Your first guild
A walk through the operator app: from an empty box to a guild with an agent that talks, remembers, wakes on a schedule, and uses a tool. Then a second guild stood up from a recipe file.
Screens are described in full in Operator app.
Open the app
http://127.0.0.1:8080/ opens on the box's operator.
With no guild yet the page offers the Builder. The sidebar on the left lists guilds. Its footer holds:
- Inbox
- Builder
- Recipes
- Databases
- the account button, with your name and organisation
The account menu sets the theme and opens Settings and Admin.
Create a guild
+ on the Guilds header opens Create a guild:
- a name
- a short description (the goal everyone in the guild shares)
- a longer description (what it is for, how it should work, constraints)
- Private, on by default, which keeps the guild to its operators
The guild opens at /guilds/<id> on its Group chat tab. The header shows its name, its agent count, and a menu.
The guild's tabs:
- Group chat
- Settings
- State
- Routines
- Issues
- Tools
- Secrets
- Usage
Add an agent
Add agent under the guild in the sidebar (or New agent in the guild menu) opens Create an agent.
- Identity — a name (changeable later) and a focus, one line on what the agent is for. The focus seeds its Task. An avatar picker gives it a mascot and one of six identity colours.
- Model — the model, and a thinking level and temperature where the model takes them. With no key yet, a banner above every page says so and the picker names the two ways to give one (Install).
- Run limits — max output tokens, max steps, max tool calls, max input tokens.
- Disk — the workspace cap, 10240 MiB by default.
- Box — Lite, the default, for scripts, notes, connectors and a terminal on the smallest machine. Or Standard, which turns coding and the browser on together and gives the agent the large machine.
- Agent capabilities — CLI, scripting, and Memory, all on by default.
A new agent opens on its Prompt tab.
The agent inspector is a sheet over the group chat. Its tabs:
- Chat
- Prompt
- State
- Routine
- Browser (with browser access)
- CLI
- Tools
- Secrets
- Usage
- Settings
Write the Task
The Prompt tab edits the standing task. That is the operator-owned part of every prompt the agent runs with.
Write what the agent does, how, and what it keeps.
Save (or Cmd/Ctrl+S) archives the previous version. The version picker shows any earlier one. Restore brings it back.
Insert embeds a note or a table in the Task as ![[display-path]]. A shared note the guild maintains is then read into the prompt on every run.
Talk to it
Back on Group chat, a message with no tag wakes every agent in the guild. @name wakes that agent alone. The @ menu in the composer lists the members.
A responding row names the agent generating. Its post arrives on the left with an Inspect run pill.
For a private conversation, open the agent and its Chat tab:
- a new session
- a composer
- a paperclip for images when the model accepts them
The clock on the tab strip opens the session drawer. It lists every run: direct, or woken by the chat or a routine.
Inspect the run
A run's transcript starts with:
- the tools the run was offered
- the composed system prompt (
<guidelines>,<task>,<state>,<memory>), each a collapsible section
Every stretch of model steps that called tools renders as a run flow:
- step and call counts
- model and tool time
- tokens and cost
- a row per step, titled with the first line the model wrote
Opening a call shows its arguments and result as a folding tree or raw text, with its duration and the tool calls left.
An ended run offers Resume, which continues the same transcript.
Notes and memory
The agent's State tab holds what it keeps between runs.
- Notes lists the agent's, the guild's, and the organisation's notes side by side. A new note gets
.mdadded and opens in a Markdown editor. A provenance strip under the header says who saved it last. Open step jumps to the run and tool call that wrote it. Agents read and write notes with their note tools. The guild's State, Files shows the shared ones. - Memory lists every version of the agent's memory, newest first, with who wrote it. The agent writes it with
memory_writewhile Memory is on. The live version is editable here. An archived one opens read-only as the change against the version before it.
Add a routine
The agent's Routine tab creates schedules:
- Cron (UTC)
- or Random window, with a from and to time, the most wake-ups a day, and a timezone pre-filled from your profile
- a description
- the instruction the wake-up carries
The form previews the timing in words.
Each routine pauses and resumes on its own. The guild's Routines tab lists them all, with one switch that holds every routine of the guild. The guild header shows a Routines off chip while it is off.
A woken run appears in the session drawer and says what woke it.
Give it a tool
Tools lives at the agent, the guild, or Settings, Tools for the whole organisation. Integrations come first. Native tools come after.
- An integration is a connector from the catalog. Its card shows its login, a switch for the server, and its tools. Connect collects a key in a dialog, or opens the provider's sign-in for an OAuth connector. A connector the server cannot sign in to is marked Not set up, with the variables the server lacks (Configure).
- Native tools come in families: web search, notes, tables, memory, scripts, routines, group chat, secrets, browser, coding, issues. Each is an on/off switch captioned with where its state comes from: set here, or inherited from the guild or the organisation. A tool that waits for a server key has no switch and names the variable.
Secrets gives agents environment variables. Add secret takes a name, a value, a description, and who gets it: the organisation, a guild, or an agent.
Values never show again. The agent's sandbox has the variable. Its scripts and shell read it.
Apply a recipe
Recipes in the sidebar footer opens the library. Import JSON takes a guilds.recipe file and opens the recipe's page:
- the plan of everything it creates and needs
- each role's Task
- Approve and apply
The apply preview shows:
- the new guild's name
- each agent's name
- for each connector, a login to reuse or one to connect afterwards
The checklist that follows lists the logins, secret values, repositories, and outside setup the team still needs. Each step is verified, confirmed, or skipped.
Activate routines unpauses the schedules the recipe brought.
A browser-enabled agent then shows a Browser login needed strip with Open Browser, where you log the agent in to its sites.
Save as recipe in a guild's menu captures your own guild the same way. Export writes the file.
From here
- Inbox — issues agents report and messages that tag you.
- Databases — the organisation's tables, their records, and who reaches them.
- Open in Studio in a guild's menu — a coach that reads the guild as evidence and drafts better Tasks (Prompt Studio).
- Builder — design a recipe from a conversation (Guild builder).