Knock is Agent-Ready to agents.
Discry independently scored how well an AI agent can discover and understand the Knock 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 · 88/100Comprehension
55% of score · 96/100What we found
- An agent looking for Knock would find a near-complete agent-readiness toolkit: a 76KB structured llms.txt, a 2.1MB llms-full.txt, per-page markdown (.md) versions, AGENTS.md files in primary repos, and a dedicated knocklabs/skills repo of packaged skills built explicitly for AI coding agents — Knock has deliberately invested in machine consumers.
- An agent could integrate without trial-and-error: docs ship multi-language, copy-pasteable code examples (Node, Python, Ruby, Java) with task-titled snippets, plus explicit error-recovery guidance (back off on 429, configurable SDK retries, idempotent replay) — the self-correction signals that separate agent-usable docs from human-only docs.
- Knock operates an official MCP server and Agent Toolkit, and publishes its API via OpenAPI (Stainless-generated, with publicly accessible specs split by API surface) — an agent can both discover the tools and consume a typed contract.
- robots.txt allows all crawlers, the sitemap indexes every API/AI/CLI page, and capability boundaries (rate limits, batch limits, data retention, MJML/partial limitations, key-length validation) are documented — an agent discovers limits by reading, not by failing.
- The only material gap is breadth-driven token efficiency: the platform spans workflows, broadcasts, guides, channels, objects, preferences and more, so fully grasping the API surface costs well above 5K tokens despite a clean overview; .well-known/mcp.json is also absent (404).
What to change
Prioritized by impact on discoverability. You (or your docs platform) deploy these — Discry never touches your API.
- 01Add a /.well-known/mcp.json manifest declaring the official MCP server's tools and auth so agents can auto-discover the server without reading prose docs (quick win, the server already exists).
- 02Publish the canonical OpenAPI spec at a stable, vendor-hosted URL (e.g. docs.knock.app/openapi.yml) and link it prominently from the API reference, rather than relying on third-party mirrors — gives agents a first-party typed contract.
- 03Submit the Knock MCP server to the major registries (Glama, Smithery, PulseMCP) so agents browsing those indexes discover it — currently no clear registry listing surfaces in search.
- 04Trim or split the 76KB llms.txt into a lean capability-overview index (<50KB) that points to llms-full.txt for depth, improving right-sized agent consumption.
- 05Add a concise 'Core API in 5 minutes' overview that lets an agent grasp the full resource model (workflows, broadcasts, guides, recipients, channels) in under 5K tokens before diving into individual concept pages.
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.
Exceptionally complete execution documentation: secret-key + signed-JWT auth, machine-parseable errors surfaced as typed SDK exceptions (RateLimitError, APIStatusError, APIConnectionError), documented rate limits including a separate batch rate-limit section, cursor-based pagination, and detailed idempotency-key semantics (255-char keys, replay window, success-only recording).