explainx / blog
pxpipe is a local proxy that turns bulky Claude Code context โ system prompt, tool docs, history โ into PNG pages for ~59โ70% lower Fable bills. How it works, benchmarks, lossy caveats, and when not to use it.
Jul 4, 2026
A new post from Claude Code engineer Thariq argues that working with Claude Fable 5 keeps re-teaching an old lesson: your prompt, skills, and context are a map of the work, not the work itself. Here's what that means for PRs, skills, and finding the unknown unknowns before your agent does.
Jul 5, 2026
Anthropic has changed Claude's usage limits at least three times since August 2025: weekly caps, a temporary off-peak doubling, and a permanent doubling of Claude Code's 5-hour limits tied to a SpaceX compute deal. Here's the dated timeline so you know which limit you're actually hitting.
Jul 4, 2026
@ClaudeDevs posted "/radio" with a strawberry-and-sun pixel art clip โ and 400K+ developers asked the same question. It is not FM radio in your terminal. It is Claude FM: a built-in slash command that opens Anthropic's lo-fi stream while you code.
What if your system prompt cost 2,700 tokens instead of 25,000 โ and the model still read every line? pxpipe (pxpipe-proxy on npm) is a local proxy that does exactly that: it intercepts Claude Code requests, reflows bulky text context into dense PNG pages, and forwards the rest unchanged. Same session. Same tools. Fraction of the input tokens.
The repo hit 2.2k GitHub stars with v0.8.0 shipping July 2026. The headline claim: ~59โ70% lower end-to-end bills on Fable 5 list pricing โ but the authors are explicit that workloads differ and the durable number is measured token cut per request, not marketing math.
| Question | Answer |
|---|---|
| What is it? | Local HTTP proxy โ text โ PNG for bulky context |
| Run how? | npx pxpipe-proxy โ ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claude |
| Dashboard | http://127.0.0.1:47821/ โ savings, conversions, kill switch |
| Default models | Fable 5 + GPT 5.6 only ยท Opus/GPT 5.5 opt-in |
| Savings | ~59โ70% end-to-end on measured Fable traffic (varies) |
| Lossy? | Yes โ hex/IDs not byte-safe in images |
| Output touched? | No โ compresses request only |
| License | MIT ยท TypeScript ยท Node + Cloudflare Workers |
LLM providers bill text tokens by tokenizer chunk count. Vision billing is different: an image costs tokens from resolution, not how many characters you painted into the PNG.
pxpipe's README states the gap on real Claude Code traffic:
| Mode | Dense content efficiency |
|---|---|
| Text tokens | ~1 char per billed token |
| Image tokens (dense render) | ~3.1 chars per vision token |
A 1928ร1928 page โ 4,761 vision tokens and holds โ 92,000 characters of wrapped monospace text. Text only wins above ~19 chars per text-token โ Claude Code sessions average ~1.91 (N=391 production rows), so imaging wins often.
Concrete example from the repo: ~48k characters of system prompt + tool docs โ โ25k text tokens as plain text, โ2.7k image tokens as one 1573ร1248 page. That is the visual the model sees โ instruction banner on top, โต marking original newlines.
If you are new to how providers count units, start with what are LLM tokens โ pxpipe games the input side of that ledger.
npx pxpipe-proxy
ANTHROPIC_BASE_URL=http://127.0.0.1:47821 claude
Open http://127.0.0.1:47821/ for:
Responses stream normally โ pxpipe never compresses the model's reply. Recent conversation turns stay text; static system prompt, tool docs, and older collapsed history are the usual imaging targets.
Three buckets, each behind a profitability gate (sparse prose stays text):
| Bucket | Rule of thumb |
|---|---|
Large tool_result bodies | File reads, logs, command output above ~6k chars of token-dense content |
| Older history | Turns behind the live tail โ image pages; recent turns always text |
| System prompt + tool docs slab | Static prefix reflowed into page(s), cache-friendly splice |
Passes through unchanged: your latest messages, small blocks, sparse prose, models outside allowlist, and all model output.
GPT path note: tool definitions stay native JSON; no Anthropic cache_control markers on OpenAI transforms.
Set PXPIPE_MODELS=off to disable imaging entirely. Default: claude-fable-5,gpt-5.6.
tool_result string
โ wrap at 1928px-wide columns
โ pack ~92,000 chars/page
โ PNG[]
โ splice into /v1/messages (cache-friendly)
โ forward to Anthropic
Events log to ~/.pxpipe/events.jsonl. Each row records:
count_tokens on the original uncompressed body (parallel probe)usage from the real responseThat is how the README avoids inflated "savings on the slice we touched" โ end-to-end denominator includes requests pxpipe correctly left alone, cache reads/writes, and all output tokens.
Dollar math in docs uses Fable 5 list ratios (input ร1.0, cache write ร1.25, cache read ร0.1, output ร5) applied identically to both sides so cache discounts cancel.
Side-by-side A/B demo (plain left, pxpipe right):
| Test | N | Text arm | pxpipe (image) | Token ฮ |
|---|---|---|---|---|
| Novel arithmetic, Fable 5 | 100 | 100% | 100% | โ38% |
| Novel arithmetic, Opus 4.8 | 100 | 100% | 93% | โ38% |
| Gist recall A/B (15kโ45k char sessions), Fable 5 | 98/arm | 98/98 | 98/98 | โ |
| State tracking (mutations), Fable 5 | 18/arm | 18/18 | 18/18 | โ |
| Confabulation on never-stated facts (lower better), Fable 5 | 16/arm | 0/16 | 0/16 | โ |
| Verbatim 12-char hex, dense render, Opus | 15 | 15/15 | 0/15 | โ |
| Verbatim 12-char hex, dense render, Fable 5 | 15 | โ | 13/15 | โ |
SWE-bench Lite pilot: 10/10 both arms at โ65% request size.
SWE-bench Pro: 14/19 ON vs 15/19 OFF at โ60%; verdicts agree 18/19; single split re-resolved 3/3 on replication (variance, not compression failure per FINDINGS.md).
Receipts: eval/swe-bench/, eval/swe-bench-pro/, eval/needle-haystack/, eval/gist-recall/ on GitHub.
pxpipe's README leads with failure modes โ unusual and worth copying:
CLAUDE_CODE_SUBAGENT_MODEL=claude-sonnet-4-6)keepSharp(block) in library API pins blocks as textWins on token-dense content (~1 char/token). Can lose money on sparse prose (~3.5 chars/token). Profitability gate calibrated on N=391 production rows.
import { renderTextToImages, transformAnthropicMessages } from "pxpipe-proxy";
const { pages } = await renderTextToImages(toolResultText);
// pages[i].png: Uint8Array
const { body, applied, info } = await transformAnthropicMessages({
body: requestBytes,
model: "claude-fable-5",
});
options.keepSharp(block) forces text; options.emitRecoverable returns originals of imaged blocks. Pure JS runtime (Node + edge Workers); @napi-rs/canvas is build-time only.
/compact vs smaller context windows| Approach | Mechanism | Trade-off |
|---|---|---|
/compact | Summarize + drop history | Fast, but deletes detail |
| Smaller model | Cheaper per token | Less capability |
| pxpipe | Full content as images | Lossy on exact strings; PNG CPU cost |
| Shorter prompts | Human discipline | Free, but fights agent harness growth |
After Meta's 73.7T token month and Tesla's $200/week caps, input compression is no longer a hobby project โ but end-to-end measurement matters. pxpipe's events.jsonl counterfactual pattern is the right audit habit.
Pair with Claude usage limits timeline so you know when burn is plan-gated vs dollar-gated.
From the README:
"Why does the README read like an AI wrote it? Because one did. Most of this repo's commits โ the code and the docs โ were authored by Opus/Fable agent sessions running behind pxpipe itself, reading their own collapsed history as image pages while they worked."
That is either the best dogfooding story of 2026 or the most recursive disclaimer โ possibly both.
eval/glyph-matrix/)Each ships with an n or gets cut, per maintainers.
Does pxpipe work with Cursor / other Claude clients?
Anything that respects ANTHROPIC_BASE_URL can route through it โ Claude Code is the documented path.
Is this cheating prompt cache?
Repo claims cache-friendly splice โ static prefix preserved so Anthropic prompt caching keeps working. Verify on your traffic via events.jsonl.
Should enterprises run this?
Pilot on non-production repos first. Lossy compression + silent hex misses is a governance conversation โ especially for regulated IDs and audit trails. See enterprise token governance.
Stars / releases?
~2.2k stars, 135 forks, v0.8.0 (July 2026), MIT license, teamchong/pxpipe.
pxpipe v0.8.0 as documented July 2026. Savings figures from project FINDINGS.md and README โ re-derive on your own ~/.pxpipe/events.jsonl before budget decisions.