Agent reference

📄️Agent reference overview

This section is the spec — for Agents. Field-by-field schemas, error codes, audit-trail formats, OIDC details, rate limits, build-path resolution algorithms. Everything an Agent needs to operate ERun without guessing.

🗃Concepts

11 items

📄️MCP protocol + tools

MCP is the typed-tool surface for an environment. Where shell-level work happens over SSH, MCP carries typed actions — inspection (idle, doctor, list, version, observe), operational wrappers around the CLI (build, push, deploy, release, expose, init, delete, and the cloud/context/platform/review families), and an escape hatch (exec_raw). Agents (and any other code that wants structured, auditable access) talk to ERun over MCP; every call lands in the same audit trail the Operator reads. The full tool index at the end of this page lists every registered tool.

📄️Agent patterns

Common patterns for Agent authors. None are required — ERun's primitives are flexible enough that an Agent can implement any of these on top of the MCP tools and the erun API. What follows is the set most teams converge on.

📄️Review loop topology

The review → comment → fix loop and the Build Agent / Review Agent row in the Workflow page both describe the shape abstractly. This page is the concrete version: two standing agent roles, each in its own environment, and exactly which review state each owns.

📄️Merge queue

The merge queue is the only path to MERGED — it is what makes two independently-green reviews that break the target branch together impossible, and it carries the one audited escape hatch in the product. It also has a lot of surface: a shape, a gate, a comment-thread check, three clients that can advance it, and a wedge-recovery path. This page is the single account of all of it; Reviews stays the wire-level endpoint reference, erun review the CLI reference, and the desktop reviews tab the app reference — each links here for the mechanics rather than repeating them.