The Sifter Agentic Proxy
Your agents are brilliant.
Your API bill is not.
Sifter is a drop-in gateway between your coding agents and the model provider. One environment variable, zero workflow changes, and the same work completes for roughly a third of the cost.
90 seconds, no audio required: what Sifter does, and how we measure it.
How it works
Three mechanisms, one gateway, no code changes.
Ship the evidence, not the repo
Agents resend their whole history every turn. Sifter selects the evidence a turn actually needs from a local index and keeps the working context bounded, deterministically and session-stable.
The right model, every turn
Read-only and narrow turns run on the fast, inexpensive tier. Edits, failures, and anything requiring deep source work route to the strong tier automatically, with escalation gated by real economics.
Engineered for the provider's cache
Providers charge far less when a request starts exactly like the last one. Sifter keeps that opening identical, small, and warm, so you pay cached prices instead of full price, turn after turn.
Why savings grow
The longer the session, the more Sifter saves.
A direct agent re-pays for its entire history on every turn, so its costs climb for the whole session. Sifter keeps the working context bounded and the cache prefix small and stable, so its costs stay flat while the direct curve keeps rising. Short sessions have little to save, and Sifter passes those through untouched rather than charging for optimization it cannot deliver.
Measured, not promised
The proof is your own invoice.
Honest benchmarks
Paired sessions, run direct and through Sifter on the same repos. Quality graded turn-by-turn by an independent judge. Any turn where Sifter's answer was worse is charged the full cost of redoing it without Sifter. Medians across replicates, methodology and corrections history available to any evaluator.
Fail-open by design
Your keys and prompts stay in your environment. If a license is missing, expired, or wrong, Sifter never blocks traffic: it degrades to a transparent passthrough and says so. Your agents keep working, always.
Prove it before you pay
The 30-day trial runs on your repos and your provider account. Compare your invoice before and after: same team, same work. If the savings are not obvious, walk away and nothing breaks.
Pricing
One price. Per developer. That's it.
Developer seats
Full gateway for each developer's agent traffic on API billing: routing, caching, compaction, per-developer telemetry. Run it locally or as a shared container; licensing is per developer either way. A developer spending $300/month on API usage typically nets several times the seat price in savings.
Sifter is for API-billed agent usage. For developers spending under roughly $100/month on API usage, the honest answer is "not yet," and the trial will tell you so with your own numbers.
Get started
Running in under ten minutes.
- Start the trial. Get a 30-day trial license and download links from the trial page. No payment details required.
- Install Sifter with the one-line installer for Windows, macOS, or Linux, or pull the container image. Point
SIFTER_LICENSE_FILEat your license. - Point your agent at Sifter. For Claude Code, set
ANTHROPIC_BASE_URLto your local gateway and keep working exactly as before. Codex CLI, Aider, Cline, and OpenAI-compatible clients connect the same way. - Watch the bill. The dashboard shows spend, model mix, and cache behavior per session. The number that matters arrives on your provider invoice.
Thirty days. Your repos. Your invoice.
If the savings are not obvious by the end of the trial, walk away. Nothing breaks: Sifter fails open.