Exa Human-Dependent

AUDIE Score: 57/100 · Audited 2026-04-14 · Website: https://exa.ai · Machine-readable: JSON

Pillar Scores

P1 Signal Architecture — 17/25

P2 Clarity Stack — 13/25

P3 Trust Envelope — 9/20

P4 Velocity Triggers — 8/10

P5 Gravity Design — 10/20

Executive Summary

Exa sits in the Human-Dependent tier (57/100) despite being one of the most technically agent-ready search APIs available. Its strongest pillar is Signal Architecture — the official MCP server, six OpenAPI specs, and a well-structured llms.txt give agents excellent API discovery and integration signals. Friction-free activation (no-key MCP, instant API key, free tier) makes autonomous onboarding possible. The critical gaps are in Trust (no audit trail, thin third-party reviews, no scoped permissions) and Clarity (zero substitution/fallback guidance, JS-rendered pricing). Exa is genuinely built for AI agents as customers; the gap between its 57 score and Agent-Ready status is primarily documentation and trust infrastructure, not product capability.

Strongest Signals

Official MCP server + 6 OpenAPI specs — Exa publishes a live MCP server at mcp.exa.ai/mcp with no API key required, plus six machine-readable OpenAPI specs. This is one of the most complete API signal architectures in this tier of tooling.
llms.txt is genuinely agent-optimized — The docs/llms.txt contains the full documentation index across all API endpoints, SDKs, integrations, and changelog. Agents evaluating Exa can retrieve a complete picture from a single structured file.
Friction-free agent activation — Free tier + instant API key + no-key MCP server + coding agent onboarding flow makes Exa the easiest AI search API for an agent to autonomously evaluate and adopt.
Precise, benchmark-grounded claims — 54.4% FRAMES accuracy, sub-180ms Exa Instant latency, exact QPS limits — agents can evaluate Exa's fit without human interpretation.

Critical Gaps

No schema.org structured data on website — Exa's homepage and pricing contain no machine-readable schema markup. Agents discovering Exa via web crawl cannot parse the offer without rendering JS and interpreting prose.
Zero substitution/fallback guidance — No documentation on what agents should do when Exa endpoints are unavailable, degraded, or at rate limit. An agent mid-task has no programmatic fallback path.
No machine-accessible audit trail — API key usage is visible in a dashboard but there's no queryable log API. Agents cannot verify what they consumed, creating a trust gap for accountability-sensitive deployments.
Trust envelope is thin below Enterprise — SOC 2 Type II exists, but there's almost no third-party review volume (1 Trustpilot review), no public SLA for non-Enterprise, and no agent-scoped permissions for PAYG users.

Priority Actions

Publish substitution and fallback documentation — +4 pts · P2 · Effort: Low
Create machine-accessible usage log API — +3 pts · P3 · Effort: Medium
Add agent-scoped API key permissions — +3 pts · P3 · Effort: Medium

All 20 Criteria

P1-A Structured Data — 1/5
No schema.org or JSON-LD markup found on exa.ai homepage or pricing page. Site appears to be a standard JS-rendered SPA with no machine-readable structured data layer.

P1-B Machine-Readable Pricing — 2/5
Pricing page exists at exa.ai/pricing with a clear HTML table (Search: $7/1K, Deep Search: $12/1K, Contents: $1/1K pages, Answer: $5/1K). However, the scraper noted "obfuscated JavaScript code preventing extraction," meaning pricing is not reliably machine-readable without rendering.

P1-C llms.txt / Agent Layer — 5/5
Official llms.txt at exa.ai/docs/llms.txt (redirected from docs.exa.ai/llms.txt). Contents include full documentation index: API reference, examples, tutorials, integrations (LangChain, LlamaIndex, CrewAI, OpenAI, Claude), 6 OpenAPI specs, SDK references, and changelog entries. Clearly structured for LLM consumption.

P1-D API / MCP Availability — 5/5
Six OpenAPI specifications published (exa-spec, monitors-spec, websets-spec, research-spec, team-management-spec, openapi.json). Official MCP server published at https://mcp.exa.ai/mcp, documented at exa.ai/mcp and exa.ai/docs/reference/exa-mcp. Works with Claude Desktop, Cursor, VS Code, Windsurf, Zed, Gemini CLI, Codex, and more. No API key required for basic MCP access.

P1-E Discoverability (GEO) — 4/5
Strong AI retrieval optimization: llms.txt present, well-indexed docs, first-party integrations with LangChain/LlamaIndex/CrewAI, cited extensively in AI infrastructure comparisons. Used by Cursor, AWS, Databricks, Groq, HubSpot. Benchmark rankings cited in AI search evaluations.

P2-A Offer Completeness — 3/5
Pricing per endpoint (search, deep search, deep-reasoning, contents, monitors, answer) available at exa.ai/pricing. Free tier (1,000 req/month), PAYG, and Enterprise tiers documented. However, pricing is JS-rendered and not in a single machine-parseable source. Enterprise pricing requires human contact.

P2-B Scope & Limits — 4/5
Rate limits explicitly documented per endpoint at exa.ai/docs/reference/rate-limits: /search 10 QPS, /findSimilar 10 QPS, /contents 100 QPS, /answer 10 QPS, /research 15 concurrent tasks. Enterprise offers custom QPS. Clear, structured table.

P2-C Substitution Rules — 0/5
No guidance found anywhere on what happens when endpoints are unavailable, what fallback behavior agents should use, or what alternative endpoints serve as substitutes.

P2-D Conditional Logic Transparency — 2/5
Free/PAYG tiers clearly scoped. Enterprise conditions (ZDR, SLA, dedicated infrastructure) require "contact us" with no disclosed thresholds. Conditions scattered across pricing, docs, and enterprise pages.

P2-E Semantic Precision — 4/5
Highly specific: 54.4% FRAMES benchmark accuracy (vs 44.5% and 21.6% for competitors), sub-180ms latency for Exa Instant, exact QPS limits per endpoint, 70M+ company database size. Minimal vague marketing language in technical documentation.

P3-A Verifiable Performance — 3/5
SOC 2 Type II certified (third-party verified). Benchmarks cited across FRAMES, Tip-of-Tongue, and Seal0 are real third-party evaluations. However, only 1 Trustpilot review found; G2 results are for "Exa Websets" not Exa AI directly. Limited independent third-party review volume for a self-reported claim above 3.

P3-B Scoped Permissions — 2/5
API key-based authentication via x-api-key header. No evidence of granular agent-scoped permissions (time-bounded, amount-bounded, action-bounded). Enterprise has custom rate limits and Zero Data Retention but no documented per-agent permission scoping.

P3-C Audit Trail — 1/5
Usage tracking visible in dashboard (dashboard.exa.ai) and API key management available. No machine-accessible audit log API for agents to query programmatically. No documented log export or webhook for usage events.

P3-D Behavioral Consistency — 3/5
API versioning documented: Exa 2.0 and Exa 2.1 release notes exist on blog. Changelog present in llms.txt docs index. OpenAPI specs version-controlled. No published notice period for breaking changes.

P4-A Friction-Free Activation — 5/5
Free tier with 1,000 requests/month, API key from dashboard.exa.ai with no human gate. Official MCP server at mcp.exa.ai/mcp requires no API key for basic use. Dedicated coding agent onboarding flow at dashboard.exa.ai/onboarding-guest. Startup/education grants ($1K in credits) for qualifying projects.

P4-B Agent Decision Signals — 3/5
Free tier enables autonomous trial. Agent-targeted documentation at /reference/search-api-guide-for-coding-agents. Specific benchmark scores provide evaluation data. But no programmatic signal for "when to upgrade" or explicit agent-legible trigger for activation decisions.

P5-A Integration Depth — 3/5
Websets (custom curated data collections), 70M+ company database, industry-specific indexes (people, companies, code docs, financial data, news), custom dataset support. LangChain, LlamaIndex, CrewAI native integrations. Switching requires rebuilding custom Websets and index configurations.

P5-B Agent Memory Layer — 2/5
Monitors feature enables scheduled recurring searches (persistent query jobs). No persistent agent session memory between requests. Websets provide some account-level data persistence. Not a true agent memory layer.

P5-C Programmatic Renewal — 2/5
Pay-as-you-go billing with automatic charge to payment method. No documented agent-accessible billing API for programmatic tier management or renewal. Startup credits are manual-apply.

P5-D Compounding Value Signal — 3/5
Monitors (scheduled search feeds) and Websets (enriched data collections) compound value with continued use. Custom indexes improve over time. No explicit agent-readable "your value is growing" signal or usage analytics API.

Rubric v1 (April 2026). Scores reflect the company's state on the audit date and may have improved since.