Quick start: first 15 minutes
The purpose of this text is to take you from the installed Pulse to the first commit made by the agent. Installation - in separate document.
1. Connect an agent (2 minutes)
Two pages in the “Agents and AI” group: Settings → Agents and Settings → Models. The settings are opened by the gear on the left.
- Agents - those that are already in your system: Claude Code, Codex, Kimi Code, OpenCode. Pulse runs them as separate processes. The program must be in
PATHand logged in (claude auth loginand analogues; the “Login” button will open Terminal with the desired command). - Models — for the built-in Pulse agent: your server on this computer or on the network, or GigaChat using a key. The names of the server programs are on the “Models” page itself.
A green pill “authorized” or “server responding” means the agent is ready. For the local model, click "Check": Pulse will ask for a list of models and make a test call to the tool - you can immediately see if the model is suitable for working with the code.
2. Describe the task (3 minutes)
Return to the board (⌘ 2). Below is the composer:
- Repository — select the folder with the git repository. Without a repository, the task will receive a draft folder with its git.
- Agent - who will work. This is also where you select your agent and swarm options.
- Prompt - what needs to be done. Write as you would assign a task to a person: what to change, what not to touch, how to check the result.
- Resources (paperclip) - files that must be attached to the task.
/- skills: type slash and a list of skills from.claude/skillsrepository and~/.claude. For the built-in agent, the menu shows only verifiedSKILL.md; The repository skill must be in the current commit in order to be included in the new working copy of the issue. Request/name your texttransmits the contents of the skill along with your text. For external agents, the menu also shows their slash commands.
Click "Run". Pulse will create a git-worktree on the branch pulse/<id>, will launch the agent and give the task a name according to the meaning of the prompt.
Button "Prepare" creates a task, but does not launch the agent - convenient when you need to first collect several tasks and then launch them at once.
3. Monitor your work (5 minutes)
The board shows for each task the branch, agent, turns, calls, tokens, cost and status.
- Click by line or
↵opens a task. There is no separate preview for the first click. - There are six tabs inside the task: Session (chat) Diff, Graph, Terminal, Files (editor and preview of working copy), Git (branches, history, pending changes).
- Graph (
Agent Graph) expands the work into a tree: task → sessions → turns → tool calls and subagents. Clicking on a node shows its contents on the right. - The bottom line is the state: what is happening now, notifications, and the graph also has a timeline: pull the slider, and the graph with chat will show the state at any moment in the past.
4. Respond to permissions
When an agent wants to write a file or execute a command, the permission card appears in the progress feed, rather than as a separate window on top of the entire screen. The “Reply” button in the task header scrolls to it. The card shows what the tool is, what command and what files. Options:
- Allow - once.
- Allow in session — do not ask about this tool again in this session; the team - about its program and subteam (
npm test). This rule is not offered to a compound command (chain, writing to files, code by the interpreter): above it the card writes what the command will write and what it will run. - Reject - the agent will be refused and try differently.
If you trust the task, enable auto mode ("Auto" chip in chat): the agent will stop asking about allowed actions. The policies continue to work: explicit Ask rules require an answer, and prohibited actions are not carried out. See Control.
5. Check the result and commit (5 minutes)
Tab Diff → Problem Result shows all changes from the original commit, including already committed work. Check out the specific version, merge the author's branches if necessary, and explicitly accept the result. The agent's response does not, by itself, complete the task.
In mode Working changes on the left is a tree of changed files, on the right is a patch. You can select multiple files (⌘-click), add to index (stage) or remove, look at pictures. The commit is written right there; You can ask the AI service for a message if it is configured.
Tab Terminal (⌘ T) opens a shell directly in the task's worktree - run tests, see git log, start the build.
When a task is not needed, delete it: Pulse will delete the worktree and branch. Archive (To the archive) hides the task from the board, but leaves the data. The history of time and expenses is retained even after the task is deleted. More details: collaboration, measurements.
What to try next
| Feature | Where | Document |
|---|---|---|
| Several agents for one task | "Team" chip in composer | Tasks |
| A swarm of roles: the architect distributes tasks | chip "Swarm" | Agent swarm |
| Your agent options with role and instructions | Settings → Agents → agent card | Agents and models |
| Prohibitions, budgets, sandbox, audit | Settings → Policies | Control |
| External tools via MCP | Settings → MCP | MCP |
| Ready-made skills and roles | Settings → Library | Library |