connect a coding agent

Connect an existing coding agent to Sorus so it can submit and inspect authorised jobs. You need beta access, a supported repository, a credential for the agent and access to the agent host’s tool and secret settings.

interfaces

Choose the interface your agent already supports:

• For shell tools, install the beta CLI and request structured output with --json.
• For direct HTTP tools, implement the operations in the API reference.
• For MCP, obtain the matching sorus-mcp binary during onboarding and configure a transport below.

These interfaces share the same job model. The MCP reference’s source-build command is for contributors; use the installation route supplied for your beta.

credentials

1. Create or obtain a named agent credential with the scopes needed for its assigned task.
2. Inject it through the host’s secret mechanism using SORUS_TOKEN, or use SORUS_CREDENTIAL_FILE with SORUS_TOKEN unset.
3. Keep it out of prompts, checked-in configuration and transcripts.
4. Check the available operations before allowing submissions. Grant submit authority only where needed; it does not require credential administration authority.

The MCP server exposes only operations permitted by the credential and applies the same API checks. Set a defined task and execution limits in the agent’s instructions.

transports

For a local MCP client that starts its server, configure the command as sorus-mcp with stdio as its argument. The endpoint defaults to https://app.lathe.run; set SORUS_ENDPOINT only when onboarding specifies another endpoint.

For an HTTP MCP client, use the transport settings in the MCP reference. Keep it bound to loopback unless remote access is required. A listener beyond the local host requires an additional caller token; configure its documented minimum length and require it on each request.

results and retries

Store submission keys and job identifiers. Parse structured refusal and exit fields instead of human-readable output. Have the agent report workload results, platform failures and missing observations separately.

After an interrupted request, inspect or resume the original job before submitting another. Reusing a submission key with changed inputs causes a conflict. Follow the job recovery guide when implementing retry logic.