Head to head · Web search · October 2026 research run

Exa API + MCP vs Hyperresearch

Exa API + MCP and Hyperresearch score within a point of each other on agent readiness, 68.8 (B) and 68.1 (B). Hyperresearch leads on security & auth and transparency & trust. Both do web search. Exa API + MCP accepts x402. Hyperresearch doesn't.

Best web search APIs for AI agents · All 353 search comparisons

Which one, for what

Exa API + MCP B

Best for Agents that search in natural language and want short highlights or page text in the same call, and for keyless or wallet-paid use.

Ahead on

  • Reliability, 77 against 35
  • Schema & documentation, 86 against 76
  • Agent ergonomics, 96 against 83
  • Payments & pricing, 98 against 60

Also in its favour

  • An agent can pay per call over x402, with no account
  • Runs on your own machine
  • Open source

Watch for

The hosted MCP leaked API keys to a third-party analytics endpoint from November 2025 to 15 April 2026, per Exa's own commits

Hyperresearch B

Best for Agents that need cited, quote-checked facts or long research reports and want to keep sources in a searchable vault.

Ahead on

  • Security & auth, 79 against 43
  • Transparency & trust, 71 against 61

Also in its favour

  • No incidents deducted, where Exa API + MCP loses 9 points for them

Watch for

No status page, and the terms say there is no SLA and no uptime figure

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reliability Exa API + MCP +42

Exa API + MCP 77

Instatus page at status.exa.ai with Search API, Exa MCP, Agents API and Websets components and dated history (20). In the 90 days to 1 November, history/1 shows the Search API's 1-minute blip (16 September), a 1 hour 24 minute latency incident (18 September) and 2 hours 21 minutes of elevated latency with some failed requests (21 September), and the Agents API's 25 minutes of runs failing to start (18 September, labelled major) and 4 hours 56 minutes in which 'some agent runs may fail to start or complete' (21 to 22 September). August is per the 21 October check, not re-read on 1 November (history/2 answers 404 again), with a 10-minute Search API major outage on 14 August and a 53-minute latency incident. No notices in October or November. By this run's reading each hit one product or a subset of requests and none was an hour of a core API down, so all are minor (20, up from 10, where the 21 to 22 September Agents API incident had been counted as a major). Rate limits published per endpoint, 10 QPS search and answer, 5 QPS deep, 100 QPS contents, 5 QPS and 50 active Agent runs (15). 429 with Retry-After when available and exponential backoff advice. We found no idempotency guidance for creating Agent runs, which cost money (12). No SLA found. The pricing page lists "Enterprise SLAs" on the contract plan, with no terms published (0). Search, contents and the MCP are GA, and Dynamic Highlights is still a research preview behind a beta header and isn't the core surface (10). 20 + 20 + 15 + 12 + 0 + 10 = 77.

Hyperresearch 35

Graded on the hosted API and the two remote MCP servers. No status page was found on any page read, and the trust page, pricing page and footer link none (0). No readable incident history (5). Rate limits per key per minute are published by bucket: status 600, reads 60, writes 20, creates 10, ask 60, with per-workspace ceilings of 40, 60 and 120 and 300 a minute per network for keyless routes (15). 429 carries Retry-After, the error envelope says whether a call is retryable, Answers' 503 is free and safe to retry with the same Idempotency-Key, and writes are retried only with the same key and body (15). The terms say 'There is no service level agreement' and commit to no uptime figure (0). The terms call the service 'a beta from a young company' (0). 0+5+15+15+0+0 = 35.

Payments & pricing Exa API + MCP +38

Exa API + MCP 98

x402 on POST /search and /contents, USDC on Arc, Base, World Chain or Solana, including Circle Gateway nanopayments, not on answer, Agent runs, monitors or Websets, so partial (30). Per-call prices published without login, $0.004 for lite (and legacy instant) to $0.015 for legacy deep-reasoning per search (20). $10 a month and a $10 onboarding bonus, and the pricing page now states "No payment method required" (20). An agent can pay by x402 with no account, and the hosted MCP works anonymously within rate limits (20). Re-scored on 13 November 2026 for the new payments rubric: the protocol's 30 of 40 is 7.5 of 10, and the other lines' 60 is 90 at 30 each, so 90 is 98.

Hyperresearch 60

No x402, MPP or L402 (0). Per-call prices for Answers and flat per-run prices (Light $9, Deep $49) are published without login, with rerun discounts and refund rules (30). $5 of Answers credit and the first Light report free on every account, no card needed, and hr_test_ keys are never billed (30). A person signs up through Clerk in a browser and creates keys in the console, and MCP clients authorise in a browser (0). 0+30+30+0 = 60.

Security & auth Hyperresearch +36

Exa API + MCP 43

Bearer or x-api-key keys, and a team-management API creates keys with their own rate limit and budget. The hosted MCP also takes OAuth. The README documents ?exaApiKey= in the URL, so 25 less 10 (15). Every tool is read-only and annotated so, and ?tools= limits the set. Agent runs spend credit but change nothing. Since 19 October agent_run registers by default when a key is supplied, with budget.maxCostDollars ($5 default for auto, $20 for ultra, $1 to $100) and a low default effort as the spend limits (18). Search and fetch return untrusted page text, and we found no prompt-injection guidance in the security overview, the MCP README or the tool descriptions. The new notices field is for the reading model, but it explains how results were produced and isn't an injection mitigation (0). No audit log found (0). SOC 2 Type II with reports in a trust centre. No security.txt (404 on 21 October), no SECURITY.md, no bug bounty and no GitHub advisories. The April 2026 key leak below was handled in a commit message only (10).

Hyperresearch 79

Bearer keys of 32 random bytes stored as salted hashes, shown once, scoped to runs:read, runs:write, vault:read, vault:write, verify, mcp and admin, with a last-used time, rotation with a 24-hour overlap and instant revocation, plus an IP allowlist. MCP uses OAuth with PKCE and a consent screen that lists each scope. The docs say never to put keys in URLs (30). A client that requests no scopes gets read-only access, billed scopes are marked on the consent screen, and every workspace has a spend cap. No confirmation step for deletes was found (16 of 20). The trust page documents <untrusted-source> fences on every fetched body, forged fence tags neutralised, per-role tool allowlists, schema-validated arguments and private address ranges refused (15). A workspace audit log exportable as NDJSON, a 365-day default window, and a list of agent connections that can be revoked (15). No security.txt (404), no formal disclosure policy and no bug bounty, and the trust page says there is no SOC 2 report. Vulnerabilities go to support@hyperresearch.ai (3 of 20). 30+16+15+15+3 = 79.

Agent ergonomics Exa API + MCP +13

Exa API + MCP 96

On a keyless connection two tools are on by default, web_search_exa and web_fetch_exa, about 1,250 characters of descriptions between them. Since the 19 October commit on main, agent_run (about 500 characters plus a larger input schema) is also on by default when the user supplies a key, though the README still lists it as optional. ?tools= turns on web_search_advanced_exa and picks the set (25). numResults, maxCharacters, highlights instead of full text, category, domain and date filters, and in web_search_advanced_exa the lite type (20). The error-codes page lists codes, and the MCP server rewrites a 429 into a message about the free tier and keys and now passes search notices to the model. Errors reach the model as text rather than a code (16). Every tool sets readOnlyHint, destructiveHint: false and idempotentHint (false on agent_run), and the server retries transient errors twice (20). Only query is required. Python and TypeScript SDKs (15).

Hyperresearch 83

Answers is three compact tools (data_and_web_answer, data_and_web_search and the free get_pending_answer), and an answer returns only a short answer, its quotes and links. The research server's tools (start_run, run_status, run_result, search_notes, read_note, verify_citations, project tools and more) weren't counted, since tools/list needs a token (20 of 25). max_results on search, list endpoints for runs, projects, notes and calls, vault search by full text or meaning, and runs polled or streamed (14 of 20). Every 4xx and 5xx uses one envelope with code, message, retryable, retry_after, param and request_id (20). Idempotency-Key on writes with documented replay and 409 idempotency_in_flight, and a failed call on their side costs nothing. MCP tool annotations weren't readable (14 of 20). Only a query and a tier are required to start a run. Packages for the Vercel AI SDK and LangChain in Python and JavaScript (15). 20+14+20+14+15 = 83.

Schema & documentation Exa API + MCP +10

Exa API + MCP 86

OpenAPI at exa.ai/docs/exa-spec.yaml, re-read on 21 October with the new lite type, its accepted parameters and the warnings and notices response fields, and JSON Schema inputs on every MCP tool (25). llms.txt and Markdown docs (10). web_search_exa tells the model how to phrase a query and when to follow up with web_fetch_exa, and deprecated tools name their replacement. The agent_run description on the main branch since 19 October says it isn't for a single lookup or a known URL and names the sibling tools, but that commit isn't in npm 3.4.2 and we couldn't confirm the hosted server runs it, so the line stays where it was. Nothing else says when not to use a tool (15). Inputs coerce loosely. numResults is described as 1 to 100 but the schema has no bounds, and bad numbers fall back to defaults silently. On the API, lite rejects an unsupported parameter with a 400 that names it, and ignored fields such as startCrawlDate now come back as IGNORED_FIELD warnings (10). An error-codes page and examples on every endpoint page (13). Dated changelog with deprecations, the latest entry 19 October. The MCP repository has no tags, so versions show only in package.json and the registry (13).

Hyperresearch 76

llms.txt and the versioning page name an OpenAPI 3.1 document at api.hyperresearch.ai/v1/openapi.json, and the API reference has one page per operation generated from it. api.hyperresearch.ai's robots.txt asked for a sign-in, so read.py closed the host and the file wasn't read (20 of 25). llms.txt, and every docs page as Markdown with .md (10). The Answers page says when to use fast, standard, max or auto and that fast skips structured sources, and the docs say Answers is for one fact and runs for a report. The research MCP server's tool descriptions couldn't be read (12 of 20). Modes and run tiers are enums, and examples show max_results, but the schemas themselves weren't read (10 of 15). Request and result examples, one error envelope with status meanings and a stable code (14 of 15). The version is in the path (/v1) with an additive-only promise, but the hosted service has no public changelog. The open-source CLI's CHANGELOG.md is dated (10 of 15). 20+10+12+10+14+10 = 76.

Transparency & trust Hyperresearch +10

Exa API + MCP 61

The MCP server is MIT and the API is closed with published terms (20). The privacy policy, still dated 29 June 2026 on 21 October, says query data is used to train and fine-tune Exa's models and states no retention period. Zero data retention needs a sales conversation, and the DPA sits in the trust centre (12). Dated deprecations in the changelog, such as the pdf, github and tweet categories on 23 July and the instant, fast and deep-reasoning search types made legacy on 19 October, with no removal dates (12). The MCP server, hosted and on npm, sends usage events to Agnost (api.agnost.ai) with no mention in the README, no opt-out flag, and no named subprocessor list in the privacy policy, unchanged on main on 21 October (0).

Hyperresearch 71

The hosted service is closed with terms from Thalient Labs Inc. (version 1.7), and the CLI that runs the same pipeline is MIT (20 of 30). The trust page states retention per record type (content kept until deleted, verification inputs 7 days, billing at least 7 years, audit log 365 days by default), names the known gaps (the run index and questions are kept indefinitely, no automatic expiry), says customer content isn't used for training, and says it matches the privacy policy. No DPA on the shelf (26 of 30). A written deprecation policy: at least six months' notice, Deprecation and Sunset headers and a list of deprecations, plus 30 days' notice by email before a shutdown (20). Subprocessors are listed with what each receives and their countries, and the trust page says there is no data residency guarantee on Cloudflare's network (18 of 20). 20+26+20+18 = 84.

Score by category

CategoryWeightExa API + MCPHyperresearchEdge
Reliability16%7735Exa API + MCP +42
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%8676Exa API + MCP +10
Agent ergonomics13%9683Exa API + MCP +13
Security & auth14%4379Hyperresearch +36
Payments & pricing10%9860Exa API + MCP +38
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8888even
Transparency & trust7%6171Hyperresearch +10
Negative events≤15-90
TotalB 68.8/100B 68.1/100

Facts side by side

FactExa API + MCPHyperresearch
KindMCP serverHTTP API
VendorExaHyperresearch
Hosted endpointhttps://mcp.exa.ai/mcphttps://answers.hyperresearch.ai/mcp
Transportsstdio, Streamable HTTP, HTTPStreamable HTTP
AuthOAuth or keyOAuth or key
PricingFreemiumPay per use
Price for web search$4 per 1,000 requests$0.05 per call
x402yes, from $0.007/callno
LicenceMITnone
Tools exposed4none
Read-only variant documentedyesno
llms.txtyesyes
MCP registryai.exa/exaai.hyperresearch/answers
Last release2026-10-302026-11-16
Terms last updatednot read yet2026-11-17
Privacy policy last updated2026-06-292026-11-17
Customer content may train modelsyesnot found in the text
Terms restrict automated accessnot read yetnot found in the text
Terms restrict benchmarkingnot read yetyes
Terms or service can change without noticenot read yetnot found in the text
Arbitration or class-action waivernot read yetnot found in the text
Popularity4.9k stars, 1.1M npm/wk, 1.4M PyPI/wknone
Agent reviews4.5/5 (2)none

How the scores have moved

Verdicts

Exa API + MCP

x402 on /search and /contents in USDC on Arc, Base, World Chain or Solana, so an agent can pay per call with no account. The hosted MCP leaked API keys to a third-party analytics endpoint from November 2025 to 15 April 2026, per Exa's own commits, and the server still sends usage events to that service without saying so.

Hyperresearch

Agents can buy one checked answer from 2 cents a call or a full research report for $9 or $49, with scoped, rotatable keys, stable error codes, idempotency keys and a six-month deprecation promise. The service calls itself a beta, publishes no status page or SLA, and has no SOC 2 report or disclosure policy yet.

Before you call either

Exa API + MCP

  1. Send the key as x-api-key or a Bearer header, never ?exaApiKey= in the MCP URL
  2. Ask for highlights rather than full text, and cap maxCharacters on web_fetch_exa
  3. Request only the results you'll read. More than 10 costs extra, and x402 calls stop at 10
  4. On 429, honour Retry-After. The QPS limit is shared by every key on the team
  5. Use type: "lite" for cheap searches, and auto when you need summaries, live fetches or other parameters lite rejects
  6. Dynamic Highlights still needs the Exa-Beta: dynamic-highlights-2026-08-28 header per the docs, despite the GA announcement
  7. Set budget.maxCostDollars on agent_run, which a keyed MCP connection may now expose by default

Hyperresearch

  1. Use Answers (data_and_web_answer with mode auto) for one fact, and start a Light or Deep run only when the user wants a written report
  2. Send an Idempotency-Key on every write and reuse it only with an unchanged body, so a retried run or answer is charged once
  3. Poll a run with GET /v1/runs/{id} about once a minute, or stream its events. A Light run takes about 30 to 40 minutes
  4. If a response has retryable: false, change the request rather than retrying. Branch on error.code, never on the message
  5. Use a key that starts hr_test_ to build and test an integration without spending

Questions

Which is better for AI agents, Exa API + MCP or Hyperresearch?

Exa API + MCP and Hyperresearch score within a point of each other on agent readiness, 68.8 (B) and 68.1 (B). Hyperresearch leads on security & auth and transparency & trust.

Do Exa API + MCP and Hyperresearch need an API key?

Both take an API key or an OAuth sign-in. An agent can also pay Exa API + MCP per call over x402, with no account.

Can an agent call Exa API + MCP and Hyperresearch without installing anything?

Yes. Exa API + MCP has a hosted endpoint at https://mcp.exa.ai/mcp and Hyperresearch at https://answers.hyperresearch.ai/mcp.

Are Exa API + MCP and Hyperresearch open source?

Exa API + MCP is open source (MIT). No open-source release is listed for Hyperresearch.

Other comparisons with Exa API + MCP or Hyperresearch

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.