Courier is Agent-Ready to agents.

What we found

• —Courier is built deliberately for agents: a dedicated /docs/tools/agent-quickstart page written 'for AI coding agents', a first-party remote MCP server at mcp.courier.com, an AGENTS.md in the courier-go SDK, and a courier-skills repo. An agent landing here finds purpose-built onboarding rather than human-only reference.
• —Machine-readable docs are first-class: /docs/llms.txt (a curated, agent-routing index with a data model, intent→endpoint table, and 'Critical Gotchas') and a 2MB /docs/llms-full.txt covering the full surface. An agent can understand core capabilities in a few thousand tokens without scraping HTML.
• —Error recovery is exemplary — the responses page gives every status code an explicit Action column, a Common Validation Errors table (Error → Cause → Fix), and concrete 429 backoff steps. An agent that hits an error is told exactly how to self-correct, not just given a code.
• —The biggest gap is the OpenAPI spec: it demonstrably exists (it powers the live 'Try it' reference and the Stainless-generated SDKs, and every page cites /openapi-specs/openapi.documented.yml), but it is not offered as a cleanly downloadable public spec file — an agent can reconstruct the API from llms-full.txt but can't grab a single machine-readable artifact.
• —Minor inconsistency: request/response fields are mostly snake_case (user_id, created_at, phone_number, list_id) but identifiers come back camelCase (requestId, messageId), a small naming seam an agent must account for when parsing responses.

What to change

Prioritized by impact on discoverability. You (or your docs platform) deploy these — Discry never touches your API.

1. 01Publish the OpenAPI spec at a stable, discoverable public URL (e.g. /docs/openapi.json or a linked download on the API Reference landing). The spec already exists and backs the live reference — exposing it as a downloadable artifact would close the single largest discovery gap and likely move Discovery from B to A.
2. 02Add /.well-known/mcp.json advertising the existing mcp.courier.com server with its tool declarations and auth, so agents can auto-discover the MCP endpoint without reading the quickstart prose.
3. 03Mirror llms.txt to the root (/llms.txt) or add a redirect — the file lives at /docs/llms.txt, so an agent probing the conventional root path gets a 404 despite the excellent index existing one level down.
4. 04Normalize response identifier casing (requestId/messageId) to match the otherwise-consistent snake_case body fields, or document the casing convention explicitly so agents parse responses without surprises.
5. 05Trim or split the llms.txt index (currently ~62KB) toward the <50KB / ~12K-token sweet spot, keeping the agent-routing table and gotchas while pushing exhaustive prose into llms-full.txt, so the index stays cheap to load.