Skip to content

The Builder

The Builder is where you create and shape an agent by describing what you want in plain language. There are no forms to fill in first — you type a request, the Builder makes the change, and you watch it happen live.

Opening an agent

Click an agent in the rail, or find it through Search, and you land on the agent itself:

  • The live agent (right). Its own working interface, exactly as the people you share it with will see it. The row above it switches between App and Canvas (once the chat has made something for the canvas) and holds Phone width, Reload surface and Open surface in a new tab.
  • Owner chat (left). The same chat and canvas as in your own chats with Donkit, running as a panel to the left of the live interface — a private conversation with your agent, as its owner. Handy for trying it out for real; drag the divider to give either side more room. The panel button in the top-left corner hides the chat and brings it back, and stays in the same spot either way.
  • The agent's menu. Click the agent's name at the top: My memory, My files and your API keys for this agent, plus Rename, Make a copy, Pause, or Delete, or copy its id. For an agent shared with you, the menu offers only what your access level allows.

You can also keep talking to the Builder while it works. Send a message during a build and it lands right away: the Builder breaks off what it was saying, or finishes the step it was running and reads your message next. Until it's read, the message sits in the chat in a lighter style and can be cancelled or edited, and pressing Stop gives the text back to the message box.

Making a copy

Make a copy creates a new agent that belongs to you, starting from this one. It carries over the prompt and settings, skills, workflows, connection slots without credentials, environment key names without values, the app and backend code, the database structure without data, and the documents of knowledge bases marked Use for everyone (re-indexed for you at your cost, so the knowledge appears gradually). It never carries secrets, data, conversations, memory, or private knowledge bases. You can copy your own agents, and agents shared with you at Can copy or Can edit. See Publish an agent.

Editing together

An owner can give a person the Can edit level (see Publish an agent). An editor opens the agent in the builder and can do everything the owner can there, except:

  • delete the agent,
  • change its publication and access settings,
  • manage the owner's API keys,
  • make a connection shared,
  • change who can read or write a table of the agent's database, or open a table that already holds records to everyone signed in,
  • permanently delete the agent's data.

On Share, an editor sees the access settings. Every control on that screen stays locked. The screen names who shared the agent and says the editor can edit. It says only that person changes access. The owner pays for the editor's builds until the editor turns on Pay from my wallet in the builder.

An editor can still update the publication, so the published agent picks up their changes.

How it works:

  • One person edits at a time. While the builder is working for someone, everyone else sees who it is and cannot send messages or save until it is done.
  • The builder chat is shared, so everyone who edits sees the same conversation. It holds only the work on the agent: your users' conversations with the agent never open in it.
  • The builder works with the agent's database as the person building. When it checks or loads records, it sees what that person would see using the agent — their own records in tables scoped to own records, nothing in tables closed to users — the owner included. Building an agent is not a way to browse the records of the people who use it. A migration the builder applies still runs with the database owner's rights, so give Can edit only to people you trust with the agent's data.
  • Only the owner and the editors can open the builder. Nobody gets in by role alone, a platform administrator included. To bring one in, share the agent with them at Can edit, like anyone else.
  • Skills and sub-agents an editor creates belong to the agent's owner. The editor also gets Can edit on each sub-agent they create, because the builder builds it for them. Connections an editor adds are the editor's own and stay theirs.
  • Who pays for an editor's builds: the owner by default, or the editor. An editor may switch to their own wallet. While the owner pays, the editor's own balance doesn't matter: an empty wallet of theirs doesn't block the builder or bring up an upgrade prompt, and if the agent's account runs out, the builder says so. The owner's own builds are always paid by the account the agent belongs to, even when the owner opens the builder from another workspace. That choice covers builds only: an editor's workflow runs are paid the way the agent's publication says, like their chats with it — see Who pays for a run.
  • Workflows. An editor builds and changes the agent's workflows as the owner does, and can delete one — which deletes every person's schedule on it (the automation is deleted from their Tasks) and removes their webhook, configuration values and runs of that workflow along with it. An editor's own runs, schedule and webhook are theirs: they see and stop only their own runs, never other people's, and their schedule and webhook are their own, not the owner's — a schedule the builder sets up during an editor's turn included. Running a workflow, and keeping a schedule or a webhook of their own, needs Let shared users run workflows on, as for everyone the agent is shared with. Starting or stopping a run isn't an edit, so it works while the builder is busy for someone else.

Share for editing only with people you trust.

Pausing an agent

Under the agent's name there is Pause agent. It is a stop button for an agent that is doing something you want to end now — repeating a failing action in a loop, or being called over and over by another agent — and every one of those turns costs credits.

While an agent is paused it accepts no work at all: no chat, typed or spoken (yours or your users'), no calls from your other agents, no scheduled workflow runs, no requests through its API. Anyone who tries is told the agent is paused by its owner. A scheduled run that comes due while the agent is paused is skipped rather than started and failed, so pausing does not fill the workflow's history with failures — and, importantly, does not switch anyone's schedule off the way a genuinely broken run does. The pause covers every schedule and webhook on the agent, whoever set it up.

Nothing is lost while an agent is paused: its configuration, knowledge, conversations and workflow schedules are untouched. Resume agent in the same menu puts it straight back to work, and the next scheduled run fires as normal. The pause also stops the work already in flight — a chat turn, a voice call, a workflow run: from its next step it can no longer use the agent's tools, knowledge, memory, files or database.

Editing: the build overlay

Click Edit agent and the Builder opens over the agent — the live agent keeps running underneath, so nothing restarts when you come and go. Under the builder-mode banner, a switch has Builder chat, Settings, Publish and Share. On a phone the switch is smaller and sits centered at the bottom of the overlay. Settings opens the list of every part of the agent. Share is who can open the agent and what they can do. Publish sets its address. Done in the banner drops you back into the agent.

Inside the overlay:

  • Builder chat. You talk to the Builder here, not to your agent. You describe the agent you want — "Make a support assistant that answers from our FAQ and can email customers" — and the Builder edits the agent's configuration for you: writing its instructions, turning on the right capabilities, generating its interface. Keep iterating in the same conversation ("make it friendlier", "add a dashboard", "connect Slack"); it remembers what's already set up.
  • Settings. The page is titled Agent settings. It lists every part of the agent. A row opens that part, and Back returns to the list — see below.

You can attach files or images to a Builder message when that helps explain what you want. A banner reminds you that you're in builder mode and not in the agent's own chat. On a phone it only says you're in builder mode.

How the Builder hands out and checks work. Larger pieces of work — the interface, the backend, the database, workflows, skills — the Builder gives to specialists, and each one reports back in the same short form you can expand in the chat: a one-line summary first, then whether it finished, which files it changed, which checks it ran and how they came out, anything that stopped it, and a question for you when a decision is yours to make. The Builder asks you that question in the chat and hands the work back with your answer. After a turn that changed the interface, the Builder opens the preview to confirm it builds and works; if it doesn't, the Builder fixes it a few times before asking you how to go on. A turn that didn't change the interface skips this check, and when a specialist already passed it, the Builder doesn't run it a second time.

In Donkit Desktop for Windows, the Builder's model picker also lists your connected My subscriptions. Choose a subscription there to build and refine the agent on your own provider plan. The Builder remembers its selection; the agent's use chat has a separate model choice. Personal subscriptions require a paid Donkit plan and currently accept text messages; use a platform model for messages with attached files or images. See Connecting your own subscription.

If your connection drops while the Builder is working, the build doesn't stop: the chat reconnects on its own and picks up where it left off, and a message you sent just as the connection failed is delivered once it's back — only once.

Donkit has no build view

The Donkit assistant is a built-in agent scoped to you, so it has no build view. Other built-in agents that come with Donkit have no Builder view of their own.

Language

Talk to the Builder in whichever language you prefer. Everything your users will read — the agent's instructions, its greeting and suggested prompts, every label in the generated interface — is written in the agent's audience language: your own by default, or a different one as soon as you say who the agent is for ("this one is for our Spanish-speaking customers").

The interface the Builder generates works on a phone as well as on a desktop, because many people open agents on their phone, in a mobile browser or in the mobile app. On a narrow screen, columns stack, side menus fold into a menu button or tabs, and wide tables scroll sideways or turn into cards. You don't need to ask for this. If an interface was built earlier and doesn't fit a phone, ask the Builder to make it work on phones.

To see how the interface looks on a phone, press Phone width next to Reload surface. The live agent is then shown in a phone-sized frame; press the button again to return to the full width of the pane. It previews width only: there is no touch input and no on-screen keyboard. Your browser remembers the choice. On a narrow window (tablet or phone width) the switch is not shown, because the window is too narrow to show the chat and the interface side by side.

The settings rail

Settings on the switch under the builder-mode banner opens this list. The page is titled Agent settings. A row opens that part of the agent, and Back returns to the list. You rarely need them for the basics — the chat can change most of these for you — but they're there for precise, manual control. The rows are grouped by what you are changing. Builder chat, Publish and Share are the other positions on that switch, so you can always get back to describing instead of clicking.

Agent

  • System prompt — the agent's core instructions: its role, tone, rules, and what it should and shouldn't do. This is the single most important setting — the agent's personality and job description.
  • Model — which AI model powers the agent (defaults to Auto).
  • Chat settings — what people see in the agent's chat before they type anything, and whether they may change its model. See The agent's chat below.
  • Voice — voice settings. See Voice.

Abilities

  • Tools — the abilities you switch on: Memory, Sub-agents, Web search, Web browser, Browser control, Send email, Image generation, Video generation, Audio transcription, Database, Workflows, and Computer use. A count badge shows how many are active. See Capabilities.
  • Skills — reusable packaged know-how you can add. See Skills.

Connections

  • MCP servers — connect external tools/services through the MCP standard.
  • Integrations — connect third-party services. See Integrations.
  • Connected agents — let this agent call other agents you've built. See Agent teams.

Data

  • Knowledge — documents the agent can search and answer from. See Documents & RAG.
  • Database — structured, per-agent data tables. See Agent database.
  • Files — files the agent works with.
  • Agent learnings — lessons the agent has drawn from its own work that help everyone who uses it; the same tab switches to your own memory with the agent. See Agent learnings & agent files.

Automation

  • Workflows — scheduled or automated multi-step routines. See Workflows. When the Builder wires up a workflow step, it can use your MCP servers, the built-in web search, and the integrations Donkit offers your agents, such as Google Workspace — including one you haven't connected yet. That is the intended order, not a gap: the Builder builds the step, tells you what to connect, and the workflow asks for it on its first run if it is still missing.

For programs

  • API — keys and endpoints for calling your agent programmatically. See API access.
  • Environment — configuration values, such as keys the agent's tools need.

Quality and history

  • Evaluate — test the agent and monitor its quality. See Evaluation.
  • Versions — snapshots of the agent over time. See Versions.

Sharing

Publish and Share follow Settings on the same switch. Publish sets the address. Share is who can open the agent. See Publish an agent.

Building from the iPhone, iPad and Mac apps

The Builder is also available in the Donkit app for iPhone, iPad and Mac — you can create and edit an agent without opening the website.

The Agents tab lists the agents you own, newest-changed first, with pull-to-refresh; New agent creates one and opens the builder on it straight away.

Build on a row, or on the agent's own chat header, opens the builder in the app: the Builder chat on the left, a live preview of your agent's surface on the right. On a phone the two are a Chat / Preview switch; on iPad and Mac they sit side by side with a draggable divider.

The preview is the real dev server, so it takes a moment to boot the first time and shows what it is doing (starting, loading, building, reloading). Reload and Open in browser are in its toolbar; on a Mac ⌘R reloads and ⌘⇧R restarts the dev server. Links to other sites open in your browser, and a file the preview offers for download is saved — see The agent's interface in the apps.

While the builder is working, the app watches what it changed and reloads or restarts the preview for you at the end of the turn — you do not have to refresh anything.

Build chats are separate from your ordinary chats: they live behind the builder's own history sheet and never appear in your chat history. You can rename the agent from the builder's title.

The settings the rail exposes on the web are available as native panels from the builder's menu: Tools, System prompt, Model, Chat, Integrations (mailbox, Telegram and Google) and Agent learnings. They edit the same configuration as the web — while a builder turn is running they are read-only, so an edit cannot race what the builder is writing.

On iPhone, use the single Back button beside the screen title to return to the previous screen.

An unsaved system prompt is kept as a draft under the same key the web uses, so a prompt you started on the phone is still there in the browser.

In the Android app the builder's menu also has MCP servers and Knowledge — see Tools & MCP and Documents & RAG. In Integrations, the mailbox and Telegram slots let you pick which of your own accounts the agent uses, or add a new one, without leaving the app; Manage in Connections opens your Connections library.

Not on iPhone and iPad: skills, connected agents, the database, evaluation, versions, API keys, publishing and the environment editor — those links open the website. The Mac app has all of them natively, see Building on the Mac. The Android app has a publish panel for access, invites and the published link; evaluation, versions and the environment editor still open the website there.

Building on the Mac

The Mac app has the whole builder, not a cut-down version of it. Everything you can do in the web builder you can do there.

One page per agent. The builder chat is on the left, and the section you picked fills the panel on the right. The sidebar lists the agent and its sections in groups:

  • Build — Chat (the builder chat on its own, full width), Preview and System prompt.
  • Agent — Model, Tools, Skills, MCP servers, Connected agents and Integrations.
  • Data — Knowledge, Database, Files and Agent learnings.
  • Behaviour — Workflows, Evaluate, API, Chat, Voice and Environment.
  • Versions, Publish, and a link that opens the agent in the browser.

Sections with items show a count next to their name. The sidebar follows back and forward, so the section you were on is the one that is lit.

Preview. The panel shows the agent as it is right now. Switch between phone and desktop width, reload with ⌘R or restart the dev server with ⌘⇧R. While the builder is working the preview says Building… and refreshes at the end of the turn.

Builder chat. Rate any answer with thumbs; a bad answer asks for a reason and optional details, which helps improve the builder. If the agent is shared with you for editing, a menu in the message box says who pays for your builds — the owner or your own wallet — as described in Editing together. Builder turns use credits like any other agent turn.

Editing lock. While someone else is building on the agent, a banner names them and every section is read-only until they finish, the same one-editor-at-a-time rule as on the web.

Use agent and Publish. Use agent in the toolbar opens the agent's own chat. Publish opens the publish section, and reads Published once the agent is live. There you choose who has access, invite people and copy invite links, decide which knowledge collections everyone who uses the agent can draw on, and see the list of things visitors are shown and asked to accept before they start. The options are the same as on the web, see Publish an agent.

Workflows. The section lists the agent's workflows with their health, cost and schedules, and a workflow opens its own page with the diagram, runs and triggers — see Workflows in the Mac app. Fix with AI and Rebuild send a message to the builder chat beside it; while the builder is still answering, the message waits in the chat for you to send.

Evaluate, API, Environment and Versions. Evaluate shows how far the last build got, the task checks tried on the agent, and how it does in real conversations. API manages the agent's API keys. Environment holds the keys the agent's code reads. Versions lists the agent's saved versions — see Versions.

New agent

New agent on the Agents page opens a window with two ways to start:

  • Describe it. Write what the agent should do, for whom and what it works with, attach files if they help, and press Build. The agent is created and the builder opens with your brief already sent. If some files fail to upload, the builder still gets the brief and you are told which part was left out. Start blank creates an empty agent instead.
  • Start from a template. Browse the gallery by category — each template's publisher chooses its category, so the filters are the ones the templates carry — open a card to see what it comes with — skills, MCP servers, workflows, knowledge collections, keys to fill in, a database, backend code and whether it has an app — and press Use template. See Templates.

Agent actions

The agent menu in the builder (its title and the agent card in the sidebar), and the one in the agent's details on the Agents page, hold the actions for one agent:

  • Rename, with an optional description.
  • Clone — a copy that belongs to you, with what Making a copy describes.
  • Pause agent / Resume agent — see Pausing an agent.
  • Publish as template… / Update template… — share the agent as a template. You pick what goes in: the core setup, skills and MCP servers, app customizations, backend code, environment key names (never their values), the database structure, and optionally the rows of chosen tables and knowledge collections. Rows can hold personal data, so pick reference tables only; collections you include become shared and visible to everyone. Updating overwrites the live template. Publishing a template is limited to platform administrators, on their own agents.
  • Delete.

In the Mac app the Agents page has no details panel: right-click an agent's row there to rename or delete it, and open the agent's builder for the rest.

The agent's chat

Everyone who talks to your agent — you, people you shared it with, visitors of its published page, the mobile apps — uses the same chat. Its look is the platform's; what it says is yours to set in the Chat settings panel:

  • Welcome title and Welcome text — the heading and the short line an empty chat opens with (for example Plan your trip and Ask about routes, prices and dates). Leave them empty to keep the standard greeting.
  • Suggestions — up to four ready-made first messages shown under the welcome. Clicking one puts its text into the message box, so the person can adjust it before sending; nothing is sent on its own.
  • Placeholder — the hint inside the empty message box.
  • Allow choosing the model in chat — lets people choose a platform model in the agent's chat. It applies to you too: while it is off, platform turns use the model set in the Model panel. When it is on, the list under it limits which platform models people may pick. An eligible owner using Donkit Desktop can still select a connected personal subscription.

The Builder fills in the welcome and the suggestions for a new agent, in the agent's language; edit them here, or ask the Builder chat to change them.

What "the config" is, from your side

You don't edit a config file. In practice, your agent is the sum of its system prompt, its model, the tools it's allowed to use, the knowledge and data it can draw on, the skills it has, and its interface (the surface you see running next to the chat). The Builder chat is simply a faster way to set all of these than doing it by hand — and everything the chat does, you can still fine-tune manually in the rail.