BitFun

English 中文

Key features

Why BitFun

Agentic Mini Apps. Most agents push every task through the same chat box. BitFun builds the task its own interface instead — a chart, a board, a form, a panel — and binds a conversation to that interface's live state. You ask about what is on screen rather than re-describing it. Community builds already range from market dashboards to domain-specific tools.

Browse the public Mini Apps gallery →

Self-hosted multi-device control. Account login, cross-device session and settings sync, and controlling one signed-in device from another all run through a relay you deploy. Nothing is brokered by a vendor's cloud — the distinction that decides whether this is allowed inside a company network at all. The relay is zero-knowledge: clients derive keys locally, and the server only ever holds Argon2id hashes and AES-GCM-wrapped material.

A runtime you can reshape. Four continuous tiers, from a single Markdown file to forking the runtime: custom Agents → MCP / Skills / Codex-compatible Hooks → Mini Apps → source-level changes. You extend BitFun using BitFun.

KV cache that actually hits. Agent cost is dominated not by generated tokens but by context re-sent every turn, and a single timestamp or reordered tool list invalidates the cache from that byte onward. Prompt assembly is byte-stable across turns: 98.67% average cache hit rate over a SWE-Bench-Pro run.

flashgrep. An agent re-searches the same repository dozens to hundreds of times per task, and cold traversal on every tool call can cost more than inference itself. A resident cross-turn index cuts search time by up to 94.6% on Chromium-scale trees — roughly 36x on average.

Install

Download a build — grab the latest installer from Releases, install it, configure your model, and you are ready to go.

Or run from source:

pnpm install
pnpm run desktop:dev

Prerequisites: Node.js 22.12+ (LTS recommended), pnpm 10.15.0 via Corepack, the Rust toolchain, and the Tauri prerequisites. More detail in CONTRIBUTING.md.

First run

1. Launch BitFun, click Open on the Welcome tab, and choose a project folder.
2. Open More options (…) → Settings → Models → Create First Configuration.
3. Choose a provider, enter its API key, select one or more models, and click Save. BitFun makes the first saved model primary and tests the connection automatically.
4. Return to the Session tab, type a concrete task, and press Enter or click Send.

What you can hand to BitFun

Two kinds of complex work: shipping code in real repositories, and turning source material into office deliverables. When a task needs the browser, desktop apps, the terminal, or a remote environment, BitFun can enter the real workspace.

Shared capabilities

• Interface layer: Mini Apps give a task its own UI, with the conversation bound to that UI's live state.
• Execution layer: the filesystem, terminals, Git, browser operation, desktop applications, Computer Use, and remote workspaces let the Agent reach the real environment when a task leaves the editor.
• Customization layer: MCP, Skills, Hooks, custom Agents, and source-level extension let BitFun keep growing around your tools, roles, and interfaces.

Agent core metrics

The data below evaluates BitFun's core Agent capabilities, all measured with Deepseek-V4-Pro.

[!NOTE] These are BitFun's initial evaluation results, with each case run once. Benchmarks fluctuate with task sampling, model versions, runtime environment, and single-run variance, so treat these as an initial sanity signal that the Agent is already reasonably capable — not as a fixed ranking claim or a final ceiling. Full benchmark details will follow.

1. Initial completion snapshot — The chart below compares the current single-run results on SWE-Bench-Pro (complex software engineering) and SWE-Bench-Verified (human-verified GitHub issue fixes).

Benchmark references: SWE-Bench-Pro / SWE-Bench-Verified

2. Token economy — Agent economy needs to be evaluated across end-to-end token consumption, execution time, and KV Cache reuse. From the same SWE-Bench-Pro round, BitFun's average KV Cache hit rate was 98.67%. The follow-up report will add broader cost and latency metrics.

3. Context retrieval at scale — Agent experience also depends on how quickly it retrieves context in very large projects. On tens-of-millions-line repositories such as Chromium, BitFun uses flashgrep to cut search time by up to 94.6%, averaging a 36.1x speedup.

Customize your BitFun

BitFun's extension paths progress continuously from light to deep customization:

You can use BitFun's Code Agent to extend BitFun itself.

Where this is going

• Lights-Out Factory (in progress): Design during the day, let tasks flow to the server and run through the night, then review the results in the morning.
• Infinite Radius (in progress): Extending from desktop and browser to mobile, wearables, and more devices, so work stays accessible and continuous.
• App Evolution: Build tailored workflows with custom Agents, MCP, Skills, Mini Apps, or source-level customization. Community contributors have already created specialized versions for short-form drama, media, and more.
• Better, Faster, Cheaper: Pursue greater efficiency, better results, and lower cost.
• Ultimate Desktop: Continuously refine an easier-to-use, more capable, and more beautiful desktop experience.

Community & contributing

Questions, ideas, and bug reports are all welcome in Discussions and Issues.

Stars, Issues, and PRs are welcome. We especially care about:

1. Code Agent, Deep Review, debugging, and long-task execution capabilities
2. Cowork, research, document, and desktop workflows
3. MCP, Skills, Mini App, LSP plugins, and new domain Agents
4. Runtime stability, performance, context efficiency, and verifiability

Please submit PRs directly to the main branch. For more details, see CONTRIBUTING.md.

Disclaimer

1. This project is spare-time exploration and research into next-generation human-machine collaboration, not a commercial profit-making project.
2. AI-assisted development is part of this project's workflow. Contributions are reviewed as code, and AI-assisted PRs should disclose their testing level; see CONTRIBUTING.md.
3. This project depends on and references many open-source projects. Thanks to all open-source authors. If your rights are affected, please contact us for remediation.