MultiAgent collects the chats of all agents into a single database, so the whole working context of the fleet is available in one list — without signing into each agent separately.
Chat List
The list shows the chats of all agents with filtering by agent and tags, and pagination. Agent (sub-agent) chats — the ones agents create for their own subtasks — are hidden from the main list and moved to a separate analytics section.
Chats can carry labels — short colored marks with an icon, for sorting the list by topic. The label set lives on the agent itself; assigning and clearing takes one click, and a label filter shows only the chats you need. Details on the Own Agent page.
Working in a Chat
Open a chat and you work with the agent as in its native interface: send messages, read replies, and stop processing with one click when the model goes off track. Everything renders in real time — tool progress and new messages arrive without reloading the page.
- tool calls (Read, Grep, Bash, Task and more) with their captions;
- the model’s thinking blocks;
- Todo tasks with completion progress;
- the raw JSON of any message.
Prompt Queue
The Prompts button in the input field opens the agent’s prompt queue — a list of prepared requests of the form “text + model + skill”. Sending a row creates a new chat on the agent with that row’s model and skill, sends the prompt, and removes the row from the queue; the created chat’s title is capped by the agent’s setting — the maximum title length. The queue is stored on the agent — MultiAgent only displays it, and every agent has its own: nothing is stored on the MultiAgent server.
The AutoSend settings are the agent’s own settings: its internal scheduler starts the accumulated prompts in new chats by itself, even when nobody is watching the interface. The menu next to the AutoSend switch carries three extra options. “Send during peak hours” — when it is off, prompts with models currently in peak hours wait in the queue, while off-peak model prompts send as usual. “Auto-commit before sending” — before each prompt is sent, the project repository is committed; the commit message is the title of the most recently finished chat. “AutoFix disabled until done” — while auto-send is working through a non-empty prompt queue, AutoFix waits for it to finish instead of injecting build errors mid-batch.
The queue is built so a half-written row is never sent by accident: when the queue empties — the last row was sent or deleted — AutoSend turns itself off. In place of the empty queue sits a draft row: the first prompt is typed right there, with no “New prompt” click, and becomes a real row on first edit. How many prompts are waiting is shown by the Prompts button’s tooltip and the dialog’s bottom line — “{count} prompt{s}, bottom first”. Edits from two open tabs do not overwrite each other: rows merge line by line, so fresh changes are never rolled back by someone else’s.
Bots
The Bots dialog, opened from the agent page, manages the agent’s bots: their configuration is read and written through the agent’s own API. MultiAgent adds no logic of its own and does not store this data — it shows the agent’s interface as is.
Scheduled Tasks
The Tasks dialog shows the agent’s schedules with run conditions — for example, “notify in the chat only if the script output contains a line” — and a log of completed runs. Like bots, tasks are the agent’s own entities: their data is read and saved through MultiAgent’s relay, and updates arrive over the same channel as chat events.
MCP Servers
The “MCP Servers” button on the agent page opens its MCP server configuration — also through the agent’s API, with no intermediate storage. All changes apply on the agent itself.
Skills and Models
Every chat has its own chosen model and skill (including an explicit “No skill”); the choice is remembered and becomes the default for the next new chat. In the first line of the first message, a skill slug (/deploy) assigns the skill to the chat and is removed from the text. The “Skills” dialog opens the agent’s skill file tree and editors — edits are sent to the agent itself.
Voice Input and Sounds
The input field has a microphone for dictating messages. Sound notifications accompany new messages, finished replies and errors — you can turn them off in the settings.
Message Statistics
Every message carries tokens, cost and reply speed (tokens per second). The same metrics are aggregated across all agents in analytics.
Context Utilization
The “Context Utilization” indicator shows how full the chat’s context window is: the percentage, spent and available tokens. When the context nears the limit, it is time to finish the chat or continue from an earlier message.
Continuing from a Message
The “Continue from this message” command unfolds a new branch of the conversation from the selected message — handy for rolling back to a moment before the model’s mistake, or experimenting with a different direction of the solution.