OCC (Claude Chain Orchestrator) is a workflow engine that lets you define multi-step AI agent chains in YAML and execute them with automatic parallel processing, dependency resolution, and real-time streaming.

Not directly. OCC uses the Claude CLI by default, which handles its own authentication (claude auth login). But you can also use other providers (OpenRouter, OpenAI, Ollama, HuggingFace) — each with their own key setup.

OCC supports 6 LLM providers:

• Claude CLI — claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5
• OpenRouter — 200+ models (GPT-4o, Gemini, Llama, DeepSeek, Mistral...)
• OpenAI — GPT-4o, GPT-4o-mini, o3-mini, o4-mini
• Ollama — Any local model (Llama, Mistral, Phi, Qwen...)
• HuggingFace — 118 free models via Inference API (DeepSeek R1, Qwen 3, Gemma 4...)
• Custom — Any OpenAI-compatible API

Specify per-step with the model field: model: "openrouter/gpt-4o" or model: "ollama/llama3.2".

Depends on your chain and provider. A typical 6-step chain costs ~0ドル.05-0.15 with Claude Sonnet. With HuggingFace free tier or Ollama local models, it costs 0ドル. See Token Optimization for cost reduction techniques.

Yes. OCC is cross-platform (macOS, Linux, Windows). CI runs on all three with Node.js 20 and 22.

Yes. OCC includes a React dashboard (frontend-react/) with:

• Chain/pipeline management
• Visual canvas editor (BLOB)
• Workflow chat (conversational chain builder)
• Real-time SSE execution monitoring with mini-terminals
• Settings panel (providers, Ollama, HuggingFace, schedules, MCP servers)
• Run at http://localhost:5173 with cd frontend-react && npm run dev

The Claude CLI is not installed or not in your PATH.

# Install Claude CLI
npm install -g @anthropic-ai/claude-code
# Verify
which claude # macOS/Linux
where claude # Windows
# Or set path explicitly
export CLAUDE_BIN=/path/to/claude

TypeScript is a dev dependency. Use npx:

cd mcp-server
npm install
npx tsc --noEmit # Type check
npm run build # Build

Change the port:

REST_PORT=8080 npm run rest

Your steps have circular depends_on references. Example:

# BAD: A depends on B, B depends on A
- id: a
 depends_on: [b]
- id: b
 depends_on: [a]

Fix: restructure so dependencies form a DAG (directed acyclic graph).

Your chain defines required inputs that weren't provided:

inputs:
 - name: topic # Required (optional: false is the default)
 - name: depth
 optional: true # This one can be omitted

Provide all required inputs when executing:

curl -X POST http://localhost:4242/execute/my-chain \
 -d '{"input": {"topic": "AI"}}'

Common causes:

1. Timeout — step took too long. Increase timeout_ms
2. Claude error — check the step's error field in the execution result
3. Wrong variable — check that {variable} names match output_var values
4. Condition false — step was skipped because its condition evaluated to false

1. Check execution status:

   curl http://localhost:4242/executions/{id}

2. Stream real-time logs:

   curl -N http://localhost:4242/executions/{id}/stream

3. Look at individual step outputs in the execution result — each step has status, output, error, and durationMs.

Yes. Add the tools field:

- id: analyze
 tools: [Read, Write, Bash, Glob, Grep, WebSearch, WebFetch]
 prompt: "Read and analyze the code at {input.path}"
 output_var: analysis

Default limit is 5. Wait for running executions to finish, or increase:

MAX_CONCURRENT_EXECUTIONS=10 npm run rest

curl -X DELETE http://localhost:4242/executions/{id}

This kills all active processes and marks pending steps as skipped.

curl -X POST http://localhost:4242/executions/{id}/resume

Completed steps are skipped. Execution resumes from the first non-completed step.

OCC sends heartbeat pings every 30s to prevent this. If you're behind a proxy (Nginx, Cloudflare), configure it to allow long-lived connections:

proxy_read_timeout 3600s;
proxy_send_timeout 3600s;

1. Check .mcp.json exists in your project root
2. Verify paths are absolute (not relative)
3. Restart Claude Code after editing .mcp.json
4. Check that npm run build was run (needs dist/index.js)

The path in .mcp.json → env.CHAINS_DIR doesn't point to a valid directory. Use an absolute path:

"CHAINS_DIR": "/Users/you/projects/OCC/chains"

• Use parallel steps — remove unnecessary depends_on so steps run simultaneously
• Use cheaper models — claude-haiku-4-5 for simple classification/extraction
• Enable caching — avoid re-running identical prompts
• Reduce timeout — set timeout_ms per step instead of waiting 30 min default
• Use transforms — json_extract instead of asking the LLM to extract

• Reduce MAX_CONCURRENT_EXECUTIONS
• Set EXECUTION_MAX_AGE_DAYS to a lower value (default: 7)
• Large step outputs (>1MB) accumulate in memory — use transforms to reduce size

Use pre_tools to read files and tools: [Write] to create them:

- id: generate
 tools: [Write]
 prompt: "Generate code and write to /tmp/output.py"
 output_var: gen_result
- id: review
 depends_on: [generate]
 pre_tools:
 - type: read_file
 path: "/tmp/output.py"
 inject_as: code
 prompt: "Review this code: {code}"
 output_var: review

pre_tools:
 - type: env_var
 var_name: "DATABASE_URL"
 inject_as: db_url
prompt: "Connect to {db_url} and..."

curl -X POST http://localhost:4242/schedules \
 -H "Content-Type: application/json" \
 -d '{
 "label": "Morning briefing",
 "chainName": "smart-briefing",
 "input": {"topic": "tech news"},
 "cron": "0 8 * * *",
 "enabled": true
 }'

• Getting Started — Initial setup
• Chain Format — YAML reference
• Configuration — Environment variables
• Token Optimization — Cost reduction