AI Overview

Octo treats AI as a first-class part of an integration, not a bolt-on. LLM providers are connectors like any database or HTTP endpoint, AI capabilities are blocks you drop into a pipeline, and flows speak MCP in both directions.

LLMs are connectors

You configure an LLM provider the same way you configure Postgres or Slack: a named connector with settings. The Anthropic, OpenAI, and Gemini connectors all satisfy the same client interface, so every AI block binds to a provider by connector name and works with any of them interchangeably.

connectors:
  - name: claude
    type: llm-anthropic
    settings:
      apiKey: ${ANTHROPIC_API_KEY}

Swap llm-anthropic for llm-openai or llm-gemini and nothing else in the flow changes. See LLM Providers.

AI capabilities are blocks

AI in Octo is not a separate runtime or SDK — it is a family of blocks that compose with everything else in a flow:

• ai-agent — an LLM accomplishes a task by calling flow branches as tools, in a loop, with an iteration cap and a guardrail path.
• ai-router — an LLM picks exactly one named route for each message; the fuzzy sibling of the deterministic switch.
• ai-mapping — reshape a messy payload into a target shape, validated against a JSON Schema.
• ai-retry — protect a process chain; on failure an LLM inspects the error, revises the message, and retries.

Because they are ordinary blocks, they sit in the middle of pipelines: an HTTP source feeds an ai-router, whose route runs an ai-mapping wrapped in an ai-retry, followed by a database write. The full field reference lives at AI Blocks.

Agents are declarative

An ai-agent gets its memory, tools, and skills from YAML, not code. Per-thread conversation memory is one field (memoryThreadId); tools are process chains the model calls with JSON arguments; skills are instruction documents the model loads on demand. The whole agent — prompt, tools, memory policy, fallback behavior — is a single reviewable block in a flow file.

Flows speak MCP

The integration runs in both directions:

• Flows can be MCP servers. The mcp-router block turns flows into tools, resources, and prompts that any MCP client — Claude, IDEs, other agents — can call. See Expose an MCP Server and Secure an MCP Server with OAuth.
• The platform is an MCP server. The runtime and editor expose an MCP endpoint for AI-assisted authoring: a coding agent can draft, validate, run, and debug integrations against your environment. See The Platform MCP Server.

Explore the section