Build from Claude Code or Codex
Drive Sebati's agent builder from your own coding harness over MCP.
Everything you can build in the Console you can also build from Claude Code, Codex, Cursor, or any MCP client. Sebati exposes its builder as an MCP server, so your coding agent configures Sebati while your harness keeps your notes, specs, and the customer's documents in context.
This is the usual FDE working mode: discovery notes on one side, a live workspace on the other, and natural language in between.
Connect
Create a workspace API key. In the platform, open the workspace menu in the bottom-left corner, choose Manage workspaces, then the API Keys tab, and create a key.
The raw key (sk-ws-...) is shown once; store it safely:
export COPILOT_WS_KEY="sk-ws-..."The key is scoped to one workspace. Everything built through it lands there, and you can revoke it at any time from the same screen.
Add the server to your client.
For Claude Code, in .mcp.json:
{
"mcpServers": {
"sebati": {
"type": "http",
"url": "https://<your-deployment-host>/v1/mcp/copilot",
"headers": { "Authorization": "Bearer ${COPILOT_WS_KEY}" }
}
}
}For Codex, in ~/.codex/config.toml:
[mcp_servers.sebati]
url = "https://<your-deployment-host>/v1/mcp/copilot"
http_headers = { Authorization = "Bearer ${COPILOT_WS_KEY}" }Any other MCP client works the same way: Streamable HTTP transport, bearer auth.
Build in natural language. Your coding agent drives the tools:
Create an agent called "Vendor Review" that evaluates supplier proposals against our procurement rubric. Set up its system prompt, create a knowledge base from the files in ./rubric, bind it, attach the default model, and publish.
What the tools cover
The server exposes the full builder surface as tools:
| Family | What you can do |
|---|---|
| Agents | Create, configure, set prompts and models, toggle features, publish |
| Skills | Create workspace skills, bind and unbind them per agent |
| Knowledge | Create knowledge bases, ingest content, manage sources, bind |
| MCP servers | Register external tool servers, verify them, bind to agents |
| Cases | Create sheet-backed work agents |
The server is stateless: tools take explicit ids, and your workspace comes from the API key. Your coding agent handles id passing automatically.
Two things to know
- Draft and publish still apply. Configuration calls stage into the agent's draft; nothing goes live until a publish call. The lifecycle is the same one described in Publish and versions.
- Action policy is explicit. When registering an external MCP server for agents to use, you choose read-only or full access at registration time. Decide deliberately, the same way you would in the Console.