CLI agents
Use an agent you already run from the terminal inside your writing workspace.
Nội dung này hiện chưa có sẵn bằng ngôn ngữ của bạn.
Bring your installed agent into Oleafly. The CLI Agent mode uses that agent’s runtime and sign-in, so you can work beside the manuscript with the tools you already use.
Connect an agent
Section titled “Connect an agent”- Open Settings → AI Assistant → CLI agents and choose Check installed agents.
- Expand your agent. The Program row shows the program Oleafly found and where it is.
- Choose Test. If the card asks for a bridge, choose Install bridge first.
- Sign in through the agent’s own command line tool. The card shows its Sign-in command, and with a project open, Open sign-in terminal gives you a terminal to run it in.
- Open the assistant, switch to CLI Agent, and pick the agent.
- Start with a small task, such as explaining the project structure or revising one paragraph.
If you switch to an agent that isn’t ready yet, the assistant shows a setup card that says what’s missing. Its Set up button, such as Set up Claude Code, opens that agent in Settings. When only the bridge is missing, Install bridge on the card installs it.
Authentication, model access, and billing belong to the selected agent. An account or subscription that supports the CLI can be used according to that provider’s terms; Oleafly does not add a separate AI subscription.
Choose the program
Section titled “Choose the program”Under the Program label, the card shows the file Oleafly starts and where it came from: Found automatically, Chosen by you, or Installed by Oleafly. For some agents this row sits inside the ACP bridge details instead.
To use another copy, choose Choose… and pick the file. If Oleafly found nothing, the row says Not found on this computer and gives you a field for the location. Type or paste the path, then press Enter or choose Test. A file you pick or type is saved only after it passes the test. Use automatic detection drops your choice.
Oleafly looks in the folders where installers put agents:
- On Windows it finds agents installed with npm, pnpm, Scoop, winget and similar tools, on any drive. It reads the search path fresh each time, so Check installed agents also finds an agent you installed while Oleafly was open.
- On macOS it finds agents in folders that usually only your shell knows about, such as
~/.local/binor an nvm folder, even when you open Oleafly from the Dock or Finder.
When Oleafly finds a file with the right name that it can’t use, setup tells you what it found and why. A common case on Windows is a PowerShell script where Oleafly needs the .cmd or .exe version. Files Oleafly skipped lists each such file with its reason, for example a folder, a desktop app, a shell or runtime such as node, or a copy on a network location.
Paths in Settings stay blurred until you hover over them or tab to them, so a screenshot of Settings doesn’t show where your files live.
Test an agent
Section titled “Test an agent”Test starts the agent in an empty folder that belongs to Oleafly, away from your project, and waits for it to answer. It saves nothing. When it passes, you see the agent’s name and version. When it fails, you get one sentence saying why and one thing to try:
| What the test reports | What to try |
|---|---|
| No file there, or a folder, desktop app, or file Oleafly can’t start | Choose… and pick the program itself |
| The bridge isn’t installed yet | Install bridge |
| The agent needs Node.js, or a newer version of it | Get Node.js |
| The agent started but didn’t answer, stopped early, or sent a reply Oleafly couldn’t read | Copy details, or Open terminal to run it yourself |
If a file you chose fails for any reason other than the file itself, Use anyway saves it regardless.
Agent catalog
Section titled “Agent catalog”| Agent | CLI command |
|---|---|
| Claude Code | claude |
| Codex CLI | codex |
| Gemini CLI | gemini |
| Pi | pi |
| OpenCode | opencode |
| OpenClaw | openclaw |
| Cline | cline |
| Hermes Agent | hermes |
| CodeBuddy | codebuddy |
| Kimi Code | kimi |
| Grok Build | grok |
| Cursor | agent |
| DeepSeek Harness | dsh |
| Qoder | qodercli |
The catalog identifies supported launch paths. Available models, attachments, session resume, and usage reporting depend on what the agent exposes. Setup reports missing requirements before a session starts.
Pi runs through the pi-acp bridge, version 0.0.34, and Oleafly points the bridge at your Pi program. On Windows it handles project folders with spaces in their names.
Work beside the manuscript
Section titled “Work beside the manuscript”Give the agent a concrete request and follow its replies, tool activity, and permission prompts in the assistant panel. Open the resulting source and compile it in Oleafly.
Read the introduction and methods. List any terms that are introduced inconsistently, with file locations.
For a separate copy and final review, choose an available agent in Research tasks. That page lists the platform requirements for isolated work.
Research starting points
Section titled “Research starting points”A new conversation opens on the agent’s home page, which offers research starting points. Picking one fills in the message but doesn’t send it, so you can read and adjust it first. If no conversation is open yet, it starts one.
| Starting point | What it asks for |
|---|---|
| Literature sweep | An annotated reading list for the project’s research question, with the oleafly-literature-sweep skill |
| Related work | A related work section drafted from the reading list and the bibliography, with oleafly-related-work |
| Citation audit | A check of the manuscript’s claims against their cited sources, with oleafly-verify-claims |
| Manuscript review | A review of the full manuscript, with oleafly-review-manuscript |
| Reproducibility audit | What each result is missing to be reproduced: code, data, software versions, random seeds and build instructions |
| Figure audit | A check of every figure and table: captions, labels, cross-references, axis labels and units, readability, and colors that hold up in grayscale |
| Git review | A summary of uncommitted changes and recent commits that flags edits that look unintended. Shown only in a Git repository |
| Reference cleanup | Opens Clean the reference library, which standardizes citation keys and removes duplicates after you preview the changes |
The audits and the manuscript review ask the agent to leave manuscript files alone and write its findings to a file in research/, such as research/figure-audit.md.
A skill you pick, from a starting point or from the skill list on the home page, shows as a chip above the message. Oleafly sends the skill’s instructions along with your message, so agents that don’t get Oleafly’s tools, such as Pi, can still follow it. Click the chip to remove it. If the skill is turned off for the project, the starting point fills in the message without it.
Permission requests
Section titled “Permission requests”When the agent asks before it acts, a card says the agent needs permission and offers the agent’s own choices. If the request would change files, the card lists them. Show change opens the diff for a file so you can read it before you answer. A change too large to show says so, and you can review it after the turn. Dismiss cancels the request without allowing it. A request you leave for two minutes expires, and the agent has to ask again.
Review and undo changes
Section titled “Review and undo changes”After a turn that changed files, a line under the reply says how many. Review shows each change, and Undo and Redo work file by file or for the whole turn. Review and undo a turn explains how this works and what it leaves alone.
Conversations and export
Section titled “Conversations and export”The assistant header has New conversation, Saved conversations, and Agent setup. Each conversation records the Git commit the project was at when it started.
To keep a copy outside Oleafly, open Saved conversations and choose Export conversation…, then pick Markdown or JSON in the save dialog. The Markdown file begins with the project, agent, model, the commit the conversation started at (noting uncommitted changes), and when it started and was last updated. The messages, skills, tool calls and changed files follow. JSON holds the full record.
Custom agents
Section titled “Custom agents”Choose Find more agents to search the ACP registry for agents published by other teams, or Add a custom agent to paste a definition yourself. Review its executable or package, arguments, and connection requirements before running it. Registering a definition does not install or start the agent.
Use an external client instead
Section titled “Use an external client instead”If you prefer to stay in your agent’s own interface, enable Oleafly’s MCP server. It exposes project tools to an external client while Oleafly remains open.