MemmyHost-managed Agent
This page is for an application that embeds the Agent and approves tools itself. Standalone memmy serve does not use these settings. See OpenAI-Compatible API.
The former MEMMATE_* settings, the corresponding request header, and the memmate_permission request field are no longer recognized. Upgrade the host together with the Agent, and check the protocol below before rollout.
Environment variables
| Variable | What it does |
|---|---|
AGENT_MANAGED_EXECUTION_REQUIRED | Set to 1 to run in host-managed mode. Without it, the Agent runs standalone. |
AGENT_RUNTIME_TOKEN | At least 32 characters. The x-agent-runtime-token request header must match it exactly, or the response is 401. |
AGENT_MACOS_SANDBOX | When set to 1, the health field macosSandbox is true. |
Requests
In managed mode, POST /v1/chat/completions must be JSON. Multipart is rejected. The body includes managed_execution:
| Field | Requirement |
|---|---|
protocol | Must be 2. |
tools | A non-empty array. Each item needs string name and description, plus an object inputSchema. |
url | Only http://127.0.0.1:<port>/authorize. No username, password, query, or fragment. |
token | 64 lowercase hex characters. The Agent sends it as Authorization: Bearer when it calls this URL. |
A read grant does not cover descendant paths. read_file, list_dir, find_files, grep, and authorize_read_path include fileAccessProtocol: 1. Each path that will actually be read must pass authorize_read_path before the read.
When the host returns decision: "allow", result must be a string. A denial or execution failure is handled from decision, reason, or error.kind / error.code.
Health check
When AGENT_MANAGED_EXECUTION_REQUIRED=1, GET /health adds these fields beside status: "ok":
| Field | Value |
|---|---|
permissionProtocol | 1 |
managedExecutionProtocol | 2 |
managedExecutionRequired | true |
macosSandbox | true when AGENT_MACOS_SANDBOX is 1, otherwise false |
Without host-managed mode, the health check returns only status.