Deepgram is Agent-Ready to agents.
Discry independently scored how well an AI agent can discover and understand the Deepgram 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 · 100/100What we found
- An agent landing on the docs is handed an explicit runway: llms.txt opens with an 'Instructions for AI Agents' section telling it to append .md to any URL for clean markdown, to fetch section-level llms.txt indexes, and to connect to a hosted MCP server at /_mcp/server — Deepgram treats agents as first-class consumers, not an afterthought.
- Discovery is near-complete: a full OpenAPI 3.1 spec (plus AsyncAPI 2.6 for WebSocket channels) is publicly served at /openapi.json and /openapi.yaml, robots.txt allows all AI crawlers, a 105-URL sitemap covers every doc page, and an official Deepgram MCP server is listed on Glama. An agent can find and machine-read this API through multiple independent paths.
- Comprehension is best-in-class: every endpoint page leads with method/URL/purpose, code examples use realistic values (real sample audio URLs, nova-3 model, snake_case JSON) across six languages, and multi-step guides like Voice Agent Message Flow walk the full WebSocket handshake → settings → conversation loop with copy-pasteable JSON.
- Error recovery is written for self-correcting agents, not just humans: the STT troubleshooting page maps NET-0001 to a KeepAlive fix, explains the dg-error/dg-request-id debug headers, warns against empty-byte sends, and links dedicated 'recovering from connection errors' guides — an agent can read what to DO, not just what broke.
- The only real gaps are minor: no /.well-known/mcp.json (the MCP server lives at a non-standard /_mcp/server path), and the llms.txt is a 65KB full-docs index — comprehensive but above the ~50KB size that keeps it cheap for an agent to load wholesale.
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 pointing to the existing hosted MCP server at /_mcp/server so agents can auto-discover it via the standard well-known path instead of only finding it inside llms.txt.
- 02Serve a genuinely full-content /llms-full.txt (concatenated page markdown) distinct from /llms.txt — today both URLs return the same link index, so an agent expecting the long-form variant gets a duplicate.
- 03Trim or split /llms.txt to stay under ~50KB (e.g. a lean top-level index that points to the per-section llms.txt files you already publish), so agents can load the master index without spending ~16K tokens.
- 04Surface pagination and idempotency conventions explicitly in the docs (or the API overview) — list endpoints exist but an agent has no documented contract for paging large result sets or safely retrying writes.
- 05Promote the agentic-tools assets (MCP server, deepgram/skills repo, .md convention) into a short top-of-funnel 'For AI Agents' overview page so the strongest agent-readiness features aren't only discoverable by reading llms.txt.
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.
Deepgram documents token-based auth (Authorization: Token), a machine-parseable JSON ErrorResponse schema (in OpenAPI 3.1) plus named streaming error codes (e.g. NET-0001) and debug headers (dg-request-id, dg-error), and concurrency/rate limits with explicit backoff and queueing strategies. Pagination and idempotency were not surfaced in the reviewed pages.