Foundry API Integrations

Build API integrations from the vendor's published OpenAPI spec, adapted for Foundry, with its authentication scheme configured.

For api-integrations create, always include --description (50 characters max) — the CLI still prompts for it even with --no-prompt if omitted.

Part of a suite. If development-workflow has not already run, and this is a new app or its first capability, load the development-workflow skill first — it owns the CLI prerequisite check, scaffolding order, and manifest coordination.

This skill covers exposing external APIs (third-party services or CrowdStrike Falcon APIs) to the Falcon Foundry platform via OpenAPI/Swagger specifications. These integrations make API operations available to Falcon Fusion SOAR workflows, Foundry UI extensions, Foundry Functions, and other Foundry capabilities.

API integrations are how Foundry manages credentials. There is no secrets system, no encrypted env vars, and no key vault. When you register an API integration, the platform collects credentials at install time and manages tokens automatically. Change them later under App catalog > the app > ⋮ > Modify settings; development-workflow covers integrations added after install. This is why functions MUST call third-party REST APIs through APIIntegrations().execute_command_proxy() — not via raw HTTP with env vars.

For calling Falcon APIs from within Function code, see functions-falcon-api instead.

Decision Tree

What kind of API integration?

External API (Okta, VirusTotal, ServiceNow, etc.)
├── Vendor publishes OpenAPI spec → Download it, adapt for Foundry
└── No vendor spec available      → Write minimal spec as last resort

CrowdStrike Falcon API
├── From functions → use functions-falcon-api instead
├── From workflows → use CrowdStrike auto-auth (no spec needed)
└── API family FalconPy does not wrap
    → API integration with oauth2 clientCredentials against /oauth2/token.
      The Falcon swagger at assets.falcon.*.crowdstrike.com returns AccessDenied
      without console auth, so a minimal spec written from the product docs
      is acceptable here — the one case where hand-writing is not a last resort.

Workflow: Download, Adapt, Register

Always follow this order. Hosts that load this plugin's hooks enforce adaptation (see the hook safety net below). Antigravity CLI and a skill-symlink install MUST run the script explicitly.

1. Download the vendor's spec

NEVER write an OpenAPI spec from scratch when the vendor publishes one. Hand-written specs produce incorrect response schemas, miss required parameters, and lack proper security definitions. Download the vendor's spec even if it has hundreds of endpoints — Foundry handles large specs fine. Do NOT rationalize writing a "focused" or "minimal" spec because the vendor spec is big.

1. Ask the user if they have a local copy or know where to download it

2. Browse the repo first, don't guess URLs. List its YAML and JSON files in one call (HEAD is the default branch), then download the spec with curl.

   gh api "repos/{owner}/{repo}/git/trees/HEAD?recursive=1" \
     --jq '.tree[] | select(.type=="blob" and (.path|test("\\.(ya?ml|json)$"))) | .path'

   Narrow a long list with grep, not head (versioned copies sort first); prefer a current or latest path. Never guess URLs.

3. Search locally for existing specs
4. If the vendor does not publish a spec, only then write a minimal one

Do NOT delegate spec download to Explore agents or subagents. They lack skill context and fall back to Fetch or a browser. Download inline with gh and curl.

For detailed download commands and structural fix patterns, see references/spec-adaptation-examples.md.

2. Adapt the spec for Foundry (MANDATORY)

# <skill-dir> is this SKILL.md's directory. --dry-run previews the fixes without
# saving them, though it still converts a Swagger 2.0 input to .openapi3.*.
python3 <skill-dir>/scripts/adapt_spec_for_foundry.py /tmp/VendorApi.yaml

Dependencies: The script requires pyyaml, and installing the plugin does not install Python packages. A bare pip install fails on Homebrew and system Pythons with externally-managed-environment (PEP 668), so set up a venv once in the checkout and run the script from it:

SKILLS_REPO=/path/to/foundry-skills
python3 -m venv "$SKILLS_REPO/.venv"
"$SKILLS_REPO/.venv/bin/pip" install -r "$SKILLS_REPO/requirements.txt"
"$SKILLS_REPO/.venv/bin/python" "$SKILLS_REPO/skills/api-integrations/scripts/adapt_spec_for_foundry.py" /tmp/VendorApi.yaml

The helper is bundled with this skill at scripts/adapt_spec_for_foundry.py. Resolve that path relative to this SKILL.md, not the app workspace.

The script applies fixes derived from 12 Foundry sample apps:

• Swagger 2.0 conversion: Converts to OpenAPI 3.0 via swagger2openapi (npx) and writes the result next to the input as <name>.openapi3.json or <name>.openapi3.yaml. Register that file, not the original.
• Auth fixing: Deletes an oauth2 scheme (and its security references) when it has authorizationCode and no clientCredentials flow; a scheme with both is kept unchanged. Adds bearerFormat to an apiKey-in-Authorization scheme when its description names the prefix (SSWS, Bearer, Token, Basic). Adds a global security block if missing and removes per-operation security: [] overrides.
• Server URLs: Strips https:// from variable-based URLs (Foundry adds protocol separately). Removes default from variables without enum (prevents locked dropdown).
• Parameter dedup: Removes operation-level parameters that duplicate path-level parameters (prevents Foundry's "items are equal" validation error). When an operation overrides a path-level parameter with a different definition, the operation keeps its version and the path-level one moves into the path's other operations.

Foundry's UI import handles large/complex specs. Don't trim or simplify vendor specs. The auth fixes are what matter.

3. Register with the CLI

foundry api-integrations create --name "VendorApi" --description "Vendor API" --spec /tmp/VendorApi.yaml --no-prompt

Always include --description with api-integrations create. Even with --no-prompt, the CLI still interactively prompts for the optional description if omitted, causing Error: EOF. The value is capped at 50 characters (input must be at most 50 characters long) — shorter than the 500-character limit other artifacts allow.

Done. For most integrations, this is all you need. Validate immediately after registering (foundry apps validate --no-prompt).

Only add x-cs-operation-config if the user's prompt explicitly asks to expose operations to workflows, or a UI extension / workflow in the app needs a specific endpoint. See Expose Operations to Workflows below.

Hook safety net: When the host runs this plugin's hooks (Claude Code, Codex, Copilot CLI, and Cursor), the hook runs adapt_spec_for_foundry.py automatically if step 2 is missed. It blocks a Swagger 2.0 spec instead, because conversion writes a new file: run the script yourself and register the .openapi3 file it prints. Cursor's shell tool is Shell; the hook treats that the same as Bash. Antigravity CLI and other assistants that do not load the hooks must run the script themselves.

Authentication Configuration

API key prefix: apiKey type with name: Authorization and in: header works for APIs that send tokens via the Authorization header. Add bearerFormat to specify the prefix (e.g., SSWS, Bearer, Token) — Foundry reads this field to populate the "API key parameter prefix" in the install UI. The adapt script infers the prefix from the scheme's description automatically.

Credentials go to every operation. Foundry attaches the integration's credentials to every operation it proxies, including ones the spec marks public with security: [] (verified 2026-10-19), and the adapt script removes those overrides. Don't count on an operation-level override to keep a credential away from an endpoint; put that endpoint in a separate integration with no security scheme.

For full vendor-specific auth examples, see references/auth-examples.md.

Adapting Specs for Foundry

Server URL Configuration

Use a fixed base URL when the API domain is the same for all users:

"servers": [{"url": "https://www.virustotal.com"}]

When the domain varies per customer, use a single server variable for the full domain. The Falcon console handles the protocol separately, so the URL must not include https://. The variable needs only a description — no default, no enum:

"servers": [{"url": "{yourDomain}", "variables": {"yourDomain": {"description": "the \"yourDomain\" variable is replaced with a dynamic value at execution time"}}}]

Prefer a single variable for the complete domain. A split such as {instance}.service-now.com (used by foundry-sample-functions-python) works only when users enter just the subdomain; a user who types the full domain gets dev-12345.okta.com.okta.com. A default value without enum renders a dropdown instead of a free-text input.

For a fixed set of hosts, such as the Falcon API regions, use an enum with a default (see auth-examples).

Put setup steps in info.description. It's the text at the top of the install form. The console orders the fields itself and shows raw enum values.

Expose Operations to Workflows

Skip this unless the user asks for it. Most API integrations work without x-cs-operation-config. Only add it when the prompt explicitly mentions sharing operations with Falcon Fusion SOAR workflows, or when a UI extension or workflow in the app needs a specific endpoint.

Add x-cs-operation-config to the specific operations requested:

paths:
  /api/v1/users:
    get:
      operationId: listUsers
      x-cs-operation-config:
        workflow:
          name: listUsers
          description: List all users
          expose_to_workflow: true
          system: false
      summary: List all users

The workflow nesting under x-cs-operation-config is required. A flat expose_to_workflow: true directly under x-cs-operation-config will not work and causes deploy failures.

Exposing an Operation to AI Agents

agent_tools is a sibling of workflow under the same x-cs-operation-config key. Add it when a Foundry AI agent needs to call the operation:

paths:
  /files/{id}:
    get:
      operationId: getFileReport
      x-cs-operation-config:
        agent_tools:
          name: Get_a_file_report
          description: get a file report
          expose_to_agent: true
        workflow:
          name: Get a file report
          description: Get a file report
          system: false

The two blocks are independent — an operation can be exposed to agents, to workflows, to both, or to neither. expose_to_agent must be nested under agent_tools, exactly as expose_to_workflow must be nested under workflow.

The agent must also name the operation in its own tools list, by the tool's platform ID: action:mcp/gce/integration:<integration id>~<hash>. The ID exists only after the integration deploys, so deploy first, look the ID up (AgentWorks Add tools, or the AgentWorks tools API), then add the entry and deploy again. A name-based entry such as api_integrations.<name>.<agent_tools.name> fails the whole app's deploy with a bare 500, so don't add one, even as a placeholder. See ai-agents-development's manifest-schema reference for the lookup.

Exposure without the tools entry (or the reverse) yields an agent that silently cannot call the operation. See ai-agents-development.

For autocomplete dropdown patterns and the HTTP Actions vs. Functions decision framework, see references/spec-adaptation-examples.md.

Calling API Integrations

UI extensions call API integrations through Foundry-JS (sandboxed iframes block arbitrary HTTP requests). This pattern is from foundry-sample-foundryjs-demo:

import FalconApi from '@crowdstrike/foundry-js';
const falcon = new FalconApi();
await falcon.connect();

// Create integration instance
const apiIntegration = falcon.apiIntegration({
  definitionId: 'Okta',        // Matches API integration name in manifest.yml
  operationId: 'listUsers'     // Matches operationId in the OpenAPI spec
});

// Execute — use params.path, params.query, json (not body), or headers
const response = await apiIntegration.execute({
  request: {
    params: { query: { limit: 10 } }
  }
});

// Check for errors first
if (response.errors?.length > 0) {
  console.error('Request failed:', response.errors[0].message);
}

const statusCode = response.resources?.[0]?.status_code;
const body = response.resources?.[0]?.response_body;

For Python (FalconPy), Go (gofalcon), and detailed UI examples, see references/calling-patterns.md.

Efficiency Rules

Target: complete an API integration in under 5 minutes. Download, adapt, register, deploy. That's it.

Let the adapt script handle the spec's auth scheme. It handles auth conversion automatically — it was derived from 12 production Foundry apps. Don't read the spec up front to work out how auth works or reason about apiKey vs http/bearer vs SSWS; run the adapt script and register. The one exception: if the script's output, the import, or a test call shows it got a specific field wrong (auth fields included), patch only that field and tell the user what the script missed. Don't rework the rest of the spec.

NEVER use Read or sed on large spec files. Vendor specs can be 10K-80K+ lines. Reading them into context wastes millions of tokens and slows everything down. Instead:

# Find a specific operationId
grep -n 'operationId: listUsers' /tmp/VendorApi.yaml

# Add x-cs-operation-config to a specific operation (cross-platform)
python3 -c "
import json, sys
spec = json.load(open(sys.argv[1]))
for path in spec.get('paths', {}).values():
    for op in path.values():
        if isinstance(op, dict) and op.get('operationId') == 'listUsers':
            op['x-cs-operation-config'] = {'workflow': {'name': 'listUsers', 'description': 'List all users', 'expose_to_workflow': True, 'system': False}}
json.dump(spec, open(sys.argv[1], 'w'), indent=2)
" /tmp/VendorApi.json

Cross-platform note: Use python3 for spec manipulation instead of sed — it works on macOS, Linux, and Windows without syntax differences.

Don't add what wasn't asked for. If the prompt says "create an API integration for Okta," download the spec, adapt it, register it, and deploy. Don't read the spec to discover operations, don't add x-cs-operation-config, don't lint or trim. Foundry handles large specs fine.

Common Pitfalls

• Reading large spec files into context. NEVER use Read or sed on vendor specs. They can be tens of thousands of lines. Use grep to find line numbers, python3 to patch specific operations.
• Manually analyzing or fixing auth schemes. Trust the adapt script. Do not read the spec to reason about apiKey vs http/bearer vs SSWS. The adapt script handles auth conversion automatically. If it demonstrably gets a field wrong, the exception above applies: patch only that field and tell the user what the script missed.
• Adding x-cs-operation-config when not asked. Skip it unless the user's prompt explicitly mentions workflows or a UI/workflow needs a specific endpoint. Most integrations work without it.
• Writing specs from scratch when the vendor publishes one. Hand-written specs miss edge cases.
• Running spec linters before importing. Foundry's import handles vendor specs with lint errors. Linting wastes time and tempts trimming.
• Trimming vendor specs. Keep the full spec. Foundry handles large specs and unused operations gracefully.
• Skipping adapt_spec_for_foundry.py. Only hosts that load this plugin's hooks (Claude Code, Codex, Copilot CLI, and Cursor plugin installs) run it automatically. Antigravity CLI and skill-symlink installs must invoke the bundled helper. The script converts unsupported auth schemes and fixes server URLs that would otherwise block saving in the Falcon console.
• Including https:// in server URLs with variables. The Falcon console adds the protocol separately.
• Adding default to server variables for dynamic domains. This renders a dropdown instead of a text field.
• Splitting domains into {subdomain}.vendor.com without saying in the variable's description that users enter only the subdomain.
• Using oauth2 authorizationCode flow. Foundry only supports clientCredentials. The adapt script deletes an oauth2 scheme that has authorizationCode and no clientCredentials; a scheme with both flows is left for you to trim.
• Passing query values as strings or string arrays when the spec says integer. The platform validates the request against the spec's parameter schemas before proxying, so {"params": {"query": {"limit": ["1"]}}} fails with 400 request failed schema validation: /properties/params/properties/query/.... Send scalars typed per the spec: {"params": {"query": {"limit": 1}}}. See references/calling-patterns.md.
• A --description over 50 characters. api-integrations create rejects it with input must be at most 50 characters long.

Reading Guide

Use Cases

For real-world implementation patterns, see:

• use-cases/http-actions.md — HTTP Request actions vs API integrations
• use-cases/greynoise-deep-dive.md — End-to-end third-party API app
• use-cases/custom-soar-actions.md — Custom Falcon Fusion SOAR actions

Reference Implementations

• foundry-sample-foundryjs-demo: JSONPlaceholder (no auth, static URL, x-cs-operation-config)
• foundry-sample-anomali-threatstream: Anomali (type: http/scheme: bearer, dynamic URL)
• foundry-sample-functions-python: ServiceNow (type: http/scheme: basic, dynamic URL)
• foundry-sample-insider-risk-workday: Workday (type: http/scheme: basic + custom labels)
• foundry-sample-insider-risk-sailpoint: SailPoint (type: oauth2/clientCredentials)
• foundry-sample-servicenow-itsm: ServiceNow ITSM (dual auth: basic + oauth2)
• foundry-sample-openrouter-toolkit: OpenRouter (type: apiKey, static URL)
• See also: Build API Integrations with Falcon Fusion SOAR HTTP Actions and Technical Deep Dive with GreyNoise