Skip to main content
Once you write and approve a to-do, an agent running on your computer changes the code and leaves the result as a PR or a merge. This page covers the first setup and how one to-do is handled from start to finish.

Get started

Create an agent database

In the sidebar’s Docs tab, hover over the space name, press + and choose New database. If the space is empty, press New database right under the space. Only space admins (the team owner or a space owner) can create an agent database.
  1. Choose Agent database.
  2. If the space has no GitHub repository yet, Connect GitHub and pick the repositories. If it already has repositories, this step is skipped and all of them are assigned to the agent.
  3. Under How far should the agent go?, choose an approval policy and press Create.
You get a database named Agent to-dos with three sample to-dos assigned to you, an Agent guide page, and three to-do templates. The first sample (high priority) is already Approved; the other two are Awaiting approval.

Turn on your agent

Each person who uses the agent turns on their own agent on their own computer, once. At the top of the database are the Agent guide button, the settings button, and My agents. My agents at the top of an agent database, with the switch for this device turned off
  1. Turn on the switch in My agents (Turn on agent).
  2. The first time, if the repositories aren’t on this computer yet, choose where to download them (Where should repositories be downloaded? › Choose folder…). You only choose once, and you can change it in Settings.
  3. When asked whether to trust the folder, press Trust and start. If you already trusted the folder in Claude Code in the terminal, you’re asked only about the agent guide’s check commands.
After that, downloading the repositories and checking how to verify each one run by themselves. This can take a few minutes. If Claude Code or GitHub isn’t signed in, a notice and a button to fix it appear in the same place (for example Connect Claude and turn on).

Hand a to-do to your agent

To try it out, watch the sample that’s already Approved: your agent picks it up and leaves the result as a PR. Samples always stay as PRs — a person decides whether to merge them.

Write a to-do

Create a new to-do and describe what needs to change in the body. Free text is fine. To cover everything, you can pick a template — Idea, Bug report or Feature request. Bug report asks for the symptom, steps to reproduce, expected result and screenshots. You can also ask in a Bottari session, for example “Write a to-do for this bug”. Claude writes the to-do from the conversation with the Bottari tools.

Set the assignee

Set the to-do’s Assignee. The agent on the assignee’s computer handles it.

Approve the to-do

New to-dos start as Awaiting approval. Check the content and change Stage to Approved. Then the agent can pick it up. The approval counts only if it comes from the assignee or someone with full access to the database.

Watch the agent work

Your agent picks up approved to-dos one at a time, in the order set in the agent settings: Priority (High → Low), Oldest first or Due date first. To-dos already in progress finish first. Your agent’s session (“🤖 device name · database name”) appears in the sidebar, with one session per to-do under it. Open a to-do session to follow the agent’s planning, changes and checks. These sessions stay on the device the agent runs on, so you can’t hand them to another person or device. Click the person at the top of the session to see the Assignee card. To start an urgent approved to-do right away, write “Do T-12 first” in your agent’s session.

Answer questions

When the agent needs a decision, it asks in the session and the to-do session in the sidebar shows Asking. Pick an option above the input box or type your own answer (Or type your own answer), and the work continues. Open questions are also collected in the database’s Needs answer view.

Check the result

The agent runs each repository’s checks (such as tests). Under PR only for risky changes or Always merge, an ordinary change that passes is opened as a PR, merged, and shown in the session; under Always PR only, it stays a PR. Click the PR to see the changes on GitHub. A record of the cause, fix, checks, answers and PR is also added at the end of the to-do page. If the change is risky, or the agent guide has no checks for the repository, a review card appears in Needs answer before merging. Merge runs the pre-merge checks and then merges. Leave the PR only keeps the to-do In review; it changes to Done once the PR is merged on GitHub.

Agent settings

Open them with the settings button next to Agent guide at the top of the database. Settings apply to every agent in this database. Only space admins and people with full access to the database can change them.

Approval policy

Under every policy — Always merge included — changes to dependencies, migrations, CI or config files, and changes that are hard to undo or change direction in a big way, wait for a person’s review.

Agent guide

The agent guide is the page of rules the agent follows for each repository. Open it with the Agent guide button at the top of the database. Its first draft is written automatically by the computer of the person who created the database. If that didn’t happen, a computer of a space admin or someone with full access to the database writes it once that computer’s agent is running with Claude Code ready and the repositories downloaded, or such a person presses Write a draft on the guide. Until then, others see The draft shows here once an admin writes it. You can edit the guide like any page. Once the draft is ready, space admins and people with full access to the database can also press Update the guide to have the agent revise it.

Limits

  • The agent doesn’t pick up to-dos that haven’t been approved.
  • Editing the title or body after approval removes the approval. If the agent hasn’t started, the to-do goes back to Awaiting approval and must be approved again.
  • The agent doesn’t work while your computer is offline. If the computer is offline for over an hour, the agent stops; bring the computer back online to continue.
For further questions, contact support@bottari.ai. Bottari manual 13 pages (#940)