Englishะ ัƒััะบะธะน Map

cursor-bridge

Toolbox toolboxrustclaude-codellmapi-gatewaywatchlistRustMIT โ†ณ show in map Markdown
title
cursor-bridge
type
toolbox
summary
Rust binary that runs Claude Code on a Cursor subscription by proxying Anthropic calls to Cursor
tags
rust, claude-code, llm, api-gateway, watchlist
repo
https://github.com/hkc5/cursor-bridge
language
Rust
license
MIT
created
2026-07-29
updated
2026-07-29

cursor-bridge is a single Rust binary that lets you run Claude Code without an Anthropic subscription or API credits, by pointing it at Cursor's backend and paying with a Cursor subscription you already have. You run cursor-bridge instead of claude, and the session looks the same.

How it works

The binary does four things and then gets out of the way. It starts a local HTTP proxy on a random port, reads your Cursor auth token (from the macOS keychain, or from CURSOR_TOKEN on Linux, which has no keychain fallback), spawns claude with environment variables pointing at that proxy, and cleans up when the session ends. The proxy translates Anthropic API calls into invocations of Cursor's agent CLI, which must already be installed and logged in via agent login.

cursor-bridge                         # interactive session
cursor-bridge "refactor this file"    # one-shot prompt
cursor-bridge -p "list files"         # pipe mode

Install is cargo install cursor-bridge or a release binary. macOS and Linux only.

The design argument is entirely about lifecycle. The README names three existing proxies in the same category (cursor-api-proxy, cursor-composer-in-claude, cursor-proxy) and their shared shape: Node.js servers you start, keep in your process list, assign a port to, and wire environment variables into before running claude separately. cursor-bridge collapses that into one process that lives and dies with the terminal โ€” a statically compiled binary of roughly 780 KB against a Node runtime's 60+ MB, a random port instead of a chosen one, and nothing left behind to stop.

That is a real difference in kind rather than degree, and it is the same difference claude-code-router does not make: ccr is the long-lived, configurable, many-provider version of this idea, with a routing table and per-provider transformers. cursor-bridge has no configuration at all, which is the point and also the limit. cc-mirror takes the third approach in the family, cloning the Claude Code install per provider. All three are instances of the llm-api-routing-layer shape with Claude Code as the fixed client.

What you actually get

The README's pitch is "Claude Code's agent capabilities without Anthropic billing," and it is precise about the mechanism while being quiet about one consequence. What Cursor includes with a subscription is its Auto model โ€” Cursor's own routing choice, not a model you pick. So the thing you get for free is the Claude Code harness (its tool loop, file editing, shell access, permission model), running against whatever Cursor's Auto tier decides to serve. It is not a route to Opus on someone else's dime. Anyone evaluating this should read the value proposition as "keep the client, change the model," not "same experience, lower price."

The terms-of-service question

The repo does not claim permission. Its Legal section is two sentences: "This project is not affiliated with Anthropic or Cursor/Anysphere. Use at your own risk."

The factual position is that a Cursor subscription buys access through Cursor's own clients, and this routes a different vendor's agent through the agent CLI to consume that quota. Nothing in the README cites a Cursor term permitting third-party clients, and "free and unlimited" is the README's characterization of Auto rather than a quoted entitlement. Whatever fair-use, rate, or client restrictions Cursor's terms carry apply here, and the account exposed to enforcement is the user's. Claude Code's own licensing sits on the other side of the same transaction. This is worth stating plainly rather than treating as an implementation detail, because the entire value proposition is subscription arbitrage โ€” the same subsidy structure ai-subsidy-economics describes, used from the consumer side.

Limitations

  • Linux requires CURSOR_TOKEN in the environment; there is no keychain integration.
  • No workspace sandboxing. The agent runs in the current directory with whatever that implies.
  • Single account, no rotation.
  • Cursor's Auto tier picks the model per request, and KV state does not move between models, so any reroute re-prefills the session โ€” the cache cost prompt-caching-in-agents works through.
  • Depends on three moving surfaces at once: the Anthropic API shape the proxy translates from, Cursor's agent CLI it translates to, and Cursor's willingness to have it happen. Any of the three can break it without warning.

That last point is why it belongs on the watchlist: the tool is a small, clean piece of work whose lifespan is set by a business decision at another company rather than by its own maintenance.

Repo: https://github.com/hkc5/cursor-bridge (MIT). Published on crates.io; the README carries a stars badge but no count, so adoption is unmeasured from the clip.

Linked from