
Freebuff
github.com/codebuffai/freebuff- Category
- AI Tools
- Rank
- No. 2043Tools index
Previous survey · No. 2049 ·
- Type
- APP
- Use case
- Coding · Research & Education
- Interfaces
- CLI · Desktop · Web
- GitHub
- 13.1k stars
- Latest release
- v1.0.420-beta.185
- Date
About
A suite of five AI coding products: a desktop app for running parallel agents, a CLI, a web app builder, a cloud runner for GitHub repositories, and a research chat. Access to a curated catalog of models is funded by text ads rather than API keys.
What it does
Freebuff is the free, ad-supported edition of the Codebuff coding agent. You install it from npm, run it inside a project, and describe a change in plain English. A single agent loop then searches the code, edits files, runs terminal commands and checks its own work. The same engine sits behind a desktop app, a hosted app builder and a GitHub runner. Instead of an API key, you get a daily allowance of a house currency, Freebucks, spent on a rotating menu of models from Xiaomi, Zhipu, DeepSeek, Upstage, OpenAI and others. Text ads, personalised from your prompts, pay for it.
Why it's ranked here
The price is the headline: a working terminal coding agent with no subscription, no key and no setup. What backs it is a serious engine. The repository is the full Codebuff monorepo, with a reusable SDK, a published agent runtime and an end-to-end test suite covering startup, code edits, terminal commands and ad behaviour. The security writing is unusually candid: the docs state that downloaded agent code runs without a sandbox, then describe the publisher allowlist and per-folder consent prompt built to contain it. The cost is privacy, and the README states it plainly rather than hiding it.
What's good
The default agent is deliberately small: eight core tools, one loop, and a short system prompt that tells it to match project conventions, make the fewest changes, run the typecheck and tests, and avoid pushes or resets unless asked. Cloning a hostile repository is handled with care: project-level agents and tool-server configs do not run until you answer a yes or no prompt, and CI runs skip them unless you opt in. Edits cut off mid-generation are refused rather than patched, so a half-written file never lands on disk. Environment files are blocked from agent reads by default.
Tradeoffs
Free is paid for with your data. The README says prompts and messages may be analysed to personalise ads, some models may use submissions for training, and one beta model comes from an anonymous provider that retains prompts. Users outside supported regions or on a VPN get limited access and a smaller daily allowance. The model menu changes often: the README notes a retired DeepSeek model, and the CLI agent tied to it still references that model. Licensing is inconsistent, with the root manifest declaring Apache 2.0 while the Freebuff and SDK readmes say MIT. Agent folders you approve once stay trusted even if their contents later change.
How to use it well
Use it for everyday edits on personal or open-source code where cost matters more than confidentiality: bug fixes, small features, refactors and test runs from the terminal. Pick an unmetered model for long sessions so they do not eat your allowance. Keep proprietary or regulated code out, since prompts feed ad personalisation. Answer no to the trust prompt in any repository you did not write. Developers who want to build their own agents should look at the Codebuff SDK, which needs a paid Codebuff key. Freebuff does not cover teams needing a no-retention guarantee.
Technical notes+
Root package.json declares Apache-2.0, private, Bun 1.3.14, with workspaces for agents, cli, common, evals, freebuff, sdk and packages/agent-runtime, code-map and llm-providers; freebuff/README.md and sdk/README.md both state MIT. agents/base3.ts defines createBase3 with eight toolNames (read_files, str_replace, write_file, run_terminal_command, code_search, glob, list_directory, write_todos) and provider data_collection 'deny'; createBase3CliRoot appends web_search, read_url, ask_user, suggest_followups, gravity_index, render_ui and skill, and the comment explains there is no instructionsPrompt so the prompt cache stays warm. agents/base3-free-deepseek.ts imports FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID while README.md lists DeepSeek V4 Pro as retired. docs/agents-and-tools.md documents that registry handleSteps strings are eval'd unsandboxed, gated by a trusted-publisher list (CODEBUFF_TRUSTED_AGENT_PUBLISHERS), a per-directory consent store with mode 0600, and a truncated-write guard refusing a write under 60% of an existing 30+ line file with unbalanced brackets. agents/context-pruner.ts budgets 20,000 tokens for assistant and tool content and 50,000 for user content at 3 chars per token. cli/src/index.tsx loads the auth token from the OS keychain before parsing args. freebuff/package.json defines e2e suites including ads-behavior and code-edit.
Observed
- License
- Apache-2.0 in the root package.json; MIT stated in freebuff/README.md and sdk/README.md
- Language and runtime
- TypeScript monorepo, Bun runtime and package manager
- Install
- npm install -g freebuff; source build with Bun from a clone of the repository
- Interfaces
- Terminal CLI (OpenTUI + React), JS/TS SDK published as @codebuff/sdk, desktop, web and cloud products
- Funding model
- Text ads; README states prompts and messages may be analysed to personalise ads
- Access tiers
- Full access in supported regions; limited model set and smaller daily allowance outside them or over VPN
- Agent code isolation
- Registry agent step functions run unsandboxed; gated by a trusted-publisher allowlist and per-directory consent prompt
- Tests
- End-to-end suites declared in freebuff/package.json, including startup, code-edit, terminal-command and ads-behavior
Read from README.md, package.json, freebuff/README.md, freebuff/package.json, docs/agents-and-tools.md, agents/base3.ts, agents/base3-free-deepseek.ts, agents/context-pruner.ts, agents/constants.ts, sdk/README.md, sdk/src/index.ts, cli/src/index.tsx, SECURITY.md, AGENTS.md.
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.