LikeClaw: From Chat Dashboard To Mobile AI Workspace

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

LikeClaw is Fulldive’s AI agent app — a mobile AI workspace built in Flutter and shipped for Android and iOS. This post walks through the seven-week window between 2026-02-11 and 2026-03-30 during which 141 commits turned an internal chat dashboard into a branded LikeClaw release, with files, skills, tool calls, image generation, billing, an inbox, and an iOS TestFlight pipeline. We keep the scope narrow: what the git history actually shows, what the app can do today, and what it cannot.

First commit: a chat-based dashboard, not a chatbot

The starting point, on 2026-02-11, was a Flutter project labelled as a chat-based AI dashboard (9632105). The very first commits already committed to three decisions that shaped the rest of the work:

Flutter as the mobile runtime. One codebase targeting Android and iOS. The official Flutter docs explain this shared-codebase model; see flutter.dev architectural overview.
Token-preserving login (5682625), so the app could stay signed in across restarts rather than re-authenticating on every launch.
Files browser and skills screen with VFS integration (2970709). Before rich chat was polished, LikeClaw already treated virtual files and named skills as first-class surfaces.

That order matters. LikeClaw did not start as a chat product and later bolt on files. Chat, files, and skills landed in the same initial week, which is why the current UI treats them as peers rather than as a main screen and two tabs.

Making the chat layer actually usable

Chat is where most AI apps look alike and where the small things decide whether the app is usable on a phone. Between 2026-02-11 and 2026-02-15 LikeClaw landed the mobile plumbing that makes chat feel like chat rather than a web form:

c3001bf — rich chat messages, file renderers, and network error handling.
09ee58f — background task polling and message refresh, so results come in when the model or a tool takes a while.
59125d2 — pull-to-reload, the one gesture every Android user expects from a list.
882aa78 — tool-call visibility and generated image display, so the user sees what the agent did, not just what it said.

The 882aa78 commit is worth calling out. A lot of mobile AI apps hide tool calls behind a spinner. LikeClaw decided to show them. When the agent calls a skill or generates an image, the chat timeline reflects that. This is also the foundation for the limits we discuss in what a mobile AI agent can responsibly automate: visible tool calls are a prerequisite for informed oversight.

Model selection, billing, and the inbox

By mid-February the product had grown past “chat + files” into a workspace with real-world constraints:

7cf67c3 added model selection, billing, and credit-error handling. Users can choose a model, and the UI has to explain what happens when credits run out.
bf94935 introduced an inbox. Background jobs and notifications need somewhere to go when the user is not in the conversation that triggered them.
8a146d7 localized the billing flow, because billing copy is where unclear wording creates the most support load.
81420bd redesigned the file-changes card. When an agent edits files, the diff preview needs to read well on a phone screen.

The engineering detail behind billing deserves its own post — see billing and StoreKit 2 in a mobile AI app — but the headline is that LikeClaw shipped real in-app billing rather than deferring it to the web.

Rebrand and mobile UX overhaul

The app started life under a working name. The rebrand to LikeClaw (bfb1dda) added the current logo, icons, and theming. Close behind it was a mobile UX overhaul (fb41c67): navigation, spacing, typography, and interaction patterns tightened for phone ergonomics rather than for a tablet-sized chat window.

Two further commits round out this phase:

a8aef32 — a Connections screen, the surface through which external accounts and integrations are authorized and managed.
d3457d6 — AI consent simplification and account deletion. Consent and deletion are covered in depth in AI consent and account deletion in LikeClaw; the important point here is that they landed as part of the core product, not as afterthoughts.

iOS delivery and purchase verification

A mobile AI workspace that only ships on Android is not really a mobile AI workspace. Two commits close the loop on iOS:

eacfed4 — iOS build with TestFlight and App Store deploy workflows. See shipping LikeClaw to TestFlight and the App Store for the pipeline details. Apple’s own description of the TestFlight distribution path is at Apple’s TestFlight documentation.
0b9f7db — StoreKit 2 signed transaction handling, which is how the app validates purchases against Apple’s receipts. Apple documents this in the StoreKit reference.

A multi-brand support commit (72f4aa7) also landed in this window, anticipating that LikeClaw’s plumbing could serve more than one branded build if needed.

Where LikeClaw sits in the Fulldive ecosystem

LikeClaw is a newer arrival among Fulldive’s Android-first apps. The rest of the catalog — Wize AdBlock VPN, WizeUp (formerly FD Browser), Fulldive VR, One Emulator for Game Consoles — is maintained across long-running repositories with years of history. LikeClaw is the opposite: a fresh Flutter codebase with seven intense weeks of work visible end-to-end. The longer arc of the ecosystem is covered in the Fulldive company story and in what is Fulldive. For the list of apps including LikeClaw, see the Fulldive apps page.

Because LikeClaw does not yet have a dedicated project page on fulldive.com, the best entry point is the projects list and the FAQ for ecosystem-level questions.

Limits and safety

A mobile AI agent app is not a magic task runner. A few honest framings, grounded in what the repository shows:

LikeClaw exposes tool calls and generated images in the chat timeline (882aa78), but visible does not mean verified. A user still has to read what the agent did.
Billing is handled through platform in-app purchase flows and StoreKit 2 signed transactions (7cf67c3, 0b9f7db). That protects payment integrity. It does not certify that any particular model output is correct.
Consent and deletion (d3457d6) exist, but consent is an ongoing conversation, not a one-time checkbox. See the dedicated consent and deletion post.
This post does not claim LikeClaw can complete arbitrary tasks safely or accurately. We describe what the repository shows. Treat the agent as an assistant that needs oversight, especially for anything with financial, legal, or security consequences.

What is next

The most recent commits point at three directions: deeper skill and VFS integration (see skills and virtual files in LikeClaw), continued iOS delivery work on top of the StoreKit 2 foundation, and clearer consent surfaces as the agent’s reach grows. If you want to follow along, the Fulldive apps page is where LikeClaw will show up alongside the rest of the Android-first catalog, and support@fulldive.com is the fastest way to reach the team.

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