Cut LLM bills without touching a single line of code.
Promptflint sits between your application and Anthropic or OpenAI, collapsing bulky system prompts, tool definitions, and file histories into dense representations before they reach the model. Streamed responses pass through unmodified — latency stays the same, the bill does not.
- Latency
- unchanged
- Code
- untouched
- TTFB
- edge-served
Before
58,420 tok
$0.234 billed
After
14,210 tok
$0.057 billed
BASE_URL
https://api.promptflint.dev
API_KEY
your provider key
Sample shown — actual compression varies per workload. Cached tool definitions scored at 10,000+ tokens ship for $0.
How it works
Two env vars. Streamed responses pass through unmodified.
Point your existing Anthropic or OpenAI client at api.promptflint.dev with your provider key attached. The proxy compresses on the way in, the model responds, the proxy forwards — your app code never changes.
- 01
Swap the BASE_URL
Replace the provider host with the proxy. The integration is two lines; the contract is the same OpenAI / Anthropic shape.
- 02
Compress at the edge
A stack of fit-for-purpose compressors collapses prose, rasterises reference docs, and forwards IDs/JOSN/code byte-exact — picked per request.
- 03
Bill the diff
Tokens Billed minus Tokens Saved becomes a running dollar figure in the dashboard. Toggle conservative rules per team, per repo, per model.
The engine
A stack of fit-for-purpose compressors, not a single heuristic.
Each request is scored and routed to the compressor best suited to its payload. Nothing is shrunk that should not be — the bypass keeps what can't be safely compressed.
LLMLingua-family pruning
Semantic pruning collapses bulky instructions, tool descriptions, and prior turns into dense representations — the model keeps the signal, your prompt shrinks.
Fixed-cost PNG rendering
Long-form docs and reference material that don't change between turns are rasterised once and re-injected as a stable prefix — no per-call token tax.
Byte-exact bypass
Hashes, IDs, JSON payloads, and code pass through untouched. The compressor never invents a byte it did not see — no hallucinated syntax reaches the model.
Fingerprinted Redis cache
Compressed artifacts are fingerprinted and held in Redis so a 10,000-token tool definition rendered once serves every subsequent turn in microseconds.
Designed for
Agentic coding pipelines where prompt caches compound.
Claude Code, Cursor, Aider — any loop that re-injects the same tool definitions, project files, and AGENTS.md every turn. High cache hit rates mean every percent of compression compounds into a real budget line.
Globally distributed edge workers
P99 stays flat under load.
SSE passthrough
Streams flow through unmodified end-to-end.
Fingerprinted Redis cache
One render, every subsequent turn.
Byte-exact JSON & code
Bypass — never hallucinated.
Dashboard
Tokens Billed minus Tokens Saved — in dollars, in real time.
A running dollar figure pairs with per-rule toggles for the conservative team. Start tight, loosen as the cache warms.
Conservative teams
Compress system prompts only. Skip tool bodies until you trust the cache.
Balanced teams
Compress prose, rasterise docs, cache tool definitions — bypass code and JSON.
Anomaly teams
Hit the kill switch and route direct to upstream if an agent misbehaves.
Compress system prompts only
Never compress code or JSON
Bypass for indexed tool definitions
Route direct to upstream — kill switch
Tokens saved
44,210
Billed
$0.057
Saved
$0.177
Hit rate
92%
Plug it in
A BASE_URL swap. That is the entire integration.
No SDK. No agent-side code. Your provider key keeps working against the same Anthropic or OpenAI contract — only now the compression stack is between you and the model.
Or write to promptflint@polsia.app.
Inbox
Talk to a human
Questions about fit, rollout, procurement, or a conservative start? Drop a note and we'll route the conversation.
promptflint@polsia.app