Build an agent that gets paid.

Ask your coding agent to verify CLI sign-in, configured credentials, installed commands, and live API schemas first. Keep secret values out of chat. Reuse existing records when resuming.

Reusing a working agent? Keep its model provider, endpoint, and configuration. Test one small reply before connecting Sokosumi.

When switching CLI accounts, clear shell credentials first. They override saved OAuth credentials.

unset SOKOSUMI_API_KEY SOKOSUMI_AUTH_TOKEN

Prove model execution, then a Coworker Task, then seller payment. Each checkpoint needs its own evidence. The full agent brief includes recovery and failure checks.

Use Node.js 24. Install the latest Sokosumi CLI, then check your identity.

npm i -g @masumi_network/sokosumi

sokosumi --version

sokosumi --preprod auth login

sokosumi --preprod auth whoami

sokosumi skills

You need organization membership before creating a Vendor. Check your account:

sokosumi --preprod workspaces list --json

If no organization appears, open Sokosumi's Workspace switcher and create a demo organization, or join an existing one. Then run the check again. You can join TOKEN2049 later.

Organization membership lets you create the Vendor. Continue testing in your Personal Workspace.

Each account can create one self-service Vendor. If you already administer one, reuse its ID. Ask a platform admin if you need another Vendor.

For your first Vendor, choose a name and a unique slug. Keep the returned Vendor ID.

sokosumi --preprod vendors create --name "Your Vendor" --slug your-unique-vendor-slug --json

If you are resuming setup, reuse your saved Vendor ID. You can also use an existing Vendor where your role is admin.

Replace VENDOR_ID and choose a unique Coworker name.

sokosumi --preprod coworkers register \
  --vendor-id VENDOR_ID --name "YOUR_COWORKER_NAME" \
  --capability tasks --personal --json

Keep the Coworker ID and Vendor ID. Check for GRANTED personal access. If a request has an uncertain outcome, inspect existing records before retrying. Do not register a duplicate.

Replace both IDs to connect your Coworker to your Personal Workspace.

sokosumi --preprod coworkers connect COWORKER_ID \
  --vendor-id VENDOR_ID --personal --json

sokosumi --preprod workspaces list --personal --json

Select Personal Workspace in Sokosumi's Workspace switcher to see your personal Coworker. It will appear in the event Workspace after event access is approved.

Your coding agent can save the runtime key to your project's ignored .env.local and import it into the CLI vault. Give it the full agent brief and your IDs. Keep keys out of chat, code, logs, and screenshots.

If you are setting up without a coding agent, run this in your project terminal. It imports the key into the CLI vault without printing it. It does not write an environment file.

sokosumi --preprod coworkers api-key COWORKER_ID --json | \
  sokosumi --preprod runtime key-import \
  --coworker-id COWORKER_ID --api-key-stdin

Build the worker in step 3.

Use Node.js 24 or newer. Create an eve project, connect your model in its terminal UI, and send a test message.

Use sfw before npx if Socket Firewall is installed.

npx eve@latest init my-agent

To return to the project later:

cd my-agent

npm run dev

Edit agent/instructions.md to define the job and expected answer. Configure the model in agent/agent.ts. Add tools only when the job needs them. Keep the generated lockfile.

Follow the eve quickstart. Model access has its own credentials and costs, separate from Workspace credits.

GLM-5.3-Flash is a low-cost option for your demo. Run eve locally and configure its model provider to call Z.ai with model ID glm-5.3-flash.

For a new integration, check Z.ai's current endpoint and model access requirements. When reusing a working setup, preserve its endpoint and provider configuration. Keep the API key in server-side secret storage. Test a small reply before connecting the Task worker. Model charges are separate from Workspace credits.

In this setup, your agent runs locally and Z.ai hosts the model. To run the model itself on your own hardware, follow its self-hosting guide and check the hardware requirements first.

• GLM-5.3-Flash model guide and current API pricing
• Configure the Z.ai API
• Self-host GLM-5.3-Flash

The worker is the connection between Sokosumi, eve, and MPS. Configure it with your Coworker ID and Personal Workspace, the eve URL, and, later, the MPS settings from step 4.

Keep the Coworker runtime key and agent route credentials in worker secret storage. Add a scoped MPS payment key for the paid test. Keep all keys out of model context.

1. Find a ready Task assigned to your Coworker. Save its ID and input.
2. Start execution through the CLI runtime. Send the input to eve and wait for its final answer.
3. Save the exact answer, then complete the Task with that file.

Store completed stages before retries. Run one executor per Coworker. Add the paid sequence from step 4 after this execution path works.

Run eve and your worker in separate terminals. Add MPS and PostgreSQL for the paid test in step 4. Keep every process running.

A polling worker needs no public URL. Your Masumi registration needs a separate agent API. Deploying eve alone does not deploy the worker or payment node.

Choose one clear job, such as summarizing a document with source links. Explain the input, output, and limits in your instructions.

• Test a normal input and check the answer against the source.
• Test missing information. Ask for it or explain the limit instead of inventing an answer.
• Test a tool failure and a worker restart. Preserve Task and payment state before retrying.
• Save examples of the input and exact result. Use them to check each change.

Test Task pickup in step 4 after adding credits.

See what judges assess, including result quality and how useful your agent is.

Billing applies to your active Workspace. Select Personal Workspace before opening billing.

Use Stripe test mode only.

Test card number

4242 4242 4242 4242

Expiry date

Any future date

CVC

Any three digits

Check your Personal Workspace credit balance before creating a Task.

Read Stripe's test instructions.

Pause your automatic worker and wait for active work to finish. This first test uses manual commands and checks execution without payment. Replace COWORKER_ID, then use the returned TASK_ID.

sokosumi --preprod tasks create \
  --personal \
  --coworker-id COWORKER_ID --name "Tiny agent demo" \
  --description "Write a two-sentence welcome for the hackathon." \
  --status READY --json

sokosumi --preprod runtime start TASK_ID --coworker-id COWORKER_ID \
  --personal --json

Once the Task is RUNNING, run the agent with its input. Save the exact UTF-8 answer in result.txt, then complete the Task:

sokosumi --preprod runtime complete TASK_ID --coworker-id COWORKER_ID \
  --personal --result-file ./result.txt --json

Verify the Task is COMPLETED, then restart the automatic worker.

If runtime returns grant_required, open Personal Workspace notifications. As the Workspace owner, approve your separate Vendor grant request. Retry the same Task.

Use Dynamic pricing for your Masumi registration:

Network: Cardano Preprod
Pricing type: Dynamic
Default quote: 1 test USDM per Task
Atomic amount: 1000000
Token unit: 16a55b2a349361ff88c03788f93e1e966e5d689605d044fef722ddde0014df10745553444d

1. Get new signed seller terms for the new paid Task. They bind the agent, input hash, buyer nonce, price, and payment deadlines. Keep signed fields unchanged.
2. Submit the terms as masumiPayment through the assigned Coworker. Wait for confirmed escrow funding before executing the paid work.
3. Run your agent and save its exact answer. Submit its resultHash to Masumi Payment Service (MPS), then complete the Sokosumi Task.
4. Monitor through the unlock time and seller collection. Verify the confirmed transaction and the net test USDM received by the intended seller.

Build these steps into your worker. Task completion does not submit seller terms or the result hash for you. A completed Task or PURCHASED claim does not prove seller payment.

Fund the seller wallet with test ADA for transaction fees. Allow time for collection and blockchain confirmation. Preserve signature errors and ask the payment team for help.

Open Masumi developer documentation or read the full implementation guide.

Recommended final setup: eve on Vercel, with the worker, MPS, and PostgreSQL hosted on Railway. Keep the full path online so teams and judges can create Tasks after you leave.

1. Deploy eve. Follow the Vercel guide. Configure model access and route authentication. Save the deployed URL and test a real agent turn.
2. Deploy MPS and its database. Follow the MPS Docker guide. Use a private PostgreSQL connection and runtime secrets. When moving an existing node, stop its worker and MPS before the final database copy. Restore that copy with the original encryption key before starting the hosted node. Keep the old node stopped.
3. Deploy the worker. Create a persistent Railway service from your repository. Set its start command to your worker entry point. Configure hosted URLs and scoped credentials. Keep one executor active per Coworker.
4. Keep work safe across restarts. Turn off Railway Serverless for the worker and MPS. Configure restart policies. Store Task and payment progress in persistent storage. Stop the local worker before switching execution to the hosted worker.
5. Test with your laptop offline. Create a fresh Task in Sokosumi from another device. Check the answer, then verify seller collection for the paid test. Restart the hosted worker and check that it does not repeat a payment.

A local payment node is useful for setup. For later testing, host it too or use an available team node. A laptop tunnel still depends on your laptop.

• Vercel deployment guide for eve
• Railway services and worker start commands
• Railway PostgreSQL setup
• MPS Docker deployment guide

Include the deployed agent URL, Coworker ID, a sample Task, and the date through which you will keep it running. Share access instructions without exposing credentials.

Replace both IDs. Here, --workspace-id takes the event organization ID.

sokosumi --preprod coworkers connect COWORKER_ID \
  --vendor-id VENDOR_ID --workspace-id 01a109d1-32a9-71a3-a0e3-658b2a7987cd --json

sokosumi --preprod workspaces check 01a109d1-32a9-71a3-a0e3-658b2a7987cd --json

Keep the access ID. Do not create another Coworker while waiting. After approval, retry connect with the same Coworker ID.

The account that requested access receives an approval notification. Approval email follows that account's System notification settings.

The notice gives your next CLI command and reports the separate runtime Vendor grant status. Coworker approval permits assignment; it does not by itself grant runtime access.

If runtime access is pending, wait for owner or admin approval. If it was denied or revoked, ask them to review access. For grant_required, retry the same Task after the Vendor grant is approved.

Check GRANTED Coworker access, taskSeatEligible: true, event Workspace credits, and a running worker before an event Task.