Work with an assistant in Oleafly

A useful request gives the assistant a result to check. Try “Fix the undefined citations in this project, compile it, and report which keys are still missing.” Oleafly gives the assistant access to project files, bibliography tools, compiler logs, and PDF inspection, so the work can continue beyond a suggested snippet.

Connect the assistant you use

Open Settings → AI Assistant. In Providers and keys, configure a hosted provider, a custom OpenAI-compatible endpoint, or Ollama. For a local model, start Ollama, download the model you want, and connect its local server. Once the model and project dependencies are downloaded, you can use this setup offline.

You can also choose CLI agents to configure an installed agent such as Claude Code or Codex. These use their own account and authentication. Return to the assistant and choose Oleafly Agent or CLI Agent, then select the runtime and model for the conversation. See provider setup, CLI agents, and Ollama for the exact steps.

Set the approval mode

The approval selector sits below the chat message field and is saved per project.

Mode	When to use it
Ask for approval	Let read-classified tools run and review other actions first.
Approve for me	Allow read and network tools; file changes and shell commands still ask first.
Full access	Let tool actions run without approval prompts.
Custom	Apply your saved allow/deny rules, with the standard risk policy for unmatched tools.

Choose Ask for approval for a first editing session. Read the tool operation on each approval card before accepting it. For an edit to a selected passage, inline editing gives you a focused replacement to inspect. Approval controls vary with the selected runtime; a CLI agent also has its own configuration.

Instructions and approvals explains project rules, approvals.toml, and the other configuration files. Instructions guide the work; the approval mode controls what the assistant may run.

Give it enough context

Keep the relevant chapter, bibliography, figures, and data in the project. Use Add context to point the assistant at the material for the request. For notation or style preferences, save custom instructions under Settings → AI Assistant → Instructions:

Preserve the notation macros in the preamble. Keep citation keys unchanged. After editing LaTeX, compile and check the affected PDF pages. Report anything you could not verify.

A skill supplies a reusable workflow for literature searches, figures, reviews, and other recurring work. Choose one from the assistant’s skill cards or slash menu, then add the question and output you need.

Make the result checkable

For a build repair, name the failing file and ask for a successful compile. For a figure, specify the data, dimensions, labels, and destination. For a literature search, ask for titles, source links, and a short explanation of why each paper belongs in the review.

A request for a specific section could be:

Revise the first two paragraphs of the introduction for clarity. Preserve the claims, notation, and citation keys. Compile the project and check that the revised text appears in the PDF. List your changes.

Read the result against the request. A successful build confirms that the document compiles; it does not establish that a scientific claim or citation is correct.

Give longer jobs a Research Task

Open Research workspace → Tasks → New task. Give it a title, write the instructions, and choose an available Agent and model. Create the task, then start it. The agent works in an isolated copy while you continue writing.

For example, ask it to produce reading-list.md with five relevant papers, verified identifiers, and a paragraph on each paper’s relevance. Follow Activity, open the file previews in Review, and apply the selected results when ready. This is useful for a literature sweep, an independent review, or a figure experiment you want to inspect before bringing it into the manuscript.

The Research Tasks guide covers starting, stopping, reviewing, applying, and removing tasks.

Keep useful versions

Saved checkpoints and Git have different jobs. After a successful compile, Oleafly can save a checkpoint when the source has changed and the build is eligible. Open Versioning → Saved Checkpoints to inspect or restore one. Use Source Control to make deliberate Git commits and share them with collaborators.

A checkpoint is not automatically created before every AI edit. Before a large revision, compile the current draft and confirm a checkpoint exists, or make a Git commit. See checkpoints and Git history.