Free models, paid for another way
Freebuff packages five related services: a terminal coding agent, a beta desktop app for parallel local agents, a browser app builder, a cloud agent for GitHub repositories, and a general research chat. Users do not need a subscription, credits, or an API key for the included model catalog. Text advertising funds that access.
That distinction matters. The source is public, but the included inference and hosted environments belong to an operating service that can change capacity, eligibility, and models. Freebuff is attractive when the alternative is no coding agent at all. It is less attractive when predictable throughput or a fixed provider contract matters.
The privacy exchange is unusually explicit
Freebuff says prompts, messages, code, files, and repository data are processed to provide the service. It may analyze prompts and messages, including pasted material, to personalize ads using its systems and service providers. Separate uploads and connected repositories are not sent to advertising providers. Depending on the disclosed model or feature, submissions may also be retained to develop, train, test, evaluate, fine-tune, or improve models and products.
This is clearer disclosure than a generic link to a privacy policy, and the application shows applicable session limits and model-specific data-use notices before starting. It is still a poor fit for proprietary source, credentials, regulated records, unreleased financial information, or client material whose processing terms you cannot negotiate. Redacting a prompt does not help if the agent needs to read the sensitive repository to work.
For public hobby code, coursework permitted by policy, or disposable experiments, the bargain may be reasonable. For company work, “no API key” should not override data classification. A paid bring-your-own-key tool can be cheaper than handling one confidentiality incident.
The repository is not the complete service
The contributing guide calls this repository a public mirror of the Freebuff and Codebuff source tree. A private repository is the source of truth. Accepted community changes are ported into that private repository and later exported back to GitHub. Public contribution scope covers the CLI, SDK, common packages, agents, runtime, code map, model providers, selected desktop code, scripts, and documentation.
Backend, database, billing, deployment, secret management, and the private web application are specifically excluded. Apache 2.0 applies to the published code, but cloning this repository cannot reproduce the advertised cloud and web products. It also means outside contributors do not work directly in the authoritative tree, making review history and merge timing less direct than in a normal open-source project.
The README says local service development requires Docker and .env.local, then links docs/development.md. That path returned 404 through GitHub's contents endpoint during review. The contributing guide offers Bun install and build commands, but not a complete substitute for the missing environment guide. End users avoid this complexity because the npm CLI and downloadable desktop packages use Freebuff's hosted service.
The agent design has real substance
Freebuff does not route every job through one giant prompt. File-finding agents map a codebase, implementation agents modify it, reviewers inspect results, and research agents can use documentation and a browser. Desktop workspaces isolate parallel local tasks, while web and cloud offer sandboxes, terminals, previews, and deployment workflows. The underlying Codebuff framework also exposes an SDK for custom agents.
MCP support is active, but current fixes reveal important edges. Pull request 890 says dead MCP transports were never removed from a client registry or tool cache, allowing stale connections to accumulate. It also says HTTP and SSE configuration hashes omitted headers, so two configurations with different authorization tokens could reuse one cached client. Pull request 921 fixes nested JSON Schema properties disappearing before MCP tools reached the model. These have clear patches and tests, yet operators should verify which fixes have reached the distributed binaries.
Issue 957 reports project-local MCP servers not loading when the CLI starts above a project and the project picker changes directories. A same-day pull request reloads the registry after selection and lists 41 related tests.
Desktop and long sessions need caution
Several current reports concern the bundled Bun and Electron environment. Issue 983 reproduces an illegal-instruction crash before the orchestrator starts on a Windows 11 machine with an Intel processor lacking AVX2. Issue 960 reports the same class of failure on an older Intel Mac with no baseline build. Other Windows reports describe missing Electron payload files, access violations, and a Tiny11 desktop crash even though the CLI works. Modern machines may be unaffected, but Freebuff should publish its CPU baseline prominently.
Issue 968 reports chat-messages.json reaching 275 MB after half a day, with every message, tool call, and agent output appended without a limit. Parsing that file allegedly expands memory use to between 1.5 and 2.8 GB. One user's measurement is not a benchmark, but an unbounded single-file history is a credible design concern. Back up useful sessions and watch disk and memory during extended agent runs.
The repository was pushed on August 12, 2026. Its 237 open issues and pull requests include same-day fixes, detailed reproductions, and active product questions. GitHub returned no latest release, so health should be judged from that fresh development rather than release tags. Distribution appears to happen through npm and the project's own download channels.
A useful trial, not a neutral free lunch
Freebuff lowers the financial barrier to agentic coding and gives users several serious model choices. The CLI setup is simpler than most self-hosted agents, and the public runtime contains meaningful code rather than a marketing stub. Its maintainers respond quickly to concrete reports.
The recommendation changes with the data. Use it on code you could safely send under the displayed terms, confirm limits before a long session, and prefer the CLI if desktop compatibility fails. Do not choose it because the word “free” makes the privacy, service dependency, and closed backend disappear. Those are the price, and competent teams should evaluate them explicitly.