← Search

Configure Celigo AI agents and guardrails
AI agents 🧪 Install ran in our test sandbox · Nov 11, 2026 📌 Pinned version: commit e8c094f
Install test only, on the date shown. This does not guarantee the code is safe.
🔎 Install prompt analysis · 2026-10-22 : uses API keys or tokens
⚙ Needs: a Celigo platform account (paid iPaaS…

Configure Celigo AI agents and guardrails

Build LLM-powered import steps and safety guardrails in Celigo flows — provider choice (OpenAI/Gemini), prompt design, structured output, BYOK, PII and moderation guardrails

At a glance
What
Build LLM-powered import steps and safety guardrails in Celigo flows — provider choice (OpenAI/Gemini), prompt design, structured output, BYOK, PII and moderation guardrails
Cost
Free
Needs
a Celigo platform account (paid iPaaS — AI-agent usage on Celigo AI counts against the account's monthly AI token quota); optionally your own provider API key for BYOK (mandatory for the catalog providers) — the skill is guidance the agent follows, not software
Install
Copy the installer prompt below into your Muse — your agent does the rest.

Version:

Quick answer

Configure Celigo AI agents and guardrails is a free Muse skill that builds LLM-powered import steps and safety guardrails in Celigo flows, covering provider choice between OpenAI and Gemini, prompt design, structured output, BYOK, and PII and moderation guardrails, so imported data flows through checks before it lands in your systems.

⌁

Install

Copy the install package below, then paste it into Muse🧪 Install ran in our test sandbox · Nov 11, 2026📌 Pinned version: commit e8c094f

Step 0, before installing anything: read this install prompt, summarize in a few lines what this build does and what installing it involves, then ask me to confirm before you proceed. Prerequisites: a Celigo platform account (paid iPaaS — AI-agent usage on Celigo AI counts against the account's monthly AI token quota); optionally your own provider API key for BYOK (mandatory for the catalog providers) — the skill is guidance the agent follows, not software Install "Configure Celigo AI agents and guardrails" for me. It gives my agent @celigo's workflow for building LLM-powered import steps (AI agents) and guardrails in Celigo flows: the two provider families (built-in providers — OpenAI via the Responses API, and Gemini and Anthropic Claude via the LiteLLM proxy — and the BYOK-only catalog providers Mistral, xAI, Hugging Face, DeepSeek, Cohere and Groq), prompt design, structured output (json_schema/text/blob), tool use (web search, MCP tools, prompts and resources, Celigo Tools, image generation, code execution on BYOK), one-to-many fan-out over array items, response mapping, conversation history via a stable identifier, the Celigo AI vs BYOK decision, a capability check before building, CLI commands (celigo ai-agents, celigo guardrails), a pre-submit checklist, gotchas and a common-errors table, by @celigo, MIT-licensed. Repository: https://raw.githubusercontent.com/celigo/ai/e8c094fcc8cf38f8aed63f253d7b8d46a6c7fb24/skills/configuring-ai-agents/SKILL.md (the skill is the whole skills/configuring-ai-agents/ folder at that commit: SKILL.md plus its references/ subfolder, with 36 YAML schema files across references/schemas/, references/common-schemas/ and references/guardrail-schemas/) Pinned version: install the code at commit e8c094fcc8cf38f8aed63f253d7b8d46a6c7fb24 exactly (Skill Harbor tested this version). Do not substitute a newer version. If this commit is unreachable, stop and tell me. Updated October 22, 2026: Skill Harbor regenerated this install prompt from the current upstream file (SKILL.md, 28,861 → 35,553 characters). One correction reverses what the previous prompt said, and several capabilities are new; the previous prompt is corrected wherever the file now says otherwise. What changed: (1) CORRECTION — the hosted-model check moved to save time. The previous prompt said an unhosted model on Celigo AI saves and its first run fails with 422 invalid_model, whose message lists the hosted models. The file now says the opposite: the save checks the model, and a create or update with any other model fails with 422 `Model "<model>" is not supported. Please choose a supported model and test your agent configuration.` The message does not list the hosted models, so pick a model from the examples in references/schemas/aiagent.yml or copy one from an existing working agent. With BYOK, a model off the hosted list saves and passes through to the provider as-is. (2) New tool: the provider's built-in code execution (code_execution on OpenAI, codeExecution on Gemini and Anthropic), BYOK only — on Celigo AI the save accepts it but the run is refused with 422 `Code Execution requires a BYOK connection.` Send only {"type": ...}, in the provider's own spelling; the other spelling fails the save. (3) An MCP connection can now also give the agent prompts and read-only resources, beside its tools: "prompts": [{"type": "mcp", "mcp": {"_mcpConnectionId": "…", "allowedPrompts": ["…"]}}] and "resources": [{"type": "mcp", "mcp": {"_mcpConnectionId": "…", "allowedResources": [{"name": "…", "uri": "…"}]}}] (at least one resource entry). The save checks only their shape and that the connection exists, so copy prompt names and resource URIs from the server's catalog; if the server can't be reached at run time, each record fails with mcp_connection_error before the model is called. Prompts and resources are read-only content, not actions. (4) New: one-to-many fan-out — set oneToMany: true and pathToMany to an array field to judge each item on its own; each item becomes its own record and its own billed model call, carrying the parent's other fields under _PARENT, and the items are joined back into the parent record before the next step. (5) Writing jsonSchema yourself on a built-in provider: set schemaMode: "json" beside it and send it again on every full update — a PUT that omits it clears it, and without it the UI opens the schema in its visual editor, whose next save drops keywords it doesn't keep (format, const, $ref, anyOf, additionalProperties, …) and turns integer into number. On OpenAI, a root-level additionalProperties is dropped from jsonSchema on save (nested ones stay). (6) blob output is now OpenAI and Gemini only: the save refuses it with 422 on Anthropic and on catalog providers. (7) Reasoning and sampling limits corrected: OpenAI -pro models start at reasoning.effort "medium" ("low" saves but the run is refused); Claude's thinking is _overrides.anthropic.thinkingConfig — type: "enabled" with budgetTokens (1024 or more; the save refuses enabled without it) or type: "adaptive" with an effort; Anthropic accepts temperature only from 0 to 1 (a higher value fails the save), OpenAI and Gemini go up to 2; leave topK off Claude models released after Claude Opus 4.6, which refuse it at run time. Anthropic's web search tool is webSearch, with allowedDomains or blockedDomains, not both. (8) The system prompt cap is now stated as up to 1,000,000 characters on the built-in providers (the file previously said 50 KB). An agent has no trigger or schedule of its own: it runs when a flow, API or tool hands it a record, or when a chat request sends it a message — a request like "every morning" needs the parent flow built too. Input mapping is Mapper 2.0 mappings rows (with inputContext: "envelope", their extract paths read the run's envelope), with preMap/postMap/postSubmit hooks and no postAggregate hook. (9) Invoke and preview precision: a preview evaluates only the first record; a mockResponse must hold result records ([{"statusCode": 200, "_json": {...}}], or _text for text output) — any other shape fails the save with 422 `Mock Response must be in integrator.io canonical format.` — and a test run with a mock costs nothing and proves nothing about the prompt. Debug logs are read from whatever ran the agent: celigo ai-agents enable-debug <id> opens a window of up to 60 minutes, then celigo flows debug-requests, celigo apis debug-requests or celigo tools debug-requests; the chat routes and preview/compare have no CLI command, and a chat turn runs the agent alone, without its filter, mappings, hooks or response mapping, billed like an invoke, its tools acting for real. A Claude agent's preview shows empty instructions — its prompt travels in _overrides.anthropic.systemInstruction and is applied. Conversation history: the retained history can't be read back through the API and doesn't appear in the agent's Conversation history in the UI, which lists only conversations started there. 1. Fetch the whole skills/configuring-ai-agents/ folder at the pinned commit, SKILL.md and every file under references/, into a temporary folder, then compare SKILL.md with the fingerprint from Skill Harbor's review of 2026-10-22: - SKILL.md: SHA-256 1481ddfc243671906915dc2d04c5a117b8d040c3beaffc14460acc630e7d389a If it does not match, it changed since the review: do not install it yet, tell me what you found and let me decide. 2. Safety check: read SKILL.md and its reference files and review them for anything suspicious (unexpected network calls, shell commands, credential harvesting); STOP on any red flag and tell me. Tell me in plain words what the skill does when used, which services it connects to, whether it reads or sends any personal information or keys, and anything that seems unusual. Confirm these points from the files: - The skill is guidance, not software that runs: the folder holds SKILL.md plus 36 YAML schema files, with no scripts or executable code, and nothing is sent anywhere during installation itself. In use, the workflow drives the Celigo CLI (celigo ai-agents, celigo guardrails) against my Celigo account, and an agent step's model can call web search, MCP servers (their tools, and their read-only prompts and resources), Celigo Tools, image generation or code execution only as configured in that agent. - Providers come in two families with different config formats: built-in (OpenAI in aiAgent.openai{}; Gemini and Anthropic Claude via LiteLLM in aiAgent.litellm{}, with provider settings under litellm._overrides.<provider>) and catalog providers (Mistral, xAI, Hugging Face, DeepSeek, Cohere, Groq) in a generic flat format directly on aiAgent. The provider decides the format: a mismatched block fails the save with 422 invalid_provider_config_format. - Catalog providers are BYOK-only: saving one without a _connectionId fails with 422 connection_required. Built-in providers run on Celigo AI (platform-managed credentials; usage counts against the account's monthly AI token quota) or BYOK (your own API key, configured on a Celigo HTTP connection). The file declares no allowed-hosts list and no vault. - On Celigo AI, the model must be one the platform hosts for that provider, and the save checks it: a create or update with any other model fails with 422 `Model "<model>" is not supported. Please choose a supported model and test your agent configuration.` — the message does not list the hosted models, and nothing is silently swapped. With BYOK, a model off the hosted list saves and passes through to the provider as-is. Gemini model IDs are stored bare, with no gemini/ prefix. - Instructions live at a different path per family (aiAgent.openai.instructions; aiAgent.litellm._overrides.<provider>.systemInstruction; aiAgent.instructions for catalog providers), and the wrong path silently produces an agent with no instructions. adaptorType is case-sensitive: AiAgentImport, GuardrailImport. - A guardrail renders a fixed verdict (flagged: true|false) that the parent flow routes on — guardrails flag, they never block on their own. The three sub-types are pii (detect and optionally mask — mask: true returns the redacted text under a top-level masked result field, visible downstream only if the step's response mapping writes it back), moderation, and ai_agent (custom AI validation); a guardrail's input is selected by mapping, a Mapper 1.0 object, not Mapper 2.0 mappings rows. Response mapping is configured on the flow's pageProcessors[] entry, not on the agent, in the shape {"fields": [{"extract": "_json.<field>", "generate": "<field>"}]}; a json_schema response is available via _json. - Invoking is a real, billed run: celigo ai-agents invoke (a saved agent, or a preview document with no id) creates no job, but the model is called and billed — AI credits on Celigo AI, the key's owner on BYOK — and the agent's tools (MCP servers, Celigo Tools, web search) act for real. A preview evaluates only the first record. Invoke an agent whose tools can write only with my go-ahead; to test the flow around it without calling the model, use a mockResponse of canonical result records ([{"statusCode": 200, "_json": {...}}], or _text for text output — any other shape fails the save) and celigo flows test-run; the mock run costs nothing and proves nothing about the prompt. - Files input depends on the account's Intelligent document processing setting (off by default, account-wide, owner/administrator only, in the UI; an API token cannot read it — ask me). Off: the model's own support decides (PDF, PNG, JPG, WEBP and text; other types, e.g. TIFF, can fail the record). On: Celigo extracts text and tables from PDFs, scans and office files before the model sees them, charged in AI credits. An input filter on the agent step can skip the model for excluded records as a cost control; skipped records continue with the response-mapped fields empty, so later steps writing those fields need their own filter. - json_schema placement depends on the provider family: aiAgent.openai.output.format (OpenAI), aiAgent.litellm.responseFormat (Gemini, Anthropic — type, name, strict and jsonSchema sit directly in responseFormat, no inner format level), or aiAgent.output.format (catalog providers). If you write jsonSchema yourself on a built-in provider, set schemaMode: "json" beside it and send it again on every full update. blob output works on OpenAI and Gemini only; the save refuses it on Anthropic and on catalog providers. Output is capped at 5000 tokens unless set (maxOutputTokens on OpenAI, maxCompletionTokens on Gemini/Anthropic; raisable to 128000, subject to the model's own limit). - One-to-many: with oneToMany: true and pathToMany set to an array field, each item becomes its own record and its own billed model call, carrying the parent's other fields under _PARENT, and the items are joined back into the parent record before the next step. The provider's built-in code execution (code_execution on OpenAI, codeExecution on Gemini and Anthropic) is BYOK only: on Celigo AI the save accepts it but the run is refused with 422 `Code Execution requires a BYOK connection.` - Some settings save and fail only when the agent runs (reasoning.effort "low" on an OpenAI -pro model; topK on Claude models released after Claude Opus 4.6; a code-execution tool on Celigo AI), while others are refused at the save (blob on Anthropic or a catalog provider; Anthropic temperature above 1). Preview the agent on one record before wiring it into a flow. 3. When I say yes, install it as a skill: copy SKILL.md and its reference files into the agent's skills directory, in a folder named "configuring-ai-agents", keeping the references/ subfolder structure as it is. 4. Verify with no network calls: frontmatter valid, files in place. 5. Report what was installed, where, and what I still need to do myself (e.g. hand the agent a Celigo platform account with the CLI configured; for BYOK, create an HTTP connection holding your provider API key — mandatory for a catalog provider; on Celigo AI, pick a model from the examples in references/schemas/aiagent.yml or copy one from a working agent — the save refuses any model the platform doesn't host for the provider). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.

🔒 Security, provenance and what this touches
🧭 What this touches
  • It works with an API key or a token that you provide.
Install prompt analysis · 2026-10-22
🔒 Security and provenance
📌 Pinned version: commit e8c094f · 2026-10-22
🧪 Install ran in our test sandbox · Nov 11, 2026
👁 Observed during the install test · Nov 11, 2026 : no network contact
🔎 Install prompt analysis · 2026-10-22 : uses API keys or tokens
What the scan found — excerpts from the install prompt
  • uses API keys or tokens — Prerequisites: a Celigo platform account (paid iPaaS — AI-agent usage on Celigo AI counts against the account' · - Catalog providers are BYOK-only: saving one without a _connectionId fails with 422 connection_required. Buil · - Files input depends on the account's Intelligent document processing setting (off by default, account-wide,
🔎 Last checked · Nov 19, 2026 : no change detected in the files linked by this install prompt
Skill Harbor never reviews the code of a listed skill. These are dated facts about the install prompt and our install test, not a safety certification. Review a skill yourself before you install it. How we verify
Wanna be extra careful?

The install prompt below already includes the vetting steps: your agent follows the community checklist before installing anything with executable code. Want more?

How to check a skill before installing →

✓

Verification certificate

Dated facts about what was checked for this listing, and how. This is not a safety certification.

Pinned version
e8c094f · 2026-10-22
Sandbox install test
Install ran in our test sandbox · Nov 11, 2026
Creator identity
Not verified
Fingerprint watch
Last checked · Nov 19, 2026 : no change detected in the files linked by this install prompt
Install prompt analysis
2026-10-22 : uses API keys or tokens

Skill Harbor never reviews the code of a listed skill. These are dated facts about the install prompt and our install test, not a safety certification. Review a skill yourself before you install it. How we verify

?

Questions

How do I install a skill?

Every product page includes a copy-paste install prompt. Paste it into your Muse and it sets the skill up for you — no manual configuration.

Where does my money go?

Straight to the seller. Skill Harbor never processes payments: checkout happens on the seller’s own page, usually Stripe.

What does the ✓ next to a creator’s name mean?

It means we confirmed the identity of the person behind the listing. It says nothing about the code itself — always check a skill before installing it.