Your agent is the workflow designer.
Keep working in Claude Code, your preferred coding agent, or your own application. Runfabric supplies the tools to build durable automations and the evidence to operate them.
From a conversation to a published process.
Ask your agent to define the trigger, inputs, integrations, paths and desired result. Runfabric supplies a discoverable schema, an authoring skill and management tools. The agent can inspect workflows, create drafts, validate them, publish versions and start runs.
Put a human approval before a consequential write. Give agent steps a bounded tool set. Inspect a rehearsal before publishing real integrations.
MCP for your coding agent.
Use the hosted MCP endpoint with project-scoped authorization, or install the stdio server alongside your agent. Tools cover discovery, authoring, publication, runs, approvals, integrations and triggers.
Your MCP client handles browser authorization. The human chooses the project and permitted operations.
https://dev.runfabric.dev/mcp
CLI for repeatable work.
The native CLI produces structured output and accepts JSON definitions. Browser login establishes scoped access; a Project Key supports CI and private workers.
runfabric auth login runfabric workflow list runfabric run list
Agentic steps inside the automation.
An agentic activity is a bounded reasoning loop. It sends the configured prompt and input to your broker, receives tool calls, validates them against an allowlist, executes permitted integrations and sends results back to the model. It continues until the model answers or a turn or tool limit is reached.
Choose OpenRouter or another OpenAI-compatible endpoint and supply your own credential. The workflow engine does not force a model choice. Integration credentials remain separate from the prompt and graph.
Hosted agents can use configured HTTP and remote MCP integrations. On your worker they can also use local CLI and MCP tools. A model's request cannot grant itself a new destination, credential or executable.
One API for applications, too.
Use the TypeScript client to create and publish workflows, start idempotent runs, watch results and record approvals. Embed the pure validator when your own interface needs to check a definition.
import { WorkflowClient } from '@coralbeat/workflows/sdk';
const runfabric = new WorkflowClient({
baseUrl: 'https://dev.runfabric.dev',
apiKey: process.env.RUNFABRIC_API_KEY,
});
const workflows = await runfabric.listWorkflows();A human view when you need one.
Open the same project in the web app to inspect the graph, branches, loops, receipts and approvals. The UI is another view of the process; agents do not need to automate browser clicks to use the product.