将文本上下文渲染为图片，本地代理为 Claude Code 削减约 60% token 成本

pxpipe is a local proxy that rewrites bulky text content — system prompts, tool documentation, and older conversation history — into compact PNG images before they reach the Claude Code API. Since vision tokens are priced by pixel dimensions, not by the amount of text embedded, dense content like code, JSON, and command output can be packed at roughly 3.1 characters per image token versus 1 character per text token. On production traffic, this translates to an end-to-end reduction of 59–70% in token costs, depending on workload. The proxy sits transparently between your CLI and Anthropic's API, intercepting only the request side and leaving model responses untouched.

Claude Code sessions accumulate large system prompts, tool definitions, and extensive history as the session progresses. Every turn repeats the entire static prefix and much of the history, inflating input tokens linearly with session length. Traditional compression approaches that operate on raw text yield at most a ~2x improvement, because text is already relatively dense. The vision API's fixed-pixel pricing breaks this paradigm entirely: an image block costs the same in tokens regardless of how much text is rendered inside it, allowing dense context to be compressed 3–4× more effectively than any text-based method.

pxpipe intercepts /v1/messages API calls, identifies three categories of eligible input: large tool_result bodies (file reads, command output, logs) above ~6k characters of token-dense content; older collapsed history (recent turns stay text); and the static system prompt + tool docs slab. Each block is wrapped at 1928px-wide columns and packed into one or more PNG images (≈92,000 chars per 1928×1928 image, costing ≈4,761 vision tokens). Crucially, a per-request profitability estimator decides whether imaging is beneficial: if the content is sparse prose (> ~19 chars per token in text form), it remains text. The system prompt prefix is preserved to maintain prompt caching compatibility. The exact savings per request are measured against a counterfactual count_tokens probe fired in parallel with the real forward, and both are logged to ~/.pxpipe/events.jsonl for auditability.

The proxy runs as a local HTTP server (default on 127.0.0.1:47821). You point your Claude Code CLI at it via ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claude. For each request, it: (1) receives the JSON body; (2) applies the profitability gate to each eligible text block; (3) renders blocks that pass the gate into PNG images using hard-wrapped columns and monospace fonts; (4) splices the image blocks back into the request payload, preserving the static prefix and cache_control markers; (5) forwards the rewritten request to the actual Anthropic API; (6) returns the unmodified response to the CLI. The dashboard at the same port shows live savings, side-by-side conversions, a kill switch, and model chips to control which models are imaged. The source code is structured around src/core/ with the rendering logic, src/worker.ts for the proxy server, and src/dashboard/ for the web UI.

Installation is a single command: npx pxpipe-proxy starts the proxy on 127.0.0.1:47821. Then point Claude Code at it: ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claude. The dashboard at http://127.0.0.1:47821/ shows tokens saved, all conversions, a kill switch, and model chips. For library use without the proxy, the package exports renderTextToPngs (text → images) and transformAnthropicMessages (rewrite a raw request).

pxpipe shines on token-dense workloads like coding sessions with large file reads, logs, and tool outputs, where the end-to-end bill can drop 59–70%. It wins on tasks where the model needs gist understanding rather than byte-exact recall — the imaging is lossy. Exact 12-character hex strings in dense renders are recalled at 13/15 on Fable 5 but 0/15 on Opus, and misses are silent confabulations, not errors. Therefore, byte-exact values (IDs, hashes, secrets) must stay text, which pxpipe does by keeping recent turns in text. The verbatim-risk guard is not yet built, so subagents on non-allowlisted models pass through entirely. Models outside the default scope (Fable 5, GPT 5.6) are opt-in; Opus 4.8 and GPT 5.5 show measurable misread rates and are deliberately excluded. PNG encoding adds latency to large requests before they leave. ASCII/Latin-1 content is well tested; CJK works but with conservative sizing. The gains depend on workload — sparse prose at >19 chars/token may actually cost more when imaged.