Built-in tools reference

Use this reference to configure the nine built-in tools by their stable names. Every tool validates its input before execution; fields not listed in an input schema are rejected.

Common behavior

All built-ins are enabled system tools and do not require tool-level authentication. Eight are classified as READ. mock_task_creator is the only WRITE tool and requires human approval.

The registry reports four common failures:

calculator

Performs deterministic arithmetic. Category: COMPUTE. Permission: READ. Approval: not required.

Input

Output

Returns action (string), expression (string), and result (number). Results are rounded to remove floating-point noise.

{ "action": "percentage", "a": 15, "b": 240 }

The result is 36, with an expression describing the calculation. Validation rejects unknown actions, missing operands, division by zero, and the square root of a negative number.

document_search

Searches ingested documents and knowledge bases, with a PostgreSQL system-history fallback. Category: SEARCH. Permission: READ. Approval: not required.

Input

Output

Returns query, total, and results. Each result includes title, snippet, and relevance from 0 to 1. The runtime can also identify the result source.

{ "query": "support escalation policy", "limit": 5 }

An empty results array is a successful search with no matches. Validation rejects an empty or oversized query and an out-of-range limit. A database or retrieval failure can surface as an execution failure; check the tool health state.

record_lookup

Looks up supported records through NocoDB when configured, otherwise through PostgreSQL-backed history. Category: DATA. Permission: READ. Approval: not required.

Input

Output

Returns entity, query, count, and a records array of objects.

{ "entity": "orders", "id": "ORD-1042" }

Validation rejects an unsupported entity, blank identifier or search term, and an invalid limit. When NocoDB is selected but the entity table is not configured, the call fails with a configuration error. Connection and query failures surface as execution failures.

mock_task_creator

Creates a task in PostgreSQL. Category: TASK. Permission: WRITE. Approval: required.

Input

Output

Returns taskId, status, title, priority, and createdAt. When supplied, the response also includes description and dueDate. A newly created task has status TODO.

{
  "title": "Review support escalation",
  "description": "Confirm the escalation owner for ticket SUP-1042.",
  "priority": "high",
  "dueDate": "2026-10-02"
}

The action must pass its human approval gate before execution. Validation rejects unsupported priorities, malformed dates, and length violations. Database write failures surface as execution failures. Because the action persists data, confirm whether a task already exists before retrying an uncertain result.

ai_extraction

Extracts requested fields from unstructured text. Category: DATA. Permission: READ. Approval: not required.

Input

Output

Returns extractedData, confidenceScore, sourceDocument, extractedFieldCount, and method. method is llm or regex_heuristic_fallback; fallback output can also include fallbackReason. Missing extracted values can be null.

{
  "text": "Ticket SUP-1042, priority: high, owner: Casey",
  "fieldsToExtract": ["ticket", "priority", "owner"],
  "sourceDocument": "support-note"
}

Validation rejects empty text or an empty field list. If the language model provider is missing, unavailable, or returns malformed data, the tool uses a regex heuristic instead of failing the call. Check method, confidenceScore, and null values before using extracted data in a consequential action.

ai_classification

Classifies text into one of the supplied categories. Category: DATA. Permission: READ. Approval: not required.

Input

Output

Returns assignedCategory, confidence, reasoning, and categoryScores. The assigned value is one of the candidate categories.

{
  "input": "Customer cannot access the account after resetting a password.",
  "categories": ["ACCOUNT_ACCESS", "BILLING", "GENERAL"],
  "context": "Route support tickets to the correct queue."
}

Validation rejects empty input or an empty category list. If the language model call fails, the tool uses a keyword heuristic and prefixes reasoning with a fallback disclosure. Treat the fallback and low-confidence results as signals for review, not as execution failures.

deterministic_condition

Evaluates one explicit condition and identifies a workflow branch. Category: COMPUTE. Permission: READ. Approval: not required.

Input

Output

Returns conditionMet, decisionExplanation, selectedBranch, and evaluationDetails. selectedBranch is TRUE_BRANCH or FALSE_BRANCH.

{
  "field": "priority",
  "operator": "in",
  "actualValue": "high",
  "threshold": ["high", "urgent"]
}

Validation rejects missing fields and unsupported operators. Equality and containment are case-insensitive string comparisons. Relational operators compare numerically when both values parse as numbers; otherwise they use string comparison. For in and not_in, threshold can be an array or a comma-separated value.

final_report

Builds a structured Markdown report from workflow results. Category: TASK. Permission: READ. Approval: not required.

Input

Output

Returns reportId, title, status, executiveSummary, formattedMarkdown, generatedAt, and includedStepsCount.

{
  "title": "Support triage report",
  "summary": "One high-priority access issue requires review.",
  "stepResults": { "classification": { "category": "ACCOUNT_ACCESS" } },
  "recommendation": "Assign the ticket to identity support.",
  "status": "FLAGGED"
}

Validation rejects a blank title or summary, an invalid status, and unrecognized fields. The generated report ID and timestamp vary on every successful call.

code_execution

Runs JavaScript or Python in an isolated sandbox for data processing and algorithmic work. Category: COMPUTE. Permission: READ. Approval: not required.

Input

Output

Returns language, stdout, stderr, exitCode, durationMs, and timedOut. Output is limited to 100 KB by the sandbox implementation.

{
  "language": "javascript",
  "code": "const tickets = [3, 5, 2]; console.log(tickets.reduce((a, b) => a + b, 0));",
  "timeout": 5000
}

Validation rejects unsupported languages, empty or oversized code, and a timeout outside the allowed range. A program error is represented by stderr and a nonzero exitCode; a sandbox timeout sets timedOut. Sandbox startup failures surface as execution failures.

Next steps

Use Navigate the Tool Registry to inspect live health and usage. Before attaching a write-capable tool, review Configure skill permissions and Review and cancel approval requests.