Agent Guild
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| AGENT_GUILD_NO_UPDATE_CHECK | No | Turns off version checks and self-upgrade. | 1 |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Agent Guild runs coding agents (Claude Code, Codex CLI, Antigravity CLI, Grok Build, Docker Agent) and shells in real terminals on this computer, and opens them on every other screen. Use guild_status first, list_sessions to see the sessions, read_reply for the answer that ended a session's last turn, and read_screen for what its terminal shows now. Use read_replies with your own saved cursor to catch up on replies. Reads never consume them; save a reply's cursor after handling it and check for gaps. To customize a Yard, the characters and companions of a Yard or Agent Guild: World, or to install Grove Companion on the desktop, follow https://agentguild.oddessentials.ai/agents/. Ask the person before starting or stopping a session. Never ask for, print or store the Agent Guild page link or access token.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| guild_statusA | Whether Agent Guild is running on this computer, its version, and how many sessions it has. To customize a Yard, the characters and companions of a Yard or Agent Guild: World, or to install Grove Companion on the desktop, follow https://agentguild.oddessentials.ai/agents/. |
| list_sessionsB | Every session: status, recent activity (working), model, sub-agents, latest reply id, history cursor, and inputInstance for keyed input. |
| read_replyA | The whole text of the answer that ended a session's last turn, as Markdown. Only tools with reporting on (Claude Code, Codex CLI, Antigravity CLI, Grok Build, Docker Agent) have one. |
| read_repliesA | Read retained replies in order, with ids and cursors. Pass your last handled cursor as after; omit it to start at the oldest retained reply. Reads do not consume replies. Check gap for evicted or reset history and save each cursor only after handling its reply. |
| read_screenA | The text a session's terminal shows right now, as plain lines. Use it to see a question the tool is waiting on. |
| start_sessionA | Start a new session of a coding tool or shell in Agent Guild. Ask the person first. It opens on every screen connected to Agent Guild. |
| stop_sessionA | End a session's process. Ask the person first. The session stays listed, with its final screen, until it is removed in Agent Guild. |
| fireworksA | Launch fireworks over the Yard on every screen connected to Agent Guild. Use it to celebrate something the person cares about, such as tests, a build or a release that passed. Says how many pages saw them. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 8 tools
Most tools target distinct actions: status, listing, reading, starting, stopping, and celebrating. The main possible confusion is between read_reply (last turn's final answer) and read_replies (cursor-based batch of retained replies), but the descriptions make the distinction workable. guild_status versus list_sessions also overlaps slightly on session information but remains separable.
The set is mostly verb_noun (list_sessions, read_reply, read_screen, start_session, stop_session), which is readable. However, guild_status and fireworks break the verb-first pattern, and read_reply versus read_replies introduces singular/plural naming that could be clearer.
Eight tools is a well-scoped count for a session management and monitoring server. Each tool appears to have a distinct role, with no obvious redundancy or bloat.
The surface covers status, session listing, reading outputs, screen inspection, and session start/stop. But there is no tool to send keyed input or answer a prompt, even though read_screen is explicitly meant to see questions the session is waiting on; that is a significant dead end. There is also no remove_session tool, though stop_session mentions removal happens elsewhere.