For developers: your AI tools, more compute, fewer tokens

BitPull plugs into the tools you already use through MCP (npx -y bitpull mcp) and adds what they lack: shared code to reuse, open models on your own hardware, sandboxed jobs on other machines and a browser preview your agent can drive.

What you can do

Reuse instead of regenerating

search_repos, read_manifest and fetch_files find working code shared by other developers and selected permissive or GPL GitHub repositories, and fetch only the files needed.

Offload to open models

ask_open_model and use_open_model send simple tasks to open models on your machines (free) or the network (credits). bitpull env exposes an OpenAI-compatible endpoint for other tools.

Pool your GPUs

Groups split one large model across several of your computers, or your friends', so you can run models none of them fits alone.

Sandboxed jobs

run_job runs tests, builds and scripts in containers without internet or host files, on your machines first and then on the network, in parallel.

A preview your agent can drive

preview_open, preview_look, preview_click and preview_type let the agent read, click and type in the localhost site it is building, with console errors and failed requests.

Editor, terminal and CLI

Projects is an editor with an agent, Git, debugging and Open VSX themes; bitpull code and bitpull chat bring the same agent and models to your terminal.

How to start

  1. Install: npx -y bitpull, or the desktop app from the download page.
  2. Add the MCP server to your client: { "command": "npx", "args": ["-y", "bitpull", "mcp"] }.
  3. Log in with bitpull_login: you approve a code on the website, the AI never sees a password.
  4. Read llms.txt and install.json for the full tool list, requirements per OS and checksums.

Frequently asked questions

How does BitPull help save AI tokens?

Writing code from scratch makes an AI produce, and often correct, many thousands of tokens. With BitPull it first searches the shared repositories (168,381 right now), reads each one's short description and fetches only the files it needs, which usually costs far fewer tokens. Simple tasks can also go to open models on the network, paid in credits instead of your own AI's tokens. How much you save depends on the task.

Can my AI (Claude Code, Cursor, the agent in BitPull) see and test the site it is building?

Yes, from version 0.1.41. The preview shows sites served on your computer (localhost). The AI reads the page as text with numbered buttons, links and fields, gets its console errors and failed requests, clicks, types, scrolls and takes screenshots; you see each click happen in the preview. The agent in Projects and `bitpull code` use it directly; Claude Code, Cursor and other MCP clients use the tools preview_open, preview_look, preview_click, preview_type, preview_scroll and preview_screenshot while the BitPull app is open. With a model of the community, each click or text waits for your approval. Only local sites can be previewed, never sites on the internet.

Who else it is for