Exa Agent
Exa · Exa Agent: Web Research Agent API for List Building, Enrichment, and Deep Research
Exa Agent is an asynchronous web research API from Exa. You describe the data you want and the JSON shape it should come back in, and the agent plans searches, reads sources, verifies candidates, and returns schema-validated results with citations. It is aimed at developers building list building, CRM enrichment, KYC/KYB, and deep research into their own products and pipelines.
Best for
Developers and data teams who need cited, structured web research inside their own product or pipeline, such as account lists, CRM enrichment, due diligence, or KYC/KYB checks, without building their own search-and-verify agent loop
Not ideal for
People who want a chat-based research assistant, workloads that need sub-second answers, teams that need a fixed monthly cost for open-ended research, and organizations that require a self-hosted deployment
Who it's for
Developers, data engineers, and GTM, finance, and compliance teams building research or enrichment into their own tools
Capabilities
- Asynchronous runs through `POST /agent/runs` that fan out searches, read and condense pages, split list building into parallel subtasks, and verify each candidate against your criteria
- Schema-validated JSON output from any JSON Schema in `outputSchema`, plus a natural-language answer
- Field-level grounding that cites the sources behind each value in structured results
- Row enrichment and list expansion with `input.data`, and `input.exclusion` to skip entries you already have
- Seven effort levels, from `minimal` to `xhigh` at fixed per-request prices, `auto` for variable-scope work, and `ultra` for exhaustive research
- Agent Ultra budgets: a cost cap from $1 to $100 per run, a time limit from 5 minutes to 3 hours, and early stop that keeps the results found so far
- Server-sent event streaming, polling, stored event replay, run listing, and follow-up runs that carry context with `previousRunId`
- Exa Connect data partners as additional sources, called when a schema field names a specific provider
- Contact enrichment for emails and phone numbers, billed separately at $0.02 per email and $0.07 per phone number
- Python (`exa-py`) and JavaScript (`exa-js`) SDKs, an OpenAI-compatible Responses API, and the `agent_run` tool in Exa's hosted MCP server
- Zero Data Retention, enabled per team on request (listed as an Enterprise plan feature on the pricing page)
Limitations
- A developer API, not an end-user research app: results come back as JSON and text for your own code or MCP client to use
- Higher latency by design. Exa recommends its Search API for single low-latency lookups, and Ultra runs typically take about 30 minutes and up to three hours
- `auto` and `ultra` runs are metered, so cost varies with task complexity up to the default caps of $5 and $20 per run (adjustable from $1 to $100). Large `input.data` lists consume more compute
- Benchmark results on the product page and launch post are run and reported by Exa and have not been independently reproduced
- With Zero Data Retention enabled, run data is kept only up to 10 minutes after a run ends, events cannot be replayed, and `previousRunId` and Exa Connect data sources are unavailable
- Hosted only: the agent cannot be self-hosted and is not open source, although the `exa-py` SDK and the Exa MCP server are open source
- Budget time limits and early stop work only with `effort: "ultra"`
Use cases
- Building a list of every company in a market that meets several criteria, with founders, funding, and customers for each and a cited URL for every field
- Enriching CRM accounts with judgment fields such as "sells to hospitals" or "who owns the data platform"
- Researching a customer for KYC or KYB across filings, press, court records, and regulator sites into one cited record
- Expanding an existing lead list by passing the rows you already have and excluding them from new results
- Adding a cited deep-research step to an internal agent or workflow through the OpenAI-compatible API or MCP
Our take
Exa Agent is worth a look when the output you need is a table rather than an essay. Field-level citations, JSON Schema output, input rows, and exclusions are the parts that matter for building account lists, enriching a CRM, or running diligence checks, and they are what you would otherwise have to build on top of a search API. Fixed per-request prices for the lower effort levels make simple enrichment easy to budget, while `auto` and Ultra trade predictability for completeness, so set budgets on those runs. Treat the benchmark charts as Exa's own claims, and test it against your real schemas before committing.
Who should use it
Engineering and data teams building GTM lists, enrichment, due diligence, or compliance research into their own systems, and developers who want a cited research step inside an agent through MCP or an OpenAI-compatible API.
Who should skip it
Non-developers who want a ready-made research app, products that need instant answers, teams that need a fixed price for unbounded research, and organizations that can only use self-hosted tools.
Strengths
- Returns structured JSON with citations at the field level, not just a prose report
- Clear effort-based pricing, with fixed per-request prices for five levels and caps on metered runs
- Works from SDKs, an OpenAI-compatible API, and MCP clients
- Built-in list expansion, exclusions, and follow-up runs
- Free monthly credits to try it without a payment method
Weaknesses
- Developer-only: there is no end-user app for the agent itself
- Metered runs make costs vary, and Ultra runs can take hours
- Benchmark claims are vendor-reported
- No self-hosted or open-source option
Exa Agent pricing
Starter
$0
- $10 in credits every month plus a $10 onboarding bonus
- No payment method required
- Access to all endpoints, including Agent
- 50 Agent concurrency
- MCP server access
Developer
Pay as you go
- Fixed efforts: minimal $0.012, low $0.025, medium $0.10, high $0.50, xhigh $1.00 per request
- auto: metered, default cap $5 per run
- ultra: metered, default cap $20 per run
- Self-serve Exa Connect data providers
Enterprise
Custom
- Custom Agent concurrency
- Zero data retention
- Enterprise SLAs and support
- Premium Exa Connect data providers
Free tier limits: Starter plan: $10 in free credits each month plus a $10 onboarding bonus, no payment method required, and 50 Agent concurrency.
Note: Metered runs are billed by component: $0.10 per Agent Compute Unit and $0.005 per search call. Email contact enrichment costs $0.02 per email and phone enrichment $0.07 per number. `budget.maxCostDollars` sets a per-run ceiling from $1 to $100 for auto and ultra runs, and runs that finish early cost less. Exa Connect data providers are billed on top of Agent usage.
Technical specs
API pricing
Per request for fixed efforts ($0.012 to $1.00); metered for auto and ultra at $0.10 per Agent Compute Unit plus $0.005 per search, with default caps of $5 and $20 per run
Where Exa Agent excels
Building a market map for due diligence
A single `auto` or `ultra` run can search for companies that match open-ended criteria and fill fields such as founders, funding, and customers, each with a cited source, instead of a team assembling the list by hand.
Enriching CRM accounts with judgment fields
Passing existing accounts in `input.data` with a schema for fields like "sells to hospitals" returns validated values with citations, so reviewers can check the evidence behind each field before it reaches the CRM.
Adding cited web research to an internal agent
Because Exa Agent is exposed through MCP and an OpenAI-compatible Responses API, teams can give an existing assistant a structured research step without writing their own search and verification loop.
Exa Agent vs. competitors
Exa Agent vs. You.com
You.com's Research API returns cited research answers at several effort tiers, with beta JSON Schema output and a background mode for its deepest tier. Exa Agent is built around list building and enrichment: you pass input rows and exclusions, continue a run to expand a list, attach Exa Connect data partners, and use metered `auto` and `ultra` levels for open-ended jobs. You.com fits cited research answers inside apps, and Exa Agent fits building and enriching entity lists.
Exa Agent vs. Perplexity
Perplexity is primarily a consumer answer engine, with an Agent API (models plus web search tools, with JSON Schema output) and a Search API for developers. Exa Agent has no consumer app and is designed around entity lists: input rows, exclusions, follow-up runs, field-level grounding, and data-partner sources. Perplexity fits teams that also want a consumer research product, and Exa Agent fits structured research data in a pipeline.
Frequently asked questions
What is Exa Agent?
Exa Agent is an asynchronous API from Exa for multi-step web research. You send a query and, optionally, a JSON Schema, input rows, and exclusions, and the agent runs the searches, page reads, and verification needed to return a text answer, schema-validated JSON, and citations for each field.
How much does Exa Agent cost?
Fixed effort levels cost $0.012 (minimal), $0.025 (low), $0.10 (medium), $0.50 (high), and $1.00 (xhigh) per request. The `auto` and `ultra` levels are metered at $0.10 per Agent Compute Unit plus $0.005 per search, up to default caps of $5 and $20 per run, and email and phone enrichment cost $0.02 and $0.07 per contact. Exa's free Starter plan includes $10 in credits each month plus a $10 onboarding bonus.
What is Agent Ultra?
Agent Ultra is Exa Agent's highest effort level, launched on September 25, 2026, for large list building and exhaustive research. Exa says it typically finishes complex tasks in about 30 minutes and can take up to three hours. You can cap a run's cost between $1 and $100 and its duration between 5 minutes and 3 hours, or stop it early and keep what it found.
Can I use Exa Agent from Claude or Cursor?
Yes. Exa's hosted MCP server includes an `agent_run` tool that is enabled by default, so MCP clients such as Claude and Cursor can start Exa Agent runs. Developers can also call it through the Python and JavaScript SDKs or the OpenAI-compatible Responses API.
Does Exa Agent support zero data retention?
Yes. Zero Data Retention is enabled per team on request, and Exa's pricing page lists it as an Enterprise plan feature. With ZDR enabled, run data is available only while a run executes and for up to 10 minutes after it ends, stored events cannot be replayed, and follow-up runs and Exa Connect data sources are not available.
Exa Agent vs Perplexity: what is the difference?
Perplexity is mainly an end-user answer engine, with a developer Agent API that combines models with web search tools and supports JSON Schema output. Exa Agent is developer-only and built specifically for research-to-table jobs: input rows to enrich, exclusions, follow-up runs that expand a list, field-level grounding, and Exa Connect data partners.
Integrations & fit
Alternatives to consider
About Exa Agent
Exa Agent sits on top of Exa's search engine and index, but replaces the loop of search, read, and verify calls you would otherwise write yourself. A run takes a natural-language query and, optionally, a JSON Schema for the output, rows of existing data to enrich (`input.data`), entries to exclude, and Exa Connect data partners to consult. The agent splits the work into subtasks, runs subagents across many sources in parallel, uses frontier models only where a step needs them, and returns a text answer, schema-validated JSON, field-level grounding that ties each value to its source, and a cost breakdown. Runs are asynchronous: you poll for the result, stream server-sent events as the run progresses, replay stored events later, or continue from an earlier run with `previousRunId` (for example, "find 10 more"). The same agent is available through the Python and JavaScript SDKs, an OpenAI-compatible Responses API (`model: "exa-agent"`), and the `agent_run` tool in Exa's hosted MCP server, so it can be called from Claude, Cursor, and other MCP clients. Pricing follows an effort setting: five fixed levels cost $0.012 to $1.00 per request, while `auto` and `ultra` are metered by usage with default caps of $5 and $20 per run. Agent Ultra, launched on September 25, 2026, is the highest effort level for exhaustive list building and deep research; Exa says runs typically take about 30 minutes and can take up to three hours, with budget and time limits you can set and an option to stop early and keep what has been found. Exa publishes benchmark comparisons against frontier models and Perplexity, but these are vendor-reported and not independently reproduced. The tradeoffs are that it is a developer API rather than an app, it is slower than a single search call by design, metered runs make costs vary by task, and it is a hosted service with no self-hosted option.
Updates from Exa Agent
Agent Ultra is Exa Agent's highest effort level for large list building and exhaustive research. It coordinates many subagents across thousands of sources, is metered up to a default $20 per run, and accepts cost and time budgets. Exa reported higher scores than Opus 5.5, GPT-6 Astra, and Perplexity Agent on four research benchmarks, using published figures where they existed and its own runs otherwise.
Are you the founder? Claim this listing →