Add Tools

Add Tools Skill

You are an expert at creating tool configurations for Sim integrations. Your job is to read API documentation and create properly structured tool files.

Your Task

When the user asks you to create tools for a service:

1. Use Context7 or WebFetch to read the service's API documentation
2. Create the tools directory structure
3. Generate properly typed tool configurations

Hard Rule: No Guessed Response Schemas

If the docs do not clearly show the response JSON for a tool, you MUST tell the user exactly which outputs are unknown and stop short of guessing.

• Do NOT invent response field names
• Do NOT infer nested paths from nearby endpoints
• Do NOT guess array item shapes
• Do NOT write transformResponse against unverified payloads

If the response shape is unknown, do one of these instead:

1. Ask the user for sample responses
2. Ask the user for test credentials so you can verify live responses
3. Implement only the endpoints whose outputs are documented
4. Leave the tool unimplemented and explicitly say why

Directory Structure

Create files in apps/sim/tools/{service}/:

tools/{service}/
├── index.ts      # Barrel export
├── types.ts      # Parameter & response types
└── {action}.ts   # Individual tool files (one per operation)

Tool Configuration Structure

Choose the execution boundary first

Every tool must use exactly one of these configurations:

• In-process operation (preferred): use InternalToolConfig when the executor and the implementation run in the same Sim process/trust/runtime plane. Materialize typed operation.input, implement the handler under apps/sim/lib/internal/{service}/execute-tool.ts, and register every tool ID in apps/sim/lib/internal/tool-operations/registry.server.ts.
• External provider request: use ToolConfig.request only when the URL is an absolute external HTTP(S) provider endpoint.

Never set a tool URL to /api/..., construct an absolute URL back to Sim, declare request.internal, add a directExecution property (it fails bun run check:tool-request-boundary), import a route module, or create an API route merely to normalize files, authorize access, or reuse server code. A real browser/API route may remain as a thin adapter, but the route and the tool must call the same operation directly. A true cross-process/capability boundary uses an explicit server client and is not disguised as a tool self-hop.

For protected Sim resources, the internal handler calls the domain's authorized application use case with trusted execution context; use the migrate-application-operation skill.

External provider request

Use this structure only for an absolute external provider API:

import type { {ServiceName}{Action}Params } from '@/tools/{service}/types'
import type { ToolConfig } from '@/tools/types'

interface {ServiceName}{Action}Response {
  success: boolean
  output: {
    // Define output structure here
  }
}

export const {serviceName}{Action}Tool: ToolConfig<
  {ServiceName}{Action}Params,
  {ServiceName}{Action}Response
> = {
  id: '{service}_{action}',           // snake_case, matches tool name
  name: '{Service} {Action}',         // Human readable
  description: 'Brief description',   // One sentence
  version: '1.0.0',

  // OAuth config (if service uses OAuth)
  oauth: {
    required: true,
    provider: '{service}',            // Must match OAuth provider ID
  },

  params: {
    // Hidden params (system-injected, e.g. the OAuth accessToken)
    accessToken: {
      type: 'string',
      required: true,
      visibility: 'hidden',
      description: 'OAuth access token',
    },
    // User-only params (credentials, api key, IDs user must provide)
    someId: {
      type: 'string',
      required: true,
      visibility: 'user-only',
      description: 'The ID of the resource',
    },
    // User-or-LLM params (everything else, can be provided by user OR computed by LLM)
    query: {
      type: 'string',
      required: false,                // Use false for optional
      visibility: 'user-or-llm',
      description: 'Search query',
    },
  },

  request: {
    url: (params) => `https://api.service.com/v1/resource/${params.id}`,
    method: 'POST',
    headers: (params) => ({
      Authorization: `Bearer ${params.accessToken}`,
      'Content-Type': 'application/json',
    }),
    body: (params) => ({
      // Request body - only for POST/PUT/PATCH
      // Trim ID fields to prevent copy-paste whitespace errors:
      // userId: params.userId?.trim(),
    }),
  },

  transformResponse: async (response: Response) => {
    const data = await response.json()
    return {
      success: true,
      output: {
        // Map API response to output
        // Use ?? null for nullable fields
        // Use ?? [] for optional arrays
      },
    }
  },

  outputs: {
    // Define each output field
  },
}

In-process operation

import type { InternalToolConfig } from '@/tools/types'

export const {serviceName}{Action}Tool: InternalToolConfig<
  {ServiceName}{Action}Params,
  {ServiceName}{Action}Response
> = {
  id: '{service}_{action}',
  name: '{Service} {Action}',
  description: 'Brief description',
  version: '1.0.0',
  params: {
    // Same canonical metadata as an external tool.
  },
  operation: {
    input: (params) => ({
      // Map resolved tool params into the typed semantic operation input.
    }),
  },
  outputs: {
    // Define each output field.
  },
}

The registered handler accepts InternalToolOperationCall, validates request.input, uses only trusted request.context for authority, forwards request.signal, and returns the same bounded Response contract expected by the tool executor. It has no URL, method, request headers, fetch fallback, or caller-controlled _context authority.

Critical Rules for Parameters

Visibility Options

• 'hidden' - System-injected (OAuth tokens, internal params). User never sees.
• 'user-only' - User must provide (credentials, api keys, account-specific IDs)
• 'user-or-llm' - User provides OR LLM can compute (search queries, content, filters, most fall into this category)
• 'llm-only' - Computed by the LLM only; never shown as a user field

A required 'hidden' param needs an oauth declaration or hosting.apiKeyParam to supply it (bun run check:tool-param-reachability).

Parameter Types

• 'string' - Text values
• 'number' - Numeric values
• 'boolean' - True/false
• 'json' - Complex objects (NOT 'object', use 'json')
• 'file' - Single file
• 'file[]' - Multiple files

Required vs Optional

• Always explicitly set required: true or required: false
• Optional params should have required: false

Resolved Secrets and Provenance Boundaries

Classify every request field before implementing the tool.

This is opt-in, not a blanket integration migration. Add a model-input declaration only when the service's official documentation or an unambiguous local execution path proves that the exact field is consumed by an AI model. If that cannot be established, preserve existing tool behavior and leave the field unannotated.

• Ordinary provider/API input: leave it unchanged. Explicit {{...}} references resolve and are sent with their normal request semantics. A URL, domain, resource ID, control field, or opaque payload is not model-visible merely because the provider is AI-backed or may process the referenced resource later.
• Text or structured content consumed by an AI model: declare request.modelInput for an external provider request or operation.modelInput for an in-process operation, with mode: 'project' and select only the exact model-visible fields. The shared executor replaces activated Sim secrets with canonical {{NAME}} labels before request formatting. For nested or JSON-string fields, use a small shared selector plus applyProjected; verify that selecting the rebuilt params reproduces the projected selection.
• Serialized model content sent directly to an external provider: include the serialized top-level param in request.modelInput. Project the private copy before the existing request formatter parses it; keep formatter behavior deterministic when a whole-value placeholder is not valid in the serialized grammar. Do not introduce a second hard-rejection path.
• Opaque model input owned by an in-process operation such as inline audio, image, video, or document bytes: add privateInputPaths to the mode: 'project' operation model-input declaration, or use mode: 'private-provenance' with inputPaths when there is no textual projection (see the modelInput union in apps/sim/tools/types.ts). Do not select storage keys, paths, signed URLs, or ordinary remote URLs as byte provenance; the owning operation must authorize stored bytes independently at model egress. The operation must call validateOpaqueModelInputProvenance before downloading or sending content to the model and must apply the workspace-file provenance guard before reading a persisted workspace file.
• Sim-owned durable storage or internal execution handoff that can later enter a workflow/model (table cells, Agent memory, knowledge documents/chunks, workspace-file contents, or child-workflow input): transport encrypted field-scoped provenance with operation.secretProvenance. The operation validates the exact selection and trusted scope, then persists, imports, or propagates it at the owning boundary. Preserve shared legacy behavior for rows/files whose provenance marker is NULL; never invent a tool-local migration rule.

Hard rules:

• Never substitute secret plaintext into source or serialize plaintext provenance.
• Never hand-roll private provenance headers/envelopes; the shared executeTool boundary owns transport and strips private metadata from functional results.
• Never attach private provenance to an external URL. Project proven model-visible external fields with request.modelInput; otherwise preserve ordinary request semantics. Use a registered in-process operation when encrypted provenance must cross the boundary.
• Never sanitize arbitrary third-party tool results. Projection applies only to secrets activated by Sim's resolved-secret provenance for that execution/tool call.
• Do not add provenance merely because a value is persisted, returned by a tool, or appears in a filename. Require a concrete Sim {{...}} resolution path and a later model/log boundary. If an unsupported field can resolve a secret but does not justify durable tracking (for example a file_write path), reject it at that exact ingress.
• At diagnostic boundaries, project only values carrying execution-scoped provenance. Ordinary provider responses, filenames, URLs, and errors remain unchanged when Sim did not resolve a secret into them.

Run the test-audit authoring gate, then cover these risks at the boundary that owns them: named projection, ordinary identical text without provenance, nested and serialized shape handling, unchanged ordinary external inputs, malformed/incomplete private metadata failing closed, headerless legacy requests, and absence of private metadata in the public tool result. For durable sinks, also cover legacy NULL markers, exact-empty new writes, tracked secret writes, stale/missing sidecars, and scope isolation.

Critical Rules for Outputs

File Downloads and Generated Files

Internal operations return createInternalToolFileResult / createInternalToolFilesResult from lib/internal/tool-operations/file-result.ts with bounded Buffers and a callback that places the stored descriptors in the response. Their handlers preserve this result through dispatch, using InternalToolOperationHandler<InternalToolOperationResult>. Do not serialize file bytes as base64 JSON: the executor's 10 MiB response cap runs before ordinary file postprocessing or large-value externalization. The shared executor stores files using trusted run or Copilot ownership.

External endpoints that return raw binary files explicitly declare request.responseType: 'binary' and return output.file with { name, mimeType, data: buffer, size } from transformResponse. The executor applies the bounded file-transfer budget and persists the descriptor. This opt-in is for raw binary responses, not provider JSON containing base64 or tools that fetch attachments later. Keep provider-specific limits and bounded reads; a file declaration is not permission to enlarge arbitrary JSON responses.

Attachment readers that download files inside transformResponse need their own bounded reads: the first response cap does not cover subsequent fetches. Accept ToolResponseContext as the third transform argument, forward its signal, and share one AttachmentDownloadBudget across sequential downloads. Prefer raw provider endpoints over base64 metadata. Return the same file object in the declared file / file[] output and nested message associations; FileToolProcessor stores it once and replaces every alias with the same UserFile in both workflow and Copilot execution.

When a transform receives an already-stored UserFile, return it unchanged (keep id, key, url, context, type, name, size). Building a new { name, mimeType, data, size } object from it discards the stored reference; that shape is only for fresh raw-binary responses. File outputs do not need duplicate inline text/base64 aliases; the file system handles content materialization. When an existing tool explicitly exposes content aliases in its contract, preserve its legacy version and use the existing block/tool version pattern for a file-only output. Test a file over 10 MiB through executor admission, single persistence, trusted ownership, and the unchanged JSON cap. Avoid adding top-level filename, size, MIME type, URL, or success fields that merely repeat the canonical file or tool result; keep additional provider fields only when they convey distinct information.