kp_ IDs) and end-to-end workflow tutorials (e.g. drip-feed migration, M2M token → call users), so an agent can chain operations rather than guess.
- The two notable gaps are agent-tooling conventions rather than content: no AGENTS.md in the primary kinde-oss SDK repos and no /.well-known/mcp.json — both cheap to add and would close most of the remaining discovery distance.
What to change
Prioritized by impact on discoverability. You (or your docs platform) deploy these — Discry never touches your API.
- 01Add AGENTS.md to the primary kinde-oss SDK repositories (kinde-go, kinde-management-api-js, kinde-python-sdk) with build/test/auth context for coding agents — a low-effort win that currently scores zero on a rising-standard discovery check.
- 02List the official Kinde MCP server in Glama, Smithery and PulseMCP so agents discovering capabilities via registries (not just docs.kinde.com) can find it — the server already exists, it just needs distribution.
- 03Publish a /.well-known/mcp.json pointing at the Kinde MCP server with its tool declarations and auth requirements, making the existing server machine-discoverable from the docs domain.
- 04Consolidate error recovery into a single 'errors and how to handle them' reference covering the top 5-10 error scenarios (400 parameter-level causes, 403, 429) with explicit fix steps — today the excellent 429/backoff guidance is strong but recovery guidance is uneven across other error classes.
- 05Surface the OpenAPI spec at a conventional docs path (e.g. docs.kinde.com/openapi.yaml) in addition to the kinde.com spec URLs, so agents probing standard locations resolve it without parsing llms.txt first.
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.
Execution characteristics are unusually well documented for an auth platform: a shared error_response JSON schema is referenced on 400/403 responses across the OpenAPI spec, a dedicated rate-limits page documents the 429 throttle with concurrency limiter and exponential-backoff guidance, and cursor pagination (page_size + next_token / starting_after) is specified consistently. Idempotency keys are not documented.