Case study · Autonomous AI · Backend & Infrastructure
Bridge
A small Go gateway that accepts Anthropic Messages API requests and forwards them to any OpenAI-compatible Chat Completions provider, so AI coding agents can run against other models.
- Status
- Early compatibility layer · v0.3.0
- Stack
- Go (stdlib), Server-Sent Events, Node.js launcher, GitHub Actions
- Links
- Repository
- Evidence
- Translates tool_use / tool_result blocks to and from tool_calls
- Re-emits upstream streams as Anthropic-style SSE events
- Launcher verifies release checksums; keys go to the OS keychain
What it does
Bridge translates Anthropic Messages requests — system prompts, tool_use and tool_result
blocks — into OpenAI Chat Completions requests, and maps responses back, including tool_calls
to tool_use stop reasons. Streaming responses are re-emitted as Anthropic-style SSE events
(content_block_start and friends).
Engineering notes
- Binds to
127.0.0.1by default. - A configurable first-byte timeout (
BRIDGE_FIRST_BYTE_TIMEOUT, default 60 s) has a regression test for upstreams that never send response headers. - The
npxlauncher downloads the platform binary from GitHub Releases, verifies its SHA-256 checksum, and stores the provider key in macOS Keychain or the Linux Secret Service. - Tag-triggered releases cross-compile macOS, Linux and Windows binaries with a checksum file.
- CI runs
go test,go vetandnode --testwith SHA-pinned actions.
Limitations
The README describes it as an early compatibility layer: images, prompt caching and extended thinking are not translated yet. The public history is a single squashed commit.