Celigo

Each agent gets a set of tools and MCP connections that it calls to take actions in other systems while processing a record. Tools can be built from the platform's own imports, lookups and APIs, which already hold authenticated connections to applications such as NetSuite, Shopify, Amazon and HubSpot, and any MCP server can be attached with HTTP headers, token or OAuth 2.0 authentication, its tools, resources and prompts becoming available at runtime. Least privilege access limits each agent to the tools and data its job needs, and web search and image generation are optional built in tools.

Sourcedocs.celigo.com/hc/en-us/articles/53427604879643-create-an-mcp-connection-from-an-ai-agentread 2026-10-18

Agents run as steps inside integrator.io flows and APIs, the same engine that runs a company's integrations. A flow reads records, runs preMap hooks and Mapper 2.0 transformations, hands each record to the agent, runs postMap hooks and maps the agent's standard response into later steps, with guardrails, lookups and imports placed before or after it. Flows can be daisy chained, run with standard error handling, retries and auto recovery, and the same agent can be reused across any number of flows and APIs, so multistep processes mix deterministic rules and agent judgment on one runtime.

Sourcedocs.celigo.com/hc/en-us/articles/45255470264859-learn-about-celigo-ai-agentsread 2026-10-18

Context reaches an agent per record. Mapper 2.0 shapes the record and any lookups into the agent's input, tools and MCP connections read live state from connected systems during the run, and files such as images and PDFs can be attached for extraction. The prebuilt commerce agents work this way, reading inventory, orders and settlements across systems as they go. There is no knowledge base, document index or vector store that an agent searches across runs; grounding comes from the data the flow brings and the systems the agent can query.

Sourcedocs.celigo.com/hc/en-us/articles/45566487108251read 2026-10-18

Guardrails check each record in real time before or after an agent: a PII guardrail detects chosen entity types above a confidence threshold and can mask them, and a moderation guardrail screens content natively or through the customer's own provider key, blocking what fails policy. Least privilege tool access limits what an agent can touch, and Ora stages every change it makes to flows and agents for a person's approval. The agent setup itself has no step where a person signs off on a single tool call before it runs.

Sourcedocs.celigo.com/hc/en-us/articles/45523839613851-create-a-guardrailread 2026-10-18

Single sign on works with Okta and Microsoft Azure, and roles separate account owner, admin, manage all and monitor all access, with chat in the agent preview limited to manage roles. MCP servers add per user invitations, capability sets and end user credentials so a tool runs with the caller's own permissions.

Credentials are tokenized, data is encrypted in transit with TLS 1.2 or higher and at rest with AES 256, and processed data is not stored persistently. Celigo holds SOC 1 Type II and SOC 2 Type II reports, covers GDPR, the Data Privacy Framework and CCPA, and offers HIPAA readiness with a BAA. ISO 27001 and ISO 42001 certification are targeted for the first quarter of 2027, and there is no FedRAMP or PCI DSS certification.

Sourceceligo.com/platform/securityread 2026-10-18

Execution logs record what happened while an agent or guardrail ran, including the model used, tools invoked, guardrail decisions, retries and token and credit consumption, searchable on the run history page by trace key, timestamp or agent name. Guardrail evaluation details show policy results, violations, severity and enforcement. Audit logs record who created, edited or deployed an agent, tool or guardrail, error records support retry, and MCP request logs show which access token made each call. Detailed logging can be switched on for deeper diagnostics.

Sourcedocs.celigo.com/hc/en-us/articles/47623592400155-monitor-ai-agent-and-guardrail-activity-with-execution-logsread 2026-10-18

Agents process each record on its own and return a response to the flow, carrying nothing forward to the next record or run. Since the 2026.4.1 release each agent keeps a saved conversation history, but it exists so a builder can resume a test chat in the preview pane, and production runs do not read it. State that has to persist, such as a customer's order history or a reconciliation status, lives in the connected applications, in lookups the flow performs or in flow settings, and the agent sees it only when the flow passes it in or a tool fetches it.

Sourcedocs.celigo.com/hc/en-us/articles/45255470264859-learn-about-celigo-ai-agentsread 2026-10-18

Celigo runs independent data center domains in the United States, Canada, Europe and Australia. A customer picks the domain at sign up, and all account data stays in that region. Celigo Private Cloud hosts a dedicated instance in a preferred region with encryption, role based access and audit logging, and an on premise agent reaches systems behind a customer's firewall. Moving between domains means rebuilding the account, since integrations do not transfer.

Sourcedocs.celigo.com/hc/en-us/articles/16924189197467-move-accounts-between-regional-domainsread 2026-10-18

Five prebuilt commerce agents ship ready to add to a flow: Amazon settlement reconciliation, which fixes unmatched invoices and fee or refund mismatches; a weekly discount floor check that returns the maximum safe discount per SKU; a daily oversell check that names the oversold SKUs and the cause; an available to promise agent that answers wholesale buyers with a confidence level; and order velocity protection for stockout warnings. The marketplace adds integration apps, templates and MCP server templates installed in a few steps.

Sourceceligo.com/platform/ai-agentsread 2026-10-18

Flows that contain agents start on demand, on a schedule written as a cron expression, from webhook listeners that receive events from applications such as Shopify and HubSpot in real time, or when another flow finishes. APIs built in Celigo's API builder can call agents per request, and Celigo's own MCP servers let an outside assistant invoke published tools. That covers scheduled, event driven and request driven starts without a person opening a chat.

Sourcedocs.celigo.com/hc/en-us/articles/360015827372read 2026-10-18

The builder picks the model for each agent. Celigo AI offers a curated list run under a Celigo managed account, covering OpenAI's GPT 5 and GPT 4.1 families, Google Gemini 2.5 Pro and Flash, and Anthropic Claude Sonnet, Opus and Haiku models, with GPT 4.1 mini as the default. Bring your own key routes an agent through the customer's own provider connection, billed by that provider, and a capability matrix keeps agents on models that support the tools they use. Model lifecycle labels flag models nearing retirement.

Sourcedocs.celigo.com/hc/en-us/articles/45255381685915-create-an-ai-agentread 2026-10-18

Integrations run programmatically through the Celigo API, a command line interface and a Platform MCP server. Customers also publish their own MCP servers from integrator.io, each with a URL and path, API token or user authentication, selected tools derived from Celigo APIs, resources, reusable prompts, capability sets and traffic limits, so outside agents such as Claude and Cursor invoke integration capabilities while execution stays inside Celigo. API builder lets teams expose flows, lookups and agents as their own endpoints.

Sourcedocs.celigo.com/hc/en-us/articles/45255402177819-mcp-server-overviewread 2026-10-18

The preview pane runs an agent with the same configuration it will use in the flow, on the previous step's output or on mock data, lets a builder chat with it, compares responses across models side by side, and shows execution details for refining instructions. Agents and guardrails can be previewed on their own before they join a flow. Each preview is a check a builder reads by eye, with no stored test set, expected outputs or score to compare one version of an agent with the next.

Sourcedocs.celigo.com/hc/en-us/articles/45566533436955-preview-an-ai-agentread 2026-10-18

Agents act through connectors, APIs, tools and MCP servers, with web search and image generation as optional built in tools. Nothing in the agent builder drives a browser, a desktop or a screen, and the prebuilt commerce agents work entirely through the connections a flow already holds to Amazon, NetSuite, Shopify and similar systems.

An application without an API sits outside what a Celigo agent can operate, unless an MCP server the customer attaches supplies that reach. The on premise agent extends connectivity to databases and systems behind a firewall, which is network access for integrations and involves no screen at all.

Sourcedocs.celigo.com/hc/en-us/articles/45255381685915-create-an-ai-agentread 2026-10-18