A Go port of the core of pxpipe. It converts large system prompts, tool definitions, earlier conversation turns, and tool results into dense PNG pages when the vision input is estimated to cost fewer tokens than the source text.
The transform supports Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses request bodies. The Gemini/Google API surface, upstream dashboard, and offline export CLI are not included.
| Interface | Entry point | Use case |
|---|---|---|
| CLI wrapper | pxpipe <executable> |
Run Claude Code, Claude Desktop, OpenCode, or Codex through a process-scoped proxy |
| Standalone proxy | pxpipe serve |
Connect an existing client to a loopback reverse/forward proxy |
| Go package | pxpipe.NewHandler and transform functions |
Embed the proxy or request transform in a Go server |
pxpipe-go requires Go 1.26.3 or later.
go install github.com/evan-choi/pxpipe-go/cmd/pxpipe@v0.4.19
pxpipe --versiongo install writes the binary to go env GOBIN, or to
$(go env GOPATH)/bin when GOBIN is unset. Add that directory to PATH if
needed.
Add the Go package to a module with:
go get github.com/evan-choi/pxpipe-go@v0.4.19pxpipe claude --model opus
pxpipe Claude.app
PXPIPE_MODELS=gpt-5.6-sol pxpipe opencode --model openai/gpt-5.6-sol
PXPIPE_MODELS=gpt-5.6-sol pxpipe codex --model gpt-5.6-solArguments after the executable are passed to the child unchanged. Proxy and
local CA settings apply only to the child; pxpipe does not modify the system
trust store. A Codex provider must set supports_websockets=false so request
bodies use HTTP and can be transformed.
pxpipe serve
pxpipe serve --port 8080The default port is 47821. The startup view prints ready-to-run Claude and
Codex commands. The Codex command includes the generated CA path and proxy
environment.
ANTHROPIC_BASE_URL=http://localhost:47821 claudeNewHandler forwards to the public Anthropic and OpenAI APIs by default.
package main
import (
"log"
"net/http"
pxpipe "github.com/evan-choi/pxpipe-go"
)
func main() {
handler := pxpipe.NewHandler(pxpipe.HandlerOptions{})
log.Fatal(http.ListenAndServe("127.0.0.1:47821", handler))
}Point Claude Code at the handler:
ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claudeSee the usage reference for custom upstreams, route mapping, credential injection, and pure transforms.
| Setting | Behavior |
|---|---|
PXPIPE_MODELS |
CSV list of model base IDs to transform; models outside the list pass through unchanged |
PXPIPE_GPT_PROFILES |
JSON profiles for GPT-compatible models not in the built-in table |
PXPIPE_RENDER_CACHE_BYTES |
Render cache capacity; defaults to 64 MiB, and 0 disables it |
The library and CLI wrapper have a built-in allowlist of claude-fable-5 and
gemini-3.6-flash; GPT models are opt-in. pxpipe serve accepts every valid
Anthropic and OpenAI model when PXPIPE_MODELS is empty.
- Usage reference: CLI profiles, standalone proxy, Go handler, and pure transforms
- Performance and compatibility: golden parity, benchmarks, and reproduction commands
- Upstream port status: pinned revision, port scope, and fixture verification
On an Apple M1 Pro, the four measured workloads ran 72.6–282.2 times faster than the TypeScript implementation, while peak RSS for the full suite was 73.4% lower. See performance and compatibility for the measurement conditions.
Golden fixtures come only from the pinned TypeScript implementation. Tests compare text, gate decisions, page geometry, and decoded pixels.
go test ./...MIT. Portions derived from teamchong/pxpipe (MIT).