Forge

Multi-step LLM workflows as directed acyclic graphs.

Workflows

Define multi-step workflows where each step can be an LLM call, a transform, or a tool call. Steps declare dependencies and execute in topological order.

# Create a workflow
curl -X POST http://localhost:4200/api/forge/workflows \
  -d '{
    "slug": "draft-and-critique",
    "name": "Draft + Critique",
    "steps": [
      {"id":"draft","type":"llm","config":{"model":"gpt-4o-mini","prompt":"Write about {{input}}"}},
      {"id":"critique","type":"llm","depends_on":["draft"],
       "config":{"prompt":"Critique: {{steps.draft.output}}"}},
      {"id":"final","type":"transform","depends_on":["draft","critique"],
       "config":{"expression":"concat"}}
    ]
  }'

Running Workflows

# Run it
curl -X POST http://localhost:4200/api/forge/workflows/draft-and-critique/run \
  -d '{"input": "the future of AI"}'

# Check status
curl http://localhost:4200/api/forge/runs/{run_id}

Step Types

Step types: llm (calls the proxy), transform (concat, extract_json, first_line), tool (external tool calls).

Template variables: {{input}} for the run input, {{steps.step_id.output}} for dependency outputs.

Tool Registry

# List registered tools
curl http://localhost:4200/api/forge/tools

See the Forge product page and the visual workflow builder.

Creating Workflows

Forge uses DAG (directed acyclic graph) definitions to chain LLM calls, transforms, and tool invocations:

curl -X POST http://localhost:4200/api/forge/workflows \
  -H "Authorization: Bearer sy_admin_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "summarize-and-translate",
    "steps": [
      {
        "id": "summarize",
        "type": "llm",
        "model": "gpt-4o",
        "prompt": "Summarize: {{input}}"
      },
      {
        "id": "translate",
        "type": "llm",
        "model": "gpt-4o",
        "prompt": "Translate to Spanish: {{summarize.output}}",
        "depends_on": ["summarize"]
      }
    ]
  }'

{
  "id": "wf_abc123",
  "name": "summarize-and-translate",
  "steps": 2,
  "created_at": "2026-02-28T12:00:00Z"
}

Running Workflows

curl -X POST http://localhost:4200/api/forge/workflows/wf_abc123/run \
  -H "Authorization: Bearer sy_admin_..." \
  -H "Content-Type: application/json" \
  -d '{"input": "The quarterly results showed a 15% increase in revenue..."}'

{
  "run_id": "run_xyz789",
  "workflow_id": "wf_abc123",
  "status": "completed",
  "steps": {
    "summarize": {"status": "completed", "output": "Revenue increased 15% quarter-over-quarter..."},
    "translate": {"status": "completed", "output": "Los ingresos aumentaron un 15%..."}
  },
  "total_cost_usd": 0.012,
  "duration_ms": 2400
}

Listing Workflows

curl http://localhost:4200/api/forge/workflows \
  -H "Authorization: Bearer sy_admin_..."

Workflow Detail

curl http://localhost:4200/api/forge/workflows/wf_abc123 \
  -H "Authorization: Bearer sy_admin_..."

Parallel Execution

Steps without dependencies run in parallel. The multicall module handles fan-out:

{
  "name": "multi-model-compare",
  "steps": [
    {"id": "gpt", "type": "llm", "model": "gpt-4o", "prompt": "{{input}}"},
    {"id": "claude", "type": "llm", "model": "claude-sonnet-4-5", "prompt": "{{input}}"},
    {"id": "pick_best", "type": "llm", "model": "gpt-4o",
     "prompt": "Pick the better answer:\nA: {{gpt.output}}\nB: {{claude.output}}",
     "depends_on": ["gpt", "claude"]}
  ]
}

Steps gpt and claude run simultaneously. pick_best waits for both.

Deleting Workflows

curl -X DELETE http://localhost:4200/api/forge/workflows/wf_abc123 \
  -H "Authorization: Bearer sy_admin_..."

Forge Configuration

# stockyard.yaml
apps:
  forge:
    max_workflows: 200
    max_steps_per_workflow: 20
    execution_timeout: 60s
    parallel_limit: 5

Step Types

Forge supports several step types within a workflow:

Error Handling

Steps can define fallback behavior when they fail:

{
  "id": "classify",
  "type": "llm",
  "model": "gpt-4o",
  "prompt": "Classify: {{input}}",
  "on_error": {
    "action": "fallback",
    "fallback_model": "gpt-4o-mini",
    "max_retries": 2
  }
}

If the primary model fails, the step retries with the fallback model before reporting an error.

Workflow Tips

Variable References

Steps reference outputs from other steps using {{step_id.output}} syntax. The Forge engine resolves these references and passes data between steps automatically:

# Available variables in each step
{{input}}              # Original workflow input
{{step_id.output}}     # Output from a completed step
{{step_id.status}}     # Step status (completed/failed)
{{step_id.cost_usd}}   # Cost of that step

API Summary

Limits & Quotas

Adjust limits in stockyard.yaml under apps.forge. Higher limits may increase memory usage during workflow execution.