Five products share one free account
Freebuff's public pitch spans five products, while the checkout we measured occupied 34.8 MB: a terminal agent, desktop application, browser builder, hosted GitHub worker, and research chat. The CLI has the smallest front door: install freebuff globally with npm, enter a project directory, and run the command. You then describe a change while the agent finds files, edits code, and runs project commands.
The other four products change where that work happens. Desktop can isolate concurrent agents in separate workspaces. Web supplies hosted app-building environments, while Cloud operates on connected GitHub repositories. Chat handles research. The common pitch is breadth: one account reaches five surfaces rather than forcing every job through one terminal session.
Underneath, Freebuff uses specialized agents for context gathering, implementation, browser research, and review. Desktop can also invoke locally installed Claude Code and Codex agents through the user's existing provider accounts. Those connections remain separate from Freebuff's included model catalog, so bringing one of them does not turn Freebuff into a self-contained local model runner.
Free access comes with ads, quotas, and data terms
The README's legacy limit is 6 sessions of 1 hour each per day, with up to 7 through rewards. Included models require no subscription or personal API key, and text ads pay for access. Model availability and limits also depend on region, product, tier, and current capacity. That is generous for an experiment, but it is not a fixed service entitlement.
The data terms should decide which repository you open first. Freebuff processes prompts, messages, agent traces, code, files, and repository data to provide the service. It may analyze prompts and messages for ad personalization. Separately uploaded files and connected repositories are not given to advertising providers, according to the README, although restricted partners may evaluate connected Cloud repositories.
Training use depends on the notice attached to a model or feature. When the notice permits it, Freebuff or its provider may retain submissions to develop, train, test, evaluate, fine-tune, or improve models and products. A blanket rule such as "our code never contributes to training" would therefore be inaccurate. Start with a disposable or non-sensitive repository, then check the notice for the exact model you select.
What happened when we ran it
Our sandbox installed commit 68940f7 in 81 seconds. Bun added 1,077 packages, and the installed workspace occupied 795 MB. The fresh Debian container had 3 CPUs, 8 GB of RAM, Node 22, no secrets, and no elevated privileges. Dependency installation completed successfully.
The checkout itself contained 1,777 files, about 344,713 lines of source, and used 34.8 MB. Our harness found no build script or target, so it skipped the build. It also found no test script or target, so it skipped tests. Neither skipped step counts as a pass or a failure.
The scan found two CI workflow files and Bun workspaces, but no Dockerfile or tests directory. These observations describe the measured checkout and generic root path. They do not prove that individual packages lack build commands or tests. We also did not measure task quality, agent speed, model accuracy, or the safety of generated changes.
The public repository is a mirror of a private source tree
Our measured 1,777-file checkout is a public mirror whose contribution guide calls the private repository the source of truth. Public work is accepted in areas such as the CLI, SDK, common code, agents, runtime, code map, model providers, and documentation. Backend, database, billing, deployment, secret management, internal packages, and the private web application are outside the public contribution boundary.
An accepted pull request is not merged into the public tree. A maintainer ports the patch into the private repository, then a later export returns it to GitHub. The guide also says an automated bot triages daily and may close an item after 28 inactive days, with a warning 7 days beforehand. That process can keep a large queue moving, but it weakens the usual public trail from patch to merge commit.
Contributor setup reflects the larger system. The measured commit pins Bun 1.3.14, while the README also asks for Docker and .env.local. Component commands build the SDK and Freebuff binary, and the testing guide describes package-level CI guards rather than one root test target. The three-command end-user quick start and the contributor workflow are different products in practice.
A September 28 push shows activity, while 307 items show support load
GitHub listed 12,888 stars and 307 open issues and pull requests combined on September 28, 2026. The repository was pushed that day, and recent public commits were labeled as snapshots from the private tree. Current issue activity includes a report that version 0.0.204 leaves roughly 10 MB in a temporary file on each Termux launch, plus a separate SIGILL report on an older non-AVX2 Intel CPU. Those are user reports, not findings from our sandbox.
There was no latest GitHub release returned by the API. That does not make a repository with same-day pushes abandoned. It does mean the Releases page cannot map an installed Freebuff package back to a public source snapshot for you. The project looks active, but its support queue and private-first release process demand more trust than the star count suggests.
Pick Freebuff for access, not infrastructure ownership
Freebuff's global-CLI quick start targets first-time trials; our contributor install took 81 seconds. Aider is a cleaner fit for a focused bring-your-own-model Git workflow. Codex or Claude Code fits teams already committed to one provider and one terminal agent.
The decision turns on control. Our install proved that the 1,777-file workspace can resolve its dependencies in a clean container, while the absent root build and test targets stopped the verification there. Use Freebuff for convenient access to five agent surfaces. Choose another tool when you need to own the backend, the model contract, and the complete release trail.