MCP-first action workspace

Turn useful MCP work into a durable action you can run again

Keep using the external client where you already reason. Connect it to OrgSDK, turn a useful result into a saved, versioned action, and give that work a run history, durable outputs, and a path to run unattended.

Your first useful run

Prove the durable action loop before adding complexity

Start without an external service credential. One credential-free action proves the connection, save, run, artifact, and trace as separate steps you can inspect.

01

Connect your client

Use the connection details in your OrgSDK workspace, then ask the client to inspect the OrgSDK capabilities it can actually see.

02

Save useful work

Have the client create make-checklist-artifact as an action with declared inputs. A successful source change creates a version you can return to.

03

Run without credentials

Start the checklist action from your client. It creates a durable Markdown artifact without connecting an external service or model provider.

04

Inspect the proof

Open the returned run in OrgSDK. Confirm the action version, inputs, output artifact, status, and trace. If it fails, read the trace before retrying.

05

Expand only after proof

Add a real capability, revise the action, and prove another run. Add a hook only after the action has a successful trace.

What compounds

The action is the asset—not the MCP connection

MCP opens the workspace. The value is the work you save, version, run, inspect, and improve after the conversation ends.

Versioned actions

Useful work becomes editable TypeScript with declared inputs and a history of saved source versions—not another answer trapped in a chat.

Durable results

Tables and artifacts carry structured state and named outputs across runs, so later work can build on what already happened.

Runs you can explain

Every start creates an inspectable run tied to the saved version. The trace shows what ran, what it called, and where it paused or failed.

Capabilities when needed

Plugins give proven actions typed capabilities. Connect a service only when the real job requires it.

Hooks after proof

A hook can start a saved action on a schedule or event after the manual path works. It uses the same run and trace model.

Human authority

Risky work can pause for a person to review in the control UI. A normal external-client connection does not make that decision for you.

Give useful MCP work somewhere to keep working

Start with one credential-free action. Save it, run it, inspect the trace, then add the capability and hook the real job needs.