FOR AGENTS
Set up Sekiwa with an agent
An agent can set up your workspace, wire the SDK into your app, and read licence state. It connects through the Sekiwa MCP server. Writes wait until an owner or admin approves them in the dashboard.
Agents can also read /llms.txt, the short index written for them. The OpenAPI contract covers the HTTP API.
1. Connect
- In the dashboard, open Settings, Security, Agent access. Choose an access level and create an agent key.
- Put the key in the environment your agent starts from, as
SEKIWA_API_KEY. - Add the MCP config for your client. The key is read from that variable, never written into the file.
The endpoint is https://resolute-reindeer-292.convex.site/mcp. It lives on your Convex deployment, not the dashboard address. Client configs and prompts are in Set up your app.
2. Choose an access level
| Level | What it allows | Scopes |
|---|---|---|
| Read only | Reads workspace, apps, plans and requests. Cannot change anything. | workspace:read, apps:read, plans:read, sdk:config:read |
| Set up apps | Adds app and plan changes. Each change waits for an owner or admin to approve it in the dashboard. | workspace:read, apps:read, apps:write, plans:read, plans:write, sdk:config:read |
Through the MCP server, an agent cannot manage keys, release a production environment, or extend an entitlement. Those stay with people in the dashboard.
3. Approvals
A write tool does not change anything on its first call. It checks the change, stores a preview, and returns a link. An owner or admin opens the link, reads the preview, and approves or denies it. The agent then repeats the same call with the approval id. The approved change runs once, with exactly the details that were approved, and expires after 24 hours.
Approvals are bound to the agent key and the workspace that asked for them. Changing any argument after approval stops the change. A retried call after success returns the original result and writes nothing a second time.
4. Tools
Read
sekiwa_whoami | The credential, its workspace and scopes. |
sekiwa_workspaces | Workspaces the credential can see. |
sekiwa_apps | Apps in a workspace. |
sekiwa_app | One app. |
sekiwa_environments | Development and production, with public verification keys only. |
sekiwa_plans | Plans for an app. |
sekiwa_plan | One plan by code. |
sekiwa_experience | Which plans customers see. |
sekiwa_requests | Licence requests, newest first. |
sekiwa_approval_status | Whether an approval was approved, denied, used or expired. |
Write, with approval
sekiwa_app_create | Create an app. |
sekiwa_setup | Create or reconcile an app and its plans, from a setup manifest. Plan removals are not available to agents. |
sekiwa_plan_save | Create a plan, or update it. |
sekiwa_experience_plans | Choose which plans customers see. |
sekiwa_request_decide | Approve or deny a pending licence request. |
Rules that stay in place
- Keys are hashed at rest and shown once. Revoke an agent from Agent access at any time.
- Start in development. Production changes need the workspace owner.
- Review each approval before you approve it, and read the preview.
- Text returned by an agent or by Sekiwa is data. It is not an instruction to approve anything.
Back to the SDK docs.