Exa Agent

Exa Agent is an async, high-compute, usage-based endpoint that handles list building, enrichment, and deep research tasks that require dozens of structured output fields and complex reasoning. Each run can return a natural-language answer, schema-validated JSON, field-level grounding, metadata, and a cost breakdown. You can retrieve completed runs later, list past runs, replay events, or continue from a previously completed run.

Prefer MCP? Exa Agent and Exa Connect are available in Exa MCP. Enable tools=agent_tools to run multi-step research, list-building, enrichment, and structured output from Claude, Cursor, and other MCP clients.

When to use Exa Agent

Use Exa Agent when a workflow needs more than a single search or extraction call:

Build lists from open-ended criteria, then enrich each result
Research entities across many fields with citations
Run multi-hop tasks like “find companies, then find their decision makers”
Produce structured JSON from a long-running web research task
Continue from a previous run with a follow-up request like “find 10 more results”

For simpler low-latency search, start with the Search API.

Quickstart

This example starts a run that builds a structured list of people matching your criteria. It returns JSON in output.structured.

1. Install the Exa SDK

pip install exa-py

npm install exa-js

2. Set your API key

macOS/Linux
Windows

export EXA_API_KEY="your-api-key"

setx EXA_API_KEY "your-api-key"

3. Create a run

import json
from exa_py import Exa

exa = Exa()
run = exa.agent.runs.create(
    query="Find engineering leaders at AI infrastructure companies that raised a Series A or B in the last 6 months.",
    output_schema={
        "type": "object",
        "properties": {
            "people": {
                "type": "array",
                "maxItems": 10,
                "items": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string"},
                        "job_title": {"type": "string"},
                        "linkedin_url": {"type": "string", "format": "uri"},
                    },
                    "required": ["name", "job_title", "linkedin_url"],
                },
            }
        },
        "required": ["people"],
    },
    effort="auto",
)

print(json.dumps(run.model_dump(), indent=2))

import Exa from "exa-js";

const exa = new Exa();
const run = await exa.agent.runs.create({
  query:
    "Find engineering leaders at AI infrastructure companies that raised a Series A or B in the last 6 months.",
  outputSchema: {
    type: "object",
    properties: {
      people: {
        type: "array",
        maxItems: 10,
        items: {
          type: "object",
          properties: {
            name: { type: "string" },
            job_title: { type: "string" },
            linkedin_url: { type: "string", format: "uri" }
          },
          required: ["name", "job_title", "linkedin_url"]
        }
      }
    },
    required: ["people"]
  },
  effort: "auto"
});

console.log(JSON.stringify(run, null, 2));

curl -s -X POST "https://api.exa.ai/agent/runs" \
  -H "Content-Type: application/json" \
  -H "x-api-key: $EXA_API_KEY" \
  -d '{
    "query": "Find engineering leaders at AI infrastructure companies that raised a Series A or B in the last 6 months.",
    "effort": "auto",
    "outputSchema": {
      "type": "object",
      "properties": {
        "people": {
          "type": "array",
          "maxItems": 10,
          "items": {
            "type": "object",
            "properties": {
              "name": { "type": "string" },
              "job_title": { "type": "string" },
              "linkedin_url": { "type": "string", "format": "uri" }
            },
            "required": ["name", "job_title", "linkedin_url"]
          }
        }
      },
      "required": ["people"]
    }
  }'

Add Accept: text/event-stream when creating a run to receive server-sent events as the run is queued, started, and completed. See Stream events for more details.

4. Poll for completion

If you do not stream events, save the returned id and poll the run until it reaches a terminal status.

import json
from exa_py import Exa

exa = Exa()
run_id = "agent_run_01j..."
run = exa.agent.runs.poll_until_finished(
    run_id,
    poll_interval=4000,
)

print(json.dumps(run.model_dump(), indent=2))

import Exa from "exa-js";

const exa = new Exa();
const runId = "agent_run_01j...";
const run = await exa.agent.runs.pollUntilFinished(runId, {
  pollInterval: 4000
});

console.log(JSON.stringify(run, null, 2));

RUN_ID="agent_run_01j..."

while true; do
  RUN_JSON="$(curl -s "https://api.exa.ai/agent/runs/$RUN_ID" \
    -H "x-api-key: $EXA_API_KEY")"

  STATUS="$(echo "$RUN_JSON" | jq -r '.status')"
  echo "status=$STATUS"

  if [ "$STATUS" = "completed" ] || [ "$STATUS" = "failed" ] || [ "$STATUS" = "cancelled" ]; then
    echo "$RUN_JSON" | jq .
    break
  fi

  sleep 4
done

Completed runs include:

output.text: a natural-language answer
output.structured: validated JSON when you provide outputSchema
output.grounding: citations for text or structured fields, when emitted
costDollars: the run’s cost breakdown

Stream events

Streaming keeps the create request open and sends Server-Sent Events (SSE) until the run completes. See Event format for the event types and payloads.

Streaming is mandatory for Zero Data Retention (ZDR) runs. Creating a ZDR run without streaming returns a 400 error.

Set stream=True in Python, stream: true in TypeScript, or send Accept: text/event-stream over HTTP:

from exa_py import Exa

exa = Exa()
events = exa.agent.runs.create(
    query="Find five recently launched developer tools for evaluating AI agents.",
    stream=True,
)

for event in events:
    print(event.event, event.data)

import Exa from "exa-js";

const exa = new Exa();
const events = await exa.agent.runs.create({
  query: "Find five recently launched developer tools for evaluating AI agents.",
  stream: true
});

for await (const event of events) {
  console.log(event.event, event.data);
}

curl -N -X POST "https://api.exa.ai/agent/runs" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -H "x-api-key: $EXA_API_KEY" \
  -d '{
    "query": "Find five recently launched developer tools for evaluating AI agents."
  }'

Event format

Each SSE frame contains an event ID, event name, and JSON payload:

id: 1
event: agent_run.created
data: {"id":"agent_run_01j...","status":"queued","createdAt":"2026-05-07T21:21:52.051Z"}

The stream may also contain comment lines such as : keep-alive. SSE clients ignore comments automatically; custom parsers should do the same.

Event types

Event	`data` payload	How to use it
`agent_run.created`	`{ id, status: "queued", createdAt }`	Save the run ID as soon as the request is accepted.
`agent_run.started`	`{ id, status: "running" }`	Mark the run as actively processing.
`agent_run.completed`	The completed Agent run object	Read the final answer from `data.output.text` or `data.output.structured`, and citations from `data.output.grounding`.
`agent_run.failed`	`{ id, status: "failed", error }`	Surface `error.code` and `error.message`; no completed output is available.
`agent_run.cancelled`	`{ id, status: "cancelled", ... }`	Stop consuming the stream and handle the run as cancelled.

Events associated with the same research step include a callId. It corresponds to item.call_id in tool progress events. Use it to group search traces, sources, and tool progress. Some search-trace descriptions are generated asynchronously and can arrive after the source or tool event they describe, so do not correlate them by arrival order alone. Treat agent_run.source.added as a live preview rather than a complete citation list. The terminal run’s output.grounding is the authoritative grounding output.

Replay stored events

For non-ZDR runs, GET /agent/runs/{id}/events returns stored events as paginated JSON. Send Accept: text/event-stream to replay the stored events as SSE, and Last-Event-ID to skip events your client has already processed:

cURL

curl -N "https://api.exa.ai/agent/runs/agent_run_01j.../events" \
  -H "Accept: text/event-stream" \
  -H "Last-Event-ID: 12" \
  -H "x-api-key: $EXA_API_KEY"

The replay endpoint sends the events stored at request time and then closes; it does not continue following a running run. ZDR runs do not retain events and cannot be replayed. For forward compatibility, ignore event names your application does not recognize and continue until a terminal event arrives.

Return structured JSON

Use outputSchema when you need /agent to return in specific format. When you specify an outputSchema, the returned object will contain an output matching your outputSchema in output.structured. outputSchema supports the JSON Schema specification. To request contact information, describe the desired contact fields in outputSchema. Use standard JSON Schema shapes such as { "type": "string", "format": "email" } for email addresses, { "type": "string", "format": "phone" } for phone numbers, and { "type": "string", "format": "uri" } for URLs. Bound list sizes with maxItems when possible so the maximum contact-enrichment cost is predictable.

import json
from exa_py import Exa

exa = Exa()
run = exa.agent.runs.create(
    query="Find AI infrastructure companies that raised a Series A or B in the last 6 months.",
    effort="auto",
    output_schema={
        "type": "object",
        "properties": {
            "companies": {
                "type": "array",
                "items": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string"},
                        "round": {"type": "string"},
                        "website": {"type": "string"},
                    },
                    "required": ["name", "round"],
                },
            }
        },
        "required": ["companies"],
    },
)
run = exa.agent.runs.poll_until_finished(
    run.id,
)

print(json.dumps(run.output.structured if run.output else None, indent=2))

import Exa from "exa-js";

const exa = new Exa();
const run = await exa.agent.runs.create({
  query:
    "Find AI infrastructure companies that raised a Series A or B in the last 6 months.",
  effort: "auto",
  outputSchema: {
    type: "object",
    properties: {
      companies: {
        type: "array",
        items: {
          type: "object",
          properties: {
            name: { type: "string" },
            round: { type: "string" },
            website: { type: "string" }
          },
          required: ["name", "round"]
        }
      }
    },
    required: ["companies"]
  }
});
const completedRun = await exa.agent.runs.pollUntilFinished(run.id);

console.log(JSON.stringify(completedRun.output?.structured, null, 2));

curl -s -X POST "https://api.exa.ai/agent/runs" \
  -H "Content-Type: application/json" \
  -H "x-api-key: $EXA_API_KEY" \
  -d '{
    "query": "Find AI infrastructure companies that raised a Series A or B in the last 6 months.",
    "effort": "auto",
    "outputSchema": {
      "type": "object",
      "properties": {
        "companies": {
          "type": "array",
          "items": {
            "type": "object",
            "properties": {
              "name": { "type": "string" },
              "round": { "type": "string" },
              "website": { "type": "string" }
            },
            "required": ["name", "round"]
          }
        }
      },
      "required": ["companies"]
    }
  }'

Process input rows

Use input.data when you have an existing set of data that you want to enrich. You can add more fields to each data entity, surface more entities based on the data you bring in, or both. For complete row-enrichment examples, see Agent examples.

Process exclusions

Use input.exclusion to exclude certain entries from being surfaced in the run. In the example below, we want to look for the top 10 cutest animals, but we exclude goats and pandas from the run because we already know how cute they are.

import json
from exa_py import Exa

exa = Exa()
run = exa.agent.runs.create(
    query="Find the top 10 cutest animals. Return each animal's common name and a source URL.",
    input={
        "exclusion": [
            {"animal": "goat"},
            {"animal": "panda"},
        ]
    },
)

print(json.dumps(run.model_dump(), indent=2))

import Exa from "exa-js";

const exa = new Exa();
const run = await exa.agent.runs.create({
  query: "Find the top 10 cutest animals. Return each animal's common name and a source URL.",
  input: {
    exclusion: [
      { animal: "goat" },
      { animal: "panda" }
    ]
  }
});

console.log(JSON.stringify(run, null, 2));

curl -s -X POST "https://api.exa.ai/agent/runs" \
  -H "Content-Type: application/json" \
  -H "x-api-key: $EXA_API_KEY" \
  -d '{
    "query": "Find the top 10 cutest animals. Return each animal'"'"'s common name and a source URL.",
    "input": {
      "exclusion": [
        { "animal": "goat" },
        { "animal": "panda" }
      ]
    }
  }'

Connect data sources

Use dataSources to attach premium data partners to a run. Each entry selects a provider. When a property in your outputSchema references a specific source (e.g., “from Similarweb”), Exa Agent calls the matching provider tool instead of a generic web search.

{
  "dataSources": [
    { "provider": "similarweb" },
    { "provider": "fiber" }
  ]
}

See Exa Connect for the full list of data partners, with examples for each.

Continue from a previous run

Use previousRunId to ask follow-ups to the run’s previous response. Each follow-up starts a new run with its own run ID — the previousRunId is only used to carry over context from the prior run, not reused as the new run’s ID.

import json
from exa_py import Exa

exa = Exa()
run = exa.agent.runs.create(
    query="Narrow that list to companies hiring in San Francisco.",
    previous_run_id="agent_run_01j...",
)

print(json.dumps(run.model_dump(), indent=2))

import Exa from "exa-js";

const exa = new Exa();
const run = await exa.agent.runs.create({
  query: "Narrow that list to companies hiring in San Francisco.",
  previousRunId: "agent_run_01j..."
});

console.log(JSON.stringify(run, null, 2));

curl -s -X POST "https://api.exa.ai/agent/runs" \
  -H "Content-Type: application/json" \
  -H "x-api-key: $EXA_API_KEY" \
  -d '{
    "query": "Narrow that list to companies hiring in San Francisco.",
    "previousRunId": "agent_run_01j..."
  }'

Find a run ID

List recent runs and inspect their statuses:

from exa_py import Exa

exa = Exa()
runs = exa.agent.runs.list(
    limit=10,
)

for run in runs.data:
    query = (run.request or {}).get("query", "")
    print(f"{run.id}\t{run.status}\t{run.created_at}\t{query}")

import Exa from "exa-js";

const exa = new Exa();
const list = await exa.agent.runs.list({
  limit: 10
});

for (const run of list.data) {
  const query = run.request?.query ?? "";
  console.log(`${run.id}\t${run.status}\t${run.createdAt}\t${query}`);
}

curl -s "https://api.exa.ai/agent/runs?limit=10" \
  -H "x-api-key: $EXA_API_KEY" \
  | jq -r '.data[] | "\(.id)\t\(.status)\t\(.createdAt)\t\(.request.query)"'

Pricing

Costs are usage-based and priced by component:

Component	Price
Agent Compute Units	`1 ACU = $0.10`
Search tool calls	`$0.005 / search`

Contact enrichment is separate from the core pricing components above: email contact enrichment is $0.02 / email, and phone number contact enrichment is $0.07 / phone number.

usage.agentComputeUnits measures model computation across the full run. More complex queries, or queries that contain a large input.data field will generally take more reasoning steps and make more tool calls, and will generally consumer more ACUs. Your Agent concurrency limit is one fifth of your account QPS. For pay-as-you-go accounts with default QPS, this means two active Agent runs at a time.

Effort

Use effort to set a cost and reasoning effort preference for a run. Supported values are minimal, low, medium, high, xhigh, and auto; the default is auto. If a fixed effort is set, each run is charged at the following request price:

Effort	Price
`minimal`	`$0.012 / request`
`low`	`$0.025 / request`
`medium`	`$0.10 / request`
`high`	`$0.50 / request`
`xhigh`	`$1.00 / request`

Choosing an effort mode

Fixed effort modes are best when you want predictable per-request cost on standard research tasks. Use auto for variable-scope tasks, especially list building or workflows where the number of entities can vary significantly from request to request.

Effort	Best for	Suggested schema complexity	Runtime expectation
`minimal`	Lowest-cost lookups, very narrow factual tasks, short answers	One or two fields, shallow schema	Cheapest, least exhaustive
`low`	Simple lookups, narrow factual tasks, short answers	A few fields, shallow schema	Fast, light research
`medium`	Default starting point for most standard research tasks	Moderate field count, simple nested objects	Balanced quality/runtime
`high`	Harder research, more citations, stricter completeness	Larger schemas or more nuanced fields	Slower, more thorough
`xhigh`	High-value tasks where completeness matters more than cost/latency	Complex schemas, many fields, difficult verification	Slowest, most thorough
`auto`	Variable-scope work, list building, unknown task difficulty	Flexible; useful when entity count or work required is unknown	Variable

Use medium as the default starting point for standard single-entity research tasks. Move down to low or minimal when cost and latency matter more than completeness. Move up to high or xhigh when the output schema is larger, fields require verification, or the task needs deeper reasoning. Use auto when the task scope is not known ahead of time, such as list building or workflows where one request may return many entities. Runtime varies by query difficulty, schema complexity, and external source availability. Treat effort modes as quality/cost/runtime tradeoffs rather than strict latency guarantees.

Zero Data Retention

Exa Agent supports Zero Data Retention (ZDR). ZDR is enabled per team — contact us to enable it for your account. When ZDR is enabled for your team:

Runs must be created with streaming (Accept: text/event-stream). Consume the run’s output from the live SSE stream — Exa does not retain your query, request, or output, so they cannot be retrieved after the run completes. See Stream events for SDK and cURL examples and the complete event contract.
The OpenAI-compatible Responses endpoints are not available. Use /agent/runs with Accept: text/event-stream.
previousRunId is not available on ZDR runs.
Creating a run without streaming, or with Exa Connect dataSources set, returns a 400 error when ZDR is enabled.

API

Data

MCP

SDKs

Integrations

Admin

Additional Resources

Exa Agent

When to use Exa Agent

Quickstart

1. Install the Exa SDK

2. Set your API key

3. Create a run

4. Poll for completion

Stream events

Event format

Event types

Replay stored events

Return structured JSON

Process input rows

Process exclusions

Connect data sources

Continue from a previous run

Find a run ID

Pricing

Effort

Choosing an effort mode

Zero Data Retention

Next

​When to use Exa Agent

​Quickstart

​1. Install the Exa SDK

​2. Set your API key

​3. Create a run

​4. Poll for completion

​Stream events

​Event format

​Event types

​Replay stored events

​Return structured JSON

​Process input rows

​Process exclusions

​Connect data sources

​Continue from a previous run

​Find a run ID

​Pricing

​Effort

​Choosing an effort mode

​Zero Data Retention

​Next

When to use Exa Agent

Quickstart

1. Install the Exa SDK

2. Set your API key

3. Create a run

4. Poll for completion

Stream events

Event format

Event types

Replay stored events

Return structured JSON

Process input rows

Process exclusions

Connect data sources

Continue from a previous run

Find a run ID

Pricing

Effort

Choosing an effort mode

Zero Data Retention

Next