# MCP metadata inspection: exa

Endpoint: https://mcp.exa.ai/mcp
Official source: https://exa.ai/docs/get-started/exa-mcp (reviewed 2026-10-15)
Documented profile: Keyless rate-limited endpoint; authenticated agent execution is a separate profile
Observed: 2026-10-15T07:14:22.146Z
Run: 3d04b2c8-2139-4724-8536-5976519bdbb5 (completed)
Protocol: requested 2025-11-25; server reported 2025-11-25; negotiated 2025-11-25
Coverage: 4/5 checks attempted; 4 passed, 0 failed, 0 unknown, 1 not tested, 0 not applicable.
Catalog: 2 observed tools; complete-catalog. SHA-256: 6786a32a1c87da5fe37eecbdbd76d5a41d45f8a1d008e70f23f222269b6430b3.

- **protocol-negotiation: passed**. Negotiated protocol 2025-11-25; this does not establish host compatibility.
- **initialized-notification: passed**. The initialization notification received HTTP 202 with an empty body.
- **tool-catalog: passed**. Observed the complete advertised catalog: 2 tools across 1 pages.
- **tool-metadata-shape: passed**. SDK metadata shape checked for 2 observed tools. This is not complete JSON Schema validation or behavior verification.
- **oauth-login: not-tested**. OAuth discovery, login, token refresh and authenticated access are outside this anonymous inspection.

- metadata: **passed**
- authenticatedExecution: **not-tested**
- modelApi: **not-tested**
- chatgptHost: **not-tested**
- claudeHost: **not-tested**
- publicationApproval: **not-tested**

Tool executions: **0**. This report contains metadata observations only.

- Metadata-only anonymous inspection. No tools/call, OAuth flow, API model calls, host UI or publication review was executed. No global quality or safety score is produced.
- Annotations are server declarations, not verified behavior. Metadata success does not establish task success, security, usefulness or ChatGPT/Claude compatibility.
- Only active exact HTTPS endpoints in the reviewed source registry are accepted. Redirects, supplied credentials, arbitrary endpoints and OAuth URL fetching are disabled. DNS/TLS use the local runtime: this is not a hardened multi-tenant SSRF sandbox and must not be exposed as a public scanner.
- Session identifiers are used only in memory. Request/response headers, raw error messages, raw descriptions and schemas are excluded from artifacts. Tool names and aggregate metadata remain untrusted server input.
- Only the recorded protocol versions and bounded catalog prefix are supported. Reaching a bound or lacking access means unknown, not a failed server or a passing full audit.