ArcadeData is Good to agents.
Discry independently scored how well an AI agent can discover and understand the ArcadeData 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 · 76/100Comprehension
55% of score · 91/100What we found
- An agent looking for ArcadeDB would find it immediately: it ships a well-structured llms.txt AND a comprehensive 9.4KB llms-full.txt covering data models, query languages, wire protocols, and deployment — plus an official MCP server listed on PulseMCP, so LLM tooling can connect without any scraping.
- The HTTP/JSON API reference is exemplary for agents: a lead-with-the-answer endpoint table (method + path for every operation), ~50 copy-pasteable curl examples with realistic values, clean H1→H4 heading nesting, and a documented JSON error envelope with distinct 2xx/4xx/5xx shapes.
- robots.txt actively welcomes AI: it carries Content-Signal directives (`ai-train=yes, ai-input=yes`) and blocks no AI crawlers — only legal boilerplate pages are disallowed.
- The primary gap is a standards mismatch, not a content gap: the repo has a CLAUDE.md (Claude-specific) but no AGENTS.md, so cross-agent coding tools that look for the emerging AGENTS.md standard won't find agent context by convention.
- The OpenAPI 3.0.3 spec exists and is referenced (`GET /api/v1/openapi.json`), but an agent can only retrieve it from a running server instance — there is no publicly hosted static spec URL to discover the API before spinning up the database.
What to change
Prioritized by impact on discoverability. You (or your docs platform) deploy these — Discry never touches your API.
- 01Add an AGENTS.md to the ArcadeData/arcadedb repo root (alongside the existing CLAUDE.md) so cross-agent coding tools that follow the AGENTS.md convention pick up build/run/query context automatically — a near-zero-effort discovery win.
- 02Publish the OpenAPI 3.0.3 spec at a stable public URL (e.g. docs.arcadedb.com/openapi.json) or link it from llms.txt, so agents can read the full endpoint contract without first deploying a server instance.
- 03Add explicit error-recovery guidance to the HTTP API reference: for the common failure cases (401 auth, 4xx malformed command, transaction conflicts) state the concrete fix, not just the error shape — this is what turns human-readable error docs into agent-actionable ones.
- 04Add a dedicated capability-boundaries / limits page (max payload sizes, concurrency behavior, vector-index constraints, feature availability by version) so an agent learns limits by reading rather than by failing at runtime.
- 05Serve a .well-known/mcp.json declaring the built-in MCP server's tools and auth, so agents can discover ArcadeDB's MCP capabilities directly from the domain rather than only via the PulseMCP listing.
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 well documented for a self-hosted database. The HTTP/JSON API documents a machine-parseable JSON error envelope (2xx returns a `result` field; 4xx returns an `error` field; 5xx adds `detail` and `exception`), token-based auth via login/logout endpoints, and explicitly labels query endpoints as idempotent. Rate limits and API-level pagination are not documented (expected for a self-hosted engine where the operator controls resources and paging is done via SQL LIMIT/SKIP).