Skip to content
WP EngineDocumentation

Create an agent

POST
/v1/agents
<?php


$client = new \GuzzleHttp\Client();


$response = $client->request('POST', 'https://api.ai.wpengine.com/v1/agents', [
  'body' => '{ "name": "Site Concierge", "model": "anthropic/claude-sonnet-4-6", "description": "example", "system_prompt": "example", "project_id": "example", "temperature": 1, "max_tokens": 1, "reasoning_effort": "example", "scopes": [ "example" ], "settings": {}, "mcp_servers": [ {} ], "tools": [ { "type": "custom", "name": "example", "description": "example", "input_schema": {}, "label": "example", "defer_loading": false } ], "skills": [ {} ], "metadata": { "additionalProperty": "example" } }',
  'headers' => [
    'Authorization' => 'Bearer <token>',
    'Content-Type' => 'application/json',
  ],
]);


echo $response->getBody();

Creates a managed agent under project_id, which is required and must be a project the caller is authorized for. The agent stores the model, optional system prompt, inference parameters, knowledge-base collections, MCP servers, tools, skills, scopes, and feature settings (e.g. disable_kb_rag). The model must support the tools capability; models without this capability return HTTP 400. Accepts a service-token bearer authenticating a site (the project must belong to the site’s account) or an API key scoped to project_id — an API key’s own project must match exactly. Multiple agents in the same project may share a display name; identity is keyed by id.

Authorizations

Section titled “Authorizations”

Request Bodyrequired

Section titled “Request Bodyrequired”
Media typeapplication/json
object
name
required

Human-readable agent name.

string
<= 255 characters
Example
Site Concierge
model
required

Identifier of the model that powers the agent (e.g. anthropic/claude-sonnet-4-6).

string
<= 255 characters
Example
anthropic/claude-sonnet-4-6
description

Free-form description of what the agent does.

string
<= 4096 characters
system_prompt

User prompt template (site prompt) prepended to every conversation powered by this agent.

string
<= 32768 characters
project_id
required

The project the agent belongs to. Required — the caller must always supply it explicitly. For a service-token caller authenticating a site, the project must be owned by the same account that owns the site. For an API key, it must match the key’s own project exactly.

string
temperature

Sampling temperature applied to chat completions.

number format: float
<= 2
max_tokens

Hard upper bound on output tokens per agent turn, forwarded to the model provider as max_tokens on each turn. Generation stops when the limit is reached; the response may be truncated mid-content rather than rejected. On reasoning-capable models (see capabilities on GET /v1/models), internal reasoning tokens may count against this budget alongside visible output. Omitted values leave the limit unset so provider-specific defaults apply (Anthropic defaults to 8192 when unset; OpenAI-compatible routes omit the field). Compare with the model’s max_output_tokens on GET /v1/models, which is the catalog ceiling for a single response. Values above the catalog-wide maximum are rejected with 400. When a turn exhausts this budget, session.status_idle carries stop_reason max_tokens.

integer
>= 1 <= 200000
reasoning_effort

Catalog-native reasoning effort for this agent’s model. Allowed values are the model’s parameters.reasoning_effort on GET /v1/models. Omit to use the provider default. Models that do not publish an enum reject any non-empty value.

string
scopes

Permission scopes the agent operates under. Must be drawn from the site’s existing scopes (e.g. chat, search).

Array<string>
settings

Feature toggles for the agent (e.g. disable_kb_rag: true to skip knowledge-base retrieval). Capped at 32 keys and 4 KiB serialized.

object
<= 32 properties
key
additional properties
any
mcp_servers

MCP server definitions the agent may connect to.

Array<object>
<= 20 items
object
key
additional properties
any
tools

Tool declarations available to the agent. See AgentTool for the per-entry contract.

Array
<= 128 items
One of: discriminator: type
object
type

Discriminator for the tools[] union. Omit for a plain client tool (the default); custom is the only value this schema accepts.

string
Allowed values: custom
name
required

Unique tool name. Must be unique within the agent’s tools array (across every entry, regardless of type).

string
<= 64 characters /^[a-zA-Z0-9_-]{1,64}$/
description

Human-readable description of what the tool does.

string
input_schema

JSON Schema defining the tool’s input parameters.

object
label

Optional short human-readable name for the tool, used alongside name and description when the gateway matches a search query against deferred tools. Supply it when the tool’s name is a slug whose wording a natural-language query would not contain.

string
<= 64 characters
defer_loading

When true, the tool is declared but kept out of the model’s tool list until it is needed, so a large catalogue costs no context up front. The gateway offers the agent a search over the deferred tools and loads the matches, and a deferred tool the model names directly is loaded on demand. Once loaded, its input is enforced strictly against its schema — unlike a non-deferred tool, which keeps today’s lenient handling — so a malformed call is rejected with a model-readable error instead of failing downstream. Leave false (the default) for tools that must be callable from the first turn.

boolean
skills

Skills enabled for the agent.

Array<object>
<= 64 items
object
key
additional properties
any
metadata

Arbitrary string key-value metadata stored alongside the agent. The keys “scopes” and “settings” are reserved.

object
<= 16 properties
key
additional properties
string
<= 512 characters

Responses

Section titled “Responses”

201

Section titled “201”

The created agent

Media typeapplication/json
object
id
required

Unique identifier for the agent.

string
type
required
string
Allowed values: agent
project_id
required

The project the agent belongs to.

string
name
required
string
model
required
string
description
string
system_prompt

User prompt template (site prompt).

string
temperature
number format: float
max_tokens

Output-token limit per turn (see CreateAgentRequest.max_tokens). Omitted when unset on the agent.

integer
reasoning_effort

Catalog-native reasoning effort stored on the agent. Omitted when unset. Allowed values follow the agent’s current model on GET /v1/models.

string
scopes

Permission scopes the agent operates under (round-trips the request value).

Array<string>
settings

Feature toggles for the agent (round-trips the request value).

object
key
additional properties
any
mcp_servers
Array<object>
object
key
additional properties
any
tools

Tool declarations (client tools, knowledge_base grounding entries, and echo/random_number evaluation tool opt-ins) round-tripped from the request.

Array
One of: discriminator: type
object
type

Discriminator for the tools[] union. Omit for a plain client tool (the default); custom is the only value this schema accepts.

string
Allowed values: custom
name

Unique tool name. Must be unique within the agent’s tools array (across every entry, regardless of type). Always present for tools created since name validation was added; may be absent on an agent whose tools were stored before that validation existed.

string
<= 64 characters /^[a-zA-Z0-9_-]{1,64}$/
description

Human-readable description of what the tool does.

string
input_schema

JSON Schema defining the tool’s input parameters.

object
label

Optional short human-readable name for the tool, used alongside name and description when the gateway matches a search query against deferred tools. Supply it when the tool’s name is a slug whose wording a natural-language query would not contain.

string
<= 64 characters
defer_loading

When true, the tool is declared but kept out of the model’s tool list until it is needed, so a large catalogue costs no context up front. The gateway offers the agent a search over the deferred tools and loads the matches, and a deferred tool the model names directly is loaded on demand. Once loaded, its input is enforced strictly against its schema — unlike a non-deferred tool, which keeps today’s lenient handling — so a malformed call is rejected with a model-readable error instead of failing downstream. Leave false (the default) for tools that must be callable from the first turn.

boolean
skills
Array<object>
object
key
additional properties
any
metadata

User-supplied string metadata. Reserved keys (“scopes”, “settings”) are surfaced on their own top-level fields rather than appearing here.

object
key
additional properties
string
<= 512 characters
created_at
required
string format: date-time
updated_at
string format: date-time
nullable
Example
{
  "type": "agent",
  "tools": [
    {
      "type": "custom",
      "defer_loading": false
    }
  ]
}

400

Section titled “400”

Invalid request body or parameters

Media typeapplication/json
object
error
required
object
message
required
string
type
required
string
code
required
integer
request_id
required

Correlation ID for this request, matching the X-Request-ID response header — a 32-character trace ID behind the load balancer, otherwise a UUID. Quote it in support reports.

string
Example
{
  "error": {
    "type": "invalid_request_error",
    "request_id": "80f1e2c3a4b5c6d7e8f9a0b1c2d3e4f5"
  }
}

401

Section titled “401”

Missing or invalid bearer token

Media typeapplication/json
object
error
required
object
message
required
string
type
required
string
code
required
integer
request_id
required

Correlation ID for this request, matching the X-Request-ID response header — a 32-character trace ID behind the load balancer, otherwise a UUID. Quote it in support reports.

string
Example
{
  "error": {
    "type": "invalid_request_error",
    "request_id": "80f1e2c3a4b5c6d7e8f9a0b1c2d3e4f5"
  }
}

Headers

Section titled “Headers”
WWW-Authenticate
string
Example
Bearer realm="ai-services"

403

Section titled “403”

The token is valid, but the caller is not permitted to perform this action — either its scope/permission is insufficient or the authenticated user is not authorized for the target account (error type “permission_error”). On /v1/chat/completions specifically, a 403 with error type “content_policy_violation” instead means the request content was blocked by content moderation policy; that response never identifies which filter or category matched.

Media typeapplication/json
object
error
required
object
message
required
string
type
required
string
code
required
integer
request_id
required

Correlation ID for this request, matching the X-Request-ID response header — a 32-character trace ID behind the load balancer, otherwise a UUID. Quote it in support reports.

string
Example
{
  "error": {
    "type": "invalid_request_error",
    "request_id": "80f1e2c3a4b5c6d7e8f9a0b1c2d3e4f5"
  }
}

429

Section titled “429”

Rate limit exceeded. The request was throttled at the edge; wait and retry later. The response body is a short HTML page generated by the load balancer, not a structured JSON error — clients should rely on the 429 status code rather than parsing the body.

Media typetext/html
string

500

Section titled “500”

The server encountered an unexpected error while handling the request. Safe to retry once; if it persists, the cause is server-side and the request payload was not the problem.

Media typeapplication/json
object
error
required
object
message
required
string
type
required
string
code
required
integer
request_id
required

Correlation ID for this request, matching the X-Request-ID response header — a 32-character trace ID behind the load balancer, otherwise a UUID. Quote it in support reports.

string
Example
{
  "error": {
    "type": "invalid_request_error",
    "request_id": "80f1e2c3a4b5c6d7e8f9a0b1c2d3e4f5"
  }
}

503

Section titled “503”

A required dependency is temporarily unavailable. Retry after a short delay with exponential back-off.

Media typeapplication/json
object
error
required
object
message
required
string
type
required
string
code
required
integer
request_id
required

Correlation ID for this request, matching the X-Request-ID response header — a 32-character trace ID behind the load balancer, otherwise a UUID. Quote it in support reports.

string
Example
{
  "error": {
    "type": "invalid_request_error",
    "request_id": "80f1e2c3a4b5c6d7e8f9a0b1c2d3e4f5"
  }
}