Skip to content
lacause edited this page Mar 29, 2026 · 10 revisions

OCC — Claude Chain Orchestrator Wiki

Welcome to the OCC documentation. OCC is a workflow orchestration engine for Claude AI agents — define multi-step chains in YAML, execute them with parallel processing, and access everything via MCP or REST API.

14 modules, 437 tests, 40+ REST endpoints, 25 MCP tools.

Quick Navigation

Getting Started

  • Getting Started — Install, configure, run your first chain in 5 minutes
  • Configuration — Environment variables, .mcp.json, Docker, deployment

Chain Authoring

  • Chain Format — Complete YAML reference for chain definitions
  • Step Types — All 11 step types explained with examples
  • Pre-Tools — Inject data before steps (web search, API calls, MCP servers, bash)
  • Pipelines — Chain multiple chains together

Tools & API

  • CLI — Command-line interface: run, validate, dry-run, cost estimate
  • REST API — All 40+ HTTP endpoints with request/response examples
  • MCP Tools — 25 MCP tools for Claude Code / Claude Desktop
  • MCP Client — Consume 10,000+ external MCP servers in chains

Advanced

  • Architecture — 14 modules: executor, queue, storage, linter, MCP client, etc.
  • Token Optimization — Minimize cost with caching, transforms, context strategies
  • Cookbook — Practical recipes and patterns

Help

  • FAQ — Common questions and troubleshooting

Key Concepts

Concept Description
Chain A YAML file defining a multi-step workflow. Each step is an LLM call, data transform, or control flow.
Step One unit of work in a chain. 11 types: agent, router, evaluator, gate, transform, loop, merge, browser, subchain, debate, webhook.
Pipeline Multiple chains executed in sequence, with outputs passed between them.
Pre-tool Data injection before a step — web search, HTTP fetch, file read, bash, env var, MCP call. Supports on_error (inject/skip/fail).
Variable Outputs stored as output_var and interpolated into prompts via {variable_name}. Supports {key.subkey} and {key|"fallback"}.
Wave A group of steps with no dependencies between them — they execute in parallel.
Gate A step that pauses execution until a human approves or rejects via the API.
Queue Persistent job queue — if all workers are busy, new executions are queued with priority instead of rejected.
Checkpoint Each step result is saved to SQLite immediately. Crash recovery resumes from last completed step.
Execution A running instance of a chain, with its own ID, status, step results, and SSE stream.

Clone this wiki locally