Core
The general-purpose building blocks every agent can use — run code, call HTTP APIs, prompt a model, schedule future work, and read file contents — with no connector required.
AutessaScript TOOLAUTESSA_SCRIPT
Runs an AutessaScript program in a hardened JavaScript sandbox — no file system, network sockets, threads, or process access; the script works through the platform bindings instead. All runtime variables (the agent's arguments plus environment variables such as AUTESSA_COMPANY_ID) are available to the script, making this the type to reach for when a task needs logic, branching, or several platform calls chained in one tool. supportedLanguage currently has a single value, AUTESSASCRIPT.
The script's final response string is returned to the agent verbatim; console output and errors are captured in the execution's debugging info.
Example
{
"toolType": "AUTESSA_SCRIPT",
"supportedLanguage": "AUTESSASCRIPT",
"script": "const orderId = '@@{ORDER_ID}';\nconsole.log('Processing order ' + orderId);\nJSON.stringify({ orderId: orderId, status: 'processed' });"
}
Configuration fields
- Name
supportedLanguage- Type
- string (enum)
- Description
- Required string (enum) value.
- Name
script- Type
- string
- Description
- Required string value.
API Call TOOLAPI_CALL
Makes an HTTP request to an external API. Only GET, POST, and PATCH are allowed, uriParams are URL-encoded onto the query string, and a JSON Content-Type is added automatically when a body is present. Every URL is validated before the request is sent — requests to private networks, internal infrastructure, and non-HTTP(S) schemes are rejected with a validation error — and redirects (up to 5) are re-validated hop by hop. Requests time out after 60 seconds.
The result is {"statusCode": ..., "body": ..., "headers": {...}} with the raw response body as a string for the agent to interpret.
Example
{
"toolType": "API_CALL",
"url": "https://api.example.com/v1/orders",
"method": "POST",
"headers": { "Authorization": "Bearer @@{EXAMPLE_API_TOKEN}" },
"uriParams": { "expand": "customer" },
"body": "{\"orderId\": \"@@{ORDER_ID}\", \"note\": \"@@{NOTE}\"}"
}
Configuration fields
- Name
url- Type
- string
- Description
- Required string value.
- Name
method- Type
- string
- Description
- Required string value.
- Name
headers- Type
- object
- Description
- Required object value.
- Name
uriParams- Type
- object
- Description
- Required object value.
- Name
body- Type
- string
- Description
- Required string value.
LLM Call TOOLLLM_CALL
Sends promptTemplate — with its @@{...} placeholders already resolved — to a model as a single prompt and returns the completion. model picks a capability tier by alias rather than a vendor model name: SMALL_INSTRUCTION_FOLLOWING (the default), BIG_INSTRUCTION_FOLLOWING, BIG_INSTRUCTION_FOLLOWING_REASONING, VIBE_CODING, AUTESSA_120B. Calls stream and consume LLM credits — an LLM Call invoked outside a conversation (from a trigger or scheduled event) is metered as its own AI task.
imageUrl supplies an image for a vision call: pass a presigned https:// URL or a data: URL (an s3:// URI is presigned automatically), and the image is sent alongside the prompt with a vision-capable model such as AUTESSA_120B, so an image-only document with no extractable text can still be reasoned over. Calls time out after 120 seconds by default; raise the ceiling with timeoutSeconds for long generations run outside a conversation.
The result is {"success": true, "content": ..., "usage": {"promptTokens": ..., "completionTokens": ..., "totalTokens": ...}}.
Example
{
"toolType": "LLM_CALL",
"promptTemplate": "Summarize the following support ticket in three bullet points:\n\n@@{TICKET_BODY}",
"model": "SMALL_INSTRUCTION_FOLLOWING"
}
Configuration fields
- Name
promptTemplate- Type
- string
- Description
- Required string value.
- Name
model- Type
- object
- Description
- Required object value.
- Name
timeoutSeconds- Type
- integer
- Description
- Per-call timeout (seconds). Null = current default behavior (120s). Capped at 21600 (6h) so long jobs on the WORKER runtime aren't cut off.
- Name
maxOutputTokens- Type
- integer
- Description
- Generation cap for this tool's completions. Null = omit, leaving the served model's own default — which for a self-hosted vLLM checkpoint can be effectively unbounded. A prompt tool that expects a short structured answer should set this, so a runaway completion cannot keep generating for minutes and hold the call open until its timeout.
- Name
thinking- Type
- string
- Description
- Reasoning toggle for self-hosted vLLM models: `"ON"` / `"OFF"`, same values as the SIMPLE-agent row field. Null = omit `chat_template_kwargs` entirely, i.e. the server's `--default-chat-template-kwargs` decides — exactly the pre-existing behavior, so every tool that does not set this is byte-identical on the wire. Ignored (not an error) on hosted models, mirroring the agent path's gate.
- Name
imageUrl- Type
- string · optional
- Description
- Optional image for a VISION call. When set (a presigned https URL or a `data:` URL), the user message carries an `image_url` content block alongside the prompt text — pair with a vision-capable `#model` (e.g. `AutessaModel#AUTESSA_120B`) so an image document with no extractable text can still be judged. Unset/blank preserves the existing text-only behavior exactly.
- Name
imageUrls- Type
- string · optional
- Description
- Optional MULTI-image payload for a VISION call: a JSON ARRAY of url strings (e.g. `["s3://b/p1.jpg","s3://b/p2.jpg"]`), each an `s3://` URI, a presigned https URL, or a `data:` URL. Every entry becomes its own `image_url` content block after the prompt text, in array order. Merges with (does not replace) the singular `#imageUrl`: that one goes first, then these. A hard cap of 8 images total is applied — the vLLM serving config runs `--limit-mm-per-prompt {"image": 8`}, so a 9th block would be rejected by the model server. Extras are trimmed with a warning rather than failing the call, and a malformed array is warned about and treated as empty: a judgment with fewer images than hoped must degrade, not die. Held as a String rather than a `List` on purpose — it is the shape a templated `@@{...`} variable produces, and it keeps substitution identical to `#imageUrl`.
Schedule Event TOOLSCHEDULE_EVENT
Creates a scheduled event that fires in the future — this is how an agent sets up its own follow-ups and recurring jobs. scheduleSpec.pattern picks the timing style: EVERY ({"every": {"value": 30, "unit": "MINUTES"}}, also HOURS/DAYS), CRON ({"cron": {"expression": "0 9 ? * 2 *", "timezone": "America/Los_Angeles"}}, a 6-field AWS cron in an IANA timezone), or ONCE ({"once": {"at": "2026-07-04T09:00:00", "timezone": "America/Los_Angeles"}}, a future local time). actionType picks what fires — INVOKE_AGENT (runs an agent with an input, via agentConfig), INVOKE_TOOL (executes a tool with static parameters, via toolConfig), or UPDATE_AUTESSADB (inserts or updates rows, via autessaDbConfig) — and only the matching config object is read. Passing an existing eventId updates that event instead of creating a new one; the event's name and description are generated automatically from the creating agent's context.
The result includes the handle you need for later edits: {"success": true, "eventId": ..., "scheduleArn": ..., "eventName": ..., "scheduleExpression": ..., "actionType": ..., "enabled": ...}.
Example
{
"toolType": "SCHEDULE_EVENT",
"enabled": true,
"scheduleSpec": {
"pattern": "CRON",
"cron": { "expression": "0 9 ? * 2 *", "timezone": "America/Los_Angeles" }
},
"actionType": "INVOKE_AGENT",
"agentConfig": {
"agentExecutionRequest": {
"agentId": "V1StGXR8Z5",
"input": "@@{WEEKLY_DIGEST_INSTRUCTIONS}"
},
"continueOnError": false,
"description": "Monday morning digest run"
}
}
Configuration fields
- Name
eventId- Type
- string · optional
- Description
- Optional: Event ID for updating an existing scheduled event. If null, a new event is created.
- Name
eventName- Type
- string · optional
- Description
- Optional: display name for the created event. Supports `@@{var`} placeholders resolved at invocation time. When blank, the executor auto-generates a name from the invoking agent/tool context.
- Name
enabled- Type
- boolean · optional
- Description
- Optional: whether the event is enabled. Defaults to true.
- Name
scheduleSpec- Type
- object
- Description
- Required: schedule specification (EVERY / CRON / ONCE).
- Name
actionType- Type
- string (enum)
- Description
- Required: which action type fires when the event triggers.
- Name
agentConfig- Type
- object
- Description
- Required when actionType = INVOKE_AGENT.
- Name
toolConfig- Type
- object
- Description
- Required when actionType = INVOKE_TOOL.
- Name
autessaDbConfig- Type
- object
- Description
- Required when actionType = UPDATE_AUTESSADB.
Manager Signal TOOLMANAGER_SIGNAL
A control signal for sub-agents working under a manager: calling a signal tool (for example "Complete" or "Help") tells the manager the sub-agent has finished its task or needs assistance. The signal is intercepted by the manager's invocation layer rather than executed like a normal tool — the tool's name and description carry all the meaning, and the config has no fields. If one is ever executed directly, it acknowledges and does nothing.
Example
{
"toolType": "MANAGER_SIGNAL"
}
Configuration fields
This tool type has no configuration fields beyond toolType.
Get File Content TOOLGET_FILE_CONTENT
Extracts the text content of a file so the agent can read it — PDFs and documents become text rather than raw bytes. url accepts an s3:// URI (for files in Autessa storage, presigned automatically) or an existing https:// URL; external URLs pass SSRF validation first. Extraction runs server-side with a 5-minute budget for large files.
The result is {"content": ...} with the extracted text. To view an image or pass a file to the model as an attachment instead of text, use Fetch and View.
Example
{
"toolType": "GET_FILE_CONTENT",
"url": "@@{FILE_URL}"
}
Configuration fields
- Name
url- Type
- string
- Description
- Required string value.
Get File Content from Table TOOLGET_FILE_CONTENT_FROM_TABLE
Extracts text from a file stored in an AutessaDB row: it looks up rowId in the table, reads the file in fileColumnName (which must be a FILE, IMAGE, VIDEO, or AUDIO column), and runs the same server-side extraction as Get File Content — without the agent ever handling a URL.
The result is {"success": true, "rowId": ..., "filePath": ..., "content": ..., "contentLength": ...}.
Example
{
"toolType": "GET_FILE_CONTENT_FROM_TABLE",
"tableId": "V1StGXR8Z5",
"rowId": "@@{ROW_ID}",
"fileColumnName": "contract_pdf"
}
Configuration fields
- Name
tableId- Type
- string (ID)
- Description
- Required string (ID) value. References an AutessaDB table by ID.
- Name
rowId- Type
- integer
- Description
- Required integer value.
- Name
fileColumnName- Type
- string
- Description
- Required string value.