Aegis Orchestrator
Zaru

Chat

The main conversational interface at ask.myzaru.com — conversation modes, goals in Agentic mode, model selection, and the Glass Laboratory context panel.

Chat

The chat interface at ask.myzaru.com is where you interact with the AEGIS platform. Every conversation runs in a specific mode that determines the available tools and behavior.


Conversation Modes

Choose a mode in the conversation panel above the input before sending your first message. The mode is locked once the first message is sent.

ModePurpose
ChatGeneral conversation with Zaru. No platform tools are invoked — this is a standard LLM chat with Zaru's persona and knowledge.
AgenticZaru has access to AEGIS tools and can create, discover, and execute agents and workflows on your behalf. Every request for a result, however simple (a sum, a unit conversion, a lookup), is dispatched to an agent, starting with aegis.agent.list; Zaru answers directly only about the work in hand, and offers Chat through a mode switch when you want conversation alone. The full tool surface is available based on your tier.
WorkflowSpecialized mode for designing and running multi-step workflows. Zaru guides you through defining workflow states, transitions, and agent assignments.
ExecuteSingle-shot code execution. Describe a computation in plain English, optionally provide typed inputs, and get back the result. See Running Code with Execute Mode for details.

A turn sends no call of a tool the conversation's mode did not offer, whatever tool name the model writes: the call is answered to the companion as "<tool> is not offered in this conversation." and nothing is sent. No mode offers aegis.workflow.signal, operator mode included: a workflow run's Human gate is yours to answer (Human Gates of a Workflow Run).

Zaru takes one small step at a time: it says what it is doing as it does it, in the same reply that makes the call, and shows what each step returned. It never says a change was made, verified or is waiting for your approval unless a tool's answer showed it. In Agentic and Workflow mode it stops where your answer or approval is needed, with what the step showed and the one next step it would take.


Goals in Agentic Mode

When a turn in Agentic, Workflow or Execute mode first starts creating an agent, running an agent, generating a workflow or running an intent, Zaru sets a goal: the message that began the turn, word for word. A Goal set card shows it in the conversation. Every execution the turn starts is bound to the goal, so a chain such as creating an agent and then running it stays one piece of work.

In Agentic mode, after the turn has answered, the platform's judge reads the goal, the final state and output of every execution bound to it with what the run did (how many tool calls it executed, each tool call with its result status, and the files its declared outputs recorded), and the answer you were shown, and decides whether the goal is met. The judge counts a file or an action as done only when those facts show it. A Judge card shows each verdict: the round it judged, counted from 0, the outcome (met, not met or cannot be met), the score and confidence, and what the judge found missing. When the goal is not met and rounds remain, Zaru continues by itself, without another message from you: a Continuing card shows the round and what it was told, and the round picks up from the executions already started instead of starting them again. By default a goal is granted at most three continuation rounds and stays open at most 30 minutes from when it was set; each round runs at most five model calls and 10 minutes.

While the goal is open the page re-reads the conversation every 30 seconds and when you return to the tab, so new cards appear without a reload. A message you send while a continuation round runs is refused, as it is while any turn runs; a later message that starts a new creation or execution replaces the open goal with a new one. If your sign-in can no longer be refreshed, the goal pauses with a card saying so, and the next turn in the conversation carries it on while it is still open.

When Zaru cancels one execution of the goal, with aegis.task.cancel or aegis.workflow.cancel, the goal stays open and the cancelled execution stays bound to it, so Zaru can start the corrected execution under the same goal. To end the goal itself Zaru calls aegis.goal.cancel, and Stop cancels it with the turn.

In Workflow and Execute modes the goal is set and shown, but not judged or continued. No goal is set in Chat mode, in operator mode, or in a conversation on your own provider key. The cards are not part of the conversation history the model reads.


While Zaru Works

From the moment you send a message until the turn ends, the conversation shows a working indicator with words: Zaru is working on it…, or Thinking… while the model thinks, with the minutes elapsed from the first minute on (Thinking · 2 min…). Where a goal's state names what Zaru is doing, those words are shown instead. The indicator keeps its words while the page's stream is open or the page knows the turn is still in flight.

A wait tool's card (aegis.agent.wait, aegis.task.wait, aegis.execute.wait, aegis.workflow.wait, or any tool whose name ends in .wait) whose result says the execution has not ended (pending or running) is titled Still running in a neutral style. Opened, it reads "Zaru checked, and the work had not finished yet." It is never shown as done or as a failure. A wait returns before the end only at its own time limit (45 seconds on Zaru's MCP server, Waiting on Long Executions); the run goes on by itself, and Zaru says it is still running and ends its turn rather than waiting again.

Such a card then reads the run's own status every five seconds until the run is over. Once it has ended, the card's title becomes Done, Failed or Cancelled, and a line under it says how it ended, with the state and the time where the run names them: "The run failed in CODE at 2026-10-13 03:14:56 UTC." A card that cannot read the run keeps the tool's answer and says "Zaru couldn't read this run's current state." A workflow wait reads the workflow run's own status, and the run's summary card under it shows the run's end too. Earlier checks folded inside a card are not read again.

When a call is refused because something is not connected, not chosen for the conversation or not in its profile, or when Zaru asks with zaru.connection.needed what is missing (MCP Tools), a connection card appears under the call's tool card, also when the call failed. It has one line and one button: "A mailbox is not connected to Zaru." (or a calendar, GitHub, Carbon Forge (formerly Nuclear Notes) or the server's name) with Open Connections; "Your mailbox is connected but not chosen for this conversation." with Choose it under Context, which opens the conversation panel above the input; "The profile chosen for this conversation does not allow this." with Open the profile; or "The profile chosen for this conversation no longer exists." with Open Profiles. The link brings you back: once the connection is saved, on the Connections page, in its IMAP or CalDAV form or through a provider's consent screen, or the profile is saved, Zaru Web returns to this conversation where you left it, and you ask again; nothing is sent for you. Once the family is connected, or the profile admits the tool, the card collapses to "Connected.". Zaru tells you in one sentence what is missing and that the card takes you there, and ends its turn.

By default a turn makes at most 12 model calls. When the last one still calls tools, the turn ends with a card, "I used all my steps for this turn before finishing.", and a Keep going button: pressing it continues the same turn from its stored steps with a fresh allowance, without a new message from you, and the card then reads "Kept going.". A goal's continuation rounds show no such card.

Stop ends the turn on the server at once, your turn or a goal's round, even when the page is not streaming it, and unlocks the input. The model call in flight is aborted and its partial text is not kept, no further model or tool call is made, the conversation's goal is cancelled, and every execution the turn started is cancelled, whether or not the page saw it start. The conversation then shows "Stopped by you", with "Some of its work could not be stopped." when an execution could not be cancelled, and nothing of the stopped turn after it. A turn that ended without an answer, for example because the page reloaded while it ran, ends with "This reply was cut off before Zaru finished. Send your message again to retry." When Zaru's default model refuses a call before its reply starts streaming, busy (Workers AI's code 3040), an internal server error (its code 8005), or an HTTP 429 or 5xx, the call is made again on the same model after one second and then two, at most three times, never once the reply has started and never on another model. When every call was refused, the failure reads "The AI service failed; Zaru tried three times. Please try again." for an 8005, never its code, and "The AI service is temporarily busy. Please try again in a moment." for a busy model. A turn whose model call failed before it answered keeps its failure instead: the page shows the failure banner with the failure's code, as it does while the turn runs, also after a reload, until you dismiss it or run the turn again.


Human Gates of a Workflow Run

When a workflow run you started waits at a Human state (a gate, such as the two approvals of the Forge), the gate is yours to answer, and you answer it on a control of your own. It is one of your pending approvals, beside the tool calls waiting for your approval: it counts in the toolbar's pending count, it appears as a card under the run in the conversation that started it, it is listed on the Approvals page (/approvals) with a link to the run's page, and it appears on the run's execution page.

A run Zaru starts with aegis.workflow.run shows its narrative and its pending gate card under the run's tool card in the conversation. A run on a repository opens its narrative with the orchestrator's repository line, "Prepared repository <label> at /workspace/<label>: branch aegis/<id> created from <sha>" (or "checked out at" for a work branch the remote already had), before the run's steps, as the orchestrator wrote it. While the run has not ended, the page reads your pending approvals every five seconds, so a gate the run posts appears under it within one read, whether or not the turn that started it is still open.

The gate card is headed Waiting for your answer and shows the workflow and the gate's state, then the gate's prompt whole. It has one button per word the gate takes, in the gate's order and labelled with the word itself (for the Forge, yes and no), and a field What to change (optional) whose text is sent with your answer as its feedback. Where the gate gives a deadline, the card says "If nobody answers by <time>, the run fails.", or "…, the run answers <default>." for a gate with a default answer. A click sends one answer, and the buttons are disabled while it is in flight. Once you have answered, the card reads "You answered <word>." and, when the run has moved past the gate, "The run went on."; a gate whose run ended before any answer reads "The run ended before an answer."; a gate answered elsewhere, or closed by its timeout, reads the line Approvals shows for it under Recent. An answer to a gate that is no longer waiting is refused with "This gate was already answered or is no longer waiting.", and a word the gate does not take with "This gate takes <words>."

A gate card and a tool's approval card each open and close by a tap of their header, which always shows the card's state, the workflow and the gate's state or the tool, and your answer once given. A card waiting for your answer is open; an answered, timed-out or closed card is collapsed to its header, also after a reload. Your own opening or closing holds until you reload. The same cards on the Approvals page and a run's page behave the same way.

Once a gate is answered, or closes on its timeout, it stays on the Approvals page under Recent, in order with the tool approvals you answered: the workflow and the gate's state, "You answered <word>.", or for a gate that closed on its timeout "Nobody answered; the run timed out.", and Open the run, a link to the run's page.

In Workflow and Agentic modes, when Zaru starts a run it says the run has started, that the card below shows its state, and that each gate appears on that card and in Approvals for you to answer there, and it ends its turn. Nothing of Zaru watches the run after its turn, and it promises nothing for after it. When you ask about the run later, or say you have answered its gate, Zaru reads the run once; when the run waits at a gate, it shows the gate's prompt as that read answered it and says where the run waits for your answer: the card under the run in this conversation, Approvals, or the run's execution page. Zaru never answers a gate, never asks you for the answer in the chat, and never says a gate was answered unless a tool's answer shows it. When a run Zaru started reads as failed, by a wait or by a goal's check, Zaru tells you in one sentence that it failed, carrying the run's own failure sentence (the feedback of the step it failed on, otherwise the state it failed in and the first line of that state's error), and ends its turn there; it does not search, verify, retry or start the run again unless you ask.


The Conversation Panel

The line above the input names the conversation's context, model and mode, for example "Context: None · Model: Zaru · Mode: Chat". A chosen profile reads "Profile: <name>" in place of the context, a provider of yours reads by its name and the model's, and in Code mode the line names no model. A tap on the line opens the conversation panel over the conversation, on a phone and a desktop alike, with three sections in order:

  • Context: None, your profiles, and your connections by type, then Manage tokens (Profiles).
  • Provider and model: Zaru first, then each of your providers by name with its models, the current one checked. A provider with no models shows "Add models in The Vault, Providers", a link to The Vault's Providers page, and the section ends with Manage providers, a link to the same page. Code mode has no Provider and model section.
  • Mode: each mode you may choose with its description: Chat "Conversation only", Agentic "Run AI agents", Workflow "Design & run workflows", Execute "Run code from natural language", Code "Build apps with live preview" when Code is on, Live "Run TypeScript in-browser" when Live is on, and Operator "Full platform access" for an operator or admin. Code is not offered on a phone-width screen. After the first message the section shows the conversation's mode and its description and offers no choice.

A choice in any section leaves the panel open; sending, Escape and a tap outside it close it. The panel cannot be opened while a turn runs, so the provider and model are chosen between turns. The chat's header carries neither the model choice nor the mode.

Zaru is the platform-managed model: the default, with no configuration. Your own providers are added on The Vault's Providers page (/vault/providers): Zaru sends your messages to them using your own key, which stays in your vault and is sent only to that provider, and your provider bills you for what you use. A turn on one of your providers is handed an output budget of 16,384 tokens, on every protocol. BYOK users are exempt from LLM-specific rate limits.


Conversations

Each conversation is saved automatically and appears in the left sidebar. Conversations are organized chronologically with the most recent at the top.

Managing Conversations

  • Pin, Rename, Delete — hover over a conversation (or move keyboard focus to its buttons) and use the buttons that appear at its right end. Pinned conversations stay at the top of the sidebar regardless of age. Renaming gives a conversation a descriptive title; by default its title is the start of your first message. Deleting a conversation asks for confirmation first: choose the trash icon on the conversation's row, then Delete in the dialog that names it. Cancel or Escape leaves the conversation as it was. Deletion removes the conversation and all its messages and cannot be undone.
  • Each conversation is a link: open it in a new tab with a middle click or Ctrl/Cmd-click.

Each message shows its time under it, hour and minute in your own locale and time zone, and each day opens with a separator such as "Today • 8:43 PM". Your own messages, with their times, sit at the right of the conversation; Zaru's messages and the system's sit at the left.


File Uploads in Chat

Agentic and Workflow modes surface a paperclip icon next to the chat input. Use it to attach files to your next message. The affordance is hidden in Chat, Execute, and other modes — uploaded files are only meaningful when the conversation can dispatch agents or workflows that consume them.

Attaching Files

There are three ways to attach a file:

  • Click the paperclip icon and pick one or more files from your local machine.
  • Drag and drop files onto the chat input.
  • Paste a file directly with Ctrl/Cmd+V (useful for screenshots).

Multiple files per message are supported. Each upload renders a chip showing filename, progress, and a remove button. The send button stays disabled until every in-flight upload completes.

Where Uploads Live

Files land in the chat-attachments volume in your tenant. The volume is provisioned automatically the first time you upload — you do not need to create it. Uploads count against your storage quota like any other persistent volume.

You can browse, rename, move, or delete uploaded files from The Workshop at any time. Power users who want to redirect a chat upload to a different volume can use the chevron next to the paperclip to choose a destination.

How Agents See Your Files

When you send a message with attachments, Zaru passes the file references along with the dispatch as a structured attachments field — the model never has to reason about file IDs. The model is told each file's name and type with your message, for example "[Attached files this turn: 2 (DIALOGUE.ogg, audio/ogg; notes.wav, audio/wav)]", never its volume, path, size or digest, so it can name the file a workflow input asks for. Agents and workflows that declare attachments in their input schema receive the references directly and read file contents through the platform's read-only attachment tool.


The Glass Laboratory

The Glass Laboratory is the real-time context panel on the right side of the chat interface. Expand it using the panel icon in the top-right corner.

What It Shows

When an agent or workflow execution is running, the Glass Laboratory renders each step of the process as it happens:

  • Tool invocations and their outputs
  • Iteration progress (attempt number, validation feedback, refinements)
  • Workflow state transitions
  • Policy decisions made by the SEAL Gateway

Context Panel Tabs

In Agentic, Workflow, and Execute modes, the context panel provides browsable tabs for platform resources:

  • Agents — lists deployed agents. Click Execute to run one with an inline parameter form, or Send to Chat to inject a structured prompt for Zaru to assist with.
  • Workflows — lists deployed workflows with the same Execute and Send to Chat actions.
  • Tools — lists available AEGIS tools for the current mode and tier.

The context panel is hidden in Chat mode, where no platform resources are in scope.


Per-User Context

Every conversation, in every mode, opens with your Memory loaded — a single markdown document Zaru maintains about your preferences, work style, and ongoing projects. It is the reason Zaru sounds like he already knows you instead of starting fresh each time.

On this page