01
You ask
Describe the outcome
In your MCP client: “When an order arrives, validate it, notify the warehouse, and retry failures.”
MCP-native workflow automation
Your agent writes the workflow.
Taskflow proves it works.
Connect any MCP-speaking agent. It authors real TypeScript; Taskflow compiles it, runs the candidate against a real Cloudflare Workflow instance, and promotes it only after the trial succeeds.
Early product, actively developed
›
create_workflow“Retry the webhook, then archive the result.”
export class Workflow extends
WorkflowEntrypoint {
async run(event, step) {
await step.do("send webhook", ...)
await step.do("archive result", ...)
}
}- 01CompiledTypeScript is valid✓
- 02Trial completedReal instance succeeded✓
- 03PromotedLive version updatedlive
How it works
From a request to a running workflow—in three guarded steps.
02
Your agent codes
It writes real TypeScript
The calling LLM sends a complete Cloudflare Workflow class through create_workflow or update_workflow.
03
Taskflow verifies
Broken code stays out
Taskflow compiles and trial-runs the candidate for real. Only a successful trial becomes the live version—no manual gate required.
Why Taskflow
The freedom of code, with a deployment safety net.
Your agent is not limited to a canvas or a catalog of canned actions. Taskflow gives it code-level flexibility and keeps failed candidates away from what is already working.
Works where your agent works
- Standard MCP tool calls
- Create, update, and trigger
Code, not connected boxes
- Real TypeScript logic
- Authored by the calling LLM
Safe, automatic promotion
- Compile errors stop early
- Failed trials never replace live code
Organized for every workspace
- Projects keep workflows grouped
- Per-workspace API keys
Cloudflare all the way down
- Workers, Durable Objects, and D1
- Native Cloudflare Workflows
Easy to enter, scoped by design
- Google or magic-link sign-in
- Multi-tenant workspace boundaries
Build beyond the template