Skip to content

Explainer

Skills And Virtual Files In LikeClaw: How The Agent Gets Work Done

An explainer on LikeClaw's skills screen, virtual file system, tool-call display, and Connections screen — the surfaces that turn chat into a mobile AI workspace.

Skills And Virtual Files In LikeClaw: How The Agent Gets Work Done preview

Written by the Fulldive product engineering team, based on direct inspection of the repositories listed below.

LikeClaw is an AI agent app built in Flutter. The chat timeline is the surface most users see, but it is not where the work gets done. The work happens in three other surfaces: a skills screen that catalogs what the agent can actually do, a virtual file system (VFS) that gives conversations durable files, and a Connections screen that scopes external integrations. This post explains what each surface is, when it lands in the repository, and how they fit together on a phone.

Why skills and files are peers of chat, not children

A common mistake in AI apps is to treat chat as the product and everything else as an afterthought. LikeClaw’s repository shows a different choice. On day one — 2026-02-11 — commit 2970709 added a files browser and a skills screen with VFS integration, at the same time as the chat dashboard itself (9632105). Files and skills were not grafted on later; they were peers of chat from the beginning.

That matters because it shapes how the app expects you to work. Chat is the conversation. Files are the durable artifacts. Skills are the capabilities. All three are equally first-class, and each has its own tab.

The skills screen

A skill is a defined capability the agent can invoke. Think of it as the difference between “the AI might do something” and “the AI can do these specific things.” Skills are named, enumerable, and visible.

Two reasons that is important:

  • Predictability. You can see the shape of what the agent is capable of before you ask it to do something.
  • Oversight. When the agent runs a skill, the chat shows it. Commit 882aa78 added tool-call visibility and generated image display, which means the chat timeline surfaces the skill invocation and its output, not only the natural-language response.

The tool-call visibility from 882aa78 is the foundation for our broader framing in what a mobile AI agent can responsibly automate. You cannot supervise what you cannot see.

The virtual file system

The VFS is LikeClaw’s idea of “files that belong to this workspace.” Unlike a traditional file manager, it is not a view onto the Android file system. It is a workspace-scoped store that the chat, the skills, and the agent all share.

A few commits tell the story of the files surface maturing:

  • 2970709 introduced the files browser alongside the skills screen at first launch.
  • c3001bf added rich chat messages with file renderers. When a file is referenced in chat, it displays with an appropriate renderer rather than a plain filename.
  • 81420bd redesigned the file-changes card. When the agent edits a file, the diff-like preview reads well on a phone screen.

The redesign in 81420bd is the small detail that makes file editing in a phone UI tolerable. Desktop diff viewers assume a wide screen and a mouse. A file-changes card on a phone needs to communicate “what changed” in a column that fits on a 400-pixel-wide canvas. That card is now the primary way users review agent edits before accepting them.

The Connections screen

Skills and files are internal surfaces. Connections is the boundary between LikeClaw and the rest of the internet. Commit a8aef32 added the Connections screen, which is where external accounts and integrations are authorized and managed.

The Connections model pushes toward scoped permissions rather than blanket access. For example, Gmail-based integrations can be authorized using OAuth scopes limited to the minimum required operation, as Google documents at Gmail OAuth scopes. The principle transfers: grant the narrowest scope that lets a skill do its job, and revoke scopes you no longer use.

We treat the Connections screen as the place where consent becomes concrete. Broader consent framing (AI consent simplification and account deletion in d3457d6) is covered in AI consent and account deletion in LikeClaw. Connections is where that framing meets real OAuth tokens.

How the three surfaces work together

A typical flow in LikeClaw, walked through the surfaces:

  1. You start a conversation in the chat. The chat layer is the Flutter UI; Flutter’s single-codebase model is described in the Flutter architectural overview.
  2. The agent decides it needs a skill. It invokes one. The chat timeline displays the tool call (882aa78), so you see what ran.
  3. The skill reads or writes a file in the VFS. The file renderer (c3001bf) and the file-changes card (81420bd) let you see the result in-line.
  4. If the skill needs external access — a Gmail read, a calendar event — the Connections screen (a8aef32) governs whether it can, using scopes you granted.
  5. Background polling (09ee58f) keeps the timeline current if the skill takes a while, and the inbox (bf94935) surfaces the completion if you moved on.

None of these steps are flashy. Individually they are unexciting. Together, they are what separates an AI agent app from a chat toy.

Mobile-specific choices worth noting

Two choices distinguish LikeClaw’s skills-and-files model from a desktop equivalent:

  • Everything is visible in the timeline. On desktop you can afford side panels. On a phone you cannot. So tool calls, generated images, and file changes share the same vertical scroll. That is why 882aa78 and 81420bd matter so much — they decide how much of the agent’s work fits in one column.
  • Pull-to-reload is a first-class gesture. Commit 59125d2 added pull-to-reload. It is the easiest way to force a refresh when the background poll is slow or the network flakes. Small gesture, large peace-of-mind.

Where this fits in the Fulldive ecosystem

LikeClaw shares the design principle that runs through the rest of the Fulldive catalog: prefer focused, composable surfaces over a single mega-screen. The ad blocker has a Privacy Pulse. The browser has a trends surface. The emulator has a games catalog. LikeClaw has skills and files. The broader pattern is discussed in the Fulldive company story and in what is Fulldive. LikeClaw does not yet have a dedicated project page; it is listed via the Fulldive apps page.

Limits and safety

Some candid limits on what skills and the VFS provide:

  • Tool-call visibility (882aa78) is about what ran, not whether it was correct. A visible tool call can still produce a wrong answer. Read the output; do not rubber-stamp it.
  • The VFS is not a security perimeter in the strong sense. Treat files in a workspace as readable by the agent and by anyone you grant access to the workspace.
  • Connections (a8aef32) let you authorize integrations, but a granted scope is a granted scope. If you give a skill write access to a mailbox, the agent can write to the mailbox. Prefer read-only scopes where possible.
  • We do not claim LikeClaw can complete arbitrary tasks safely or accurately. We describe what the repository supports.

Last updated: 2026-04-16. Commit hashes and version numbers are drawn from Fulldive repositories inspected on 2026-04-13.