tokensitterReplace old large results
The relay replaces old, large tool results with short stubs and keeps the originals encrypted per key. The lab defaults were: keep 12 messages, results >= 1,500 estimated tokens.
Features/Reversible context
Reversible context
Less old context.
A way back to the source.
Replace eligible old tool results with short stubs, keep the originals encrypted per key, and retrieve them when the model needs them.
Add to Claude CodeOpenAI Codex: coming soon
API relay + recall tool · opt-inEvidence snapshot: 27 Oct 2026 · per-lever live deployment not verified
Inside the workflow
Replace old large resultsOld, large tool results become short stubs.
Recall the originalThe Tokensitter recall tool retrieves an original result.
Respect the cacheThe lever acts when the cache is cold or the expected saving repays the rewrite.
Transparent by designEvery claim needs a baseline
The technique
What changes.
And what does not.
Availability differs between the API relay and the local plugin. Whether each lever is deployed on the live relay has not been independently checked.
Recall the original
The Tokensitter recall tool retrieves an original result. Each recall adds an extra round trip, which belongs in the cost accounting.
Respect the cache
The lever acts when the cache is cold or the expected saving repays rewriting it. It disables itself for a conversation if recalls become too frequent.
The boundary. API-key relay only, off by default, with the recall MCP tool required. On warm-cache transcripts in the study it deliberately saved nothing.
Quality checks
Fewer tokens is only
half the question.
Exact line-number questions test source retrieval. Sessions where the model asked for the original back when the answer sat in a shortened result: 18 of 20. With so few sessions the result is compatible with a real quality drop of up to roughly 10 points, and it does not establish broad reasoning or editing equivalence.
Checks that would strengthen this evidence. They are not yet reported for this feature:
Task correctness
Check answers against an objective test, not against how plausible they look.
Full cost
Include cache writes, cache reads, retries, recalls and helper calls when they apply.
Workload boundaries
Keep the tested model, task type and cache condition beside the result.
Failure reporting
Keep observed misses and untested conditions visible.
How we compare
Same task.
A stated baseline.
5 synthetic agent sessions of 13-18 file reads each, built from files of public packages; Sonnet 5.5. The original transcript is compared with the same request passed through the lever. Warm busy sessions and cold resumes are separate conditions.
Baseline complete-task cost − feature-on complete-task cost
We aim to report alongside: net API cost per completed task · task pass rate
We aim to report alongside: net API cost per completed task · task pass rate
Figures and evidence
The result.
With its boundaries.
Results are copied from our lab reports (snapshot 27 Oct 2026) and were not re-run for this website.
Lab-63%Hono transcript, first turn after an idle gap with a cold cache: $0.1740 to $0.0644
Lab-38%Anthropic SDK transcript, first turn after an idle gap: $0.1353 to $0.0842
Lab · No saving on a warm cache$0.00Saved on a warm, busy session in all 5 transcripts: the lever held back by design
Lab30 of 30Line-number answers correct with the lever (30 of 30 without it)
Read this alongside the figures. The cuts are a spread across five transcripts, not a confidence interval. No warm-cache saving was observed because the lever deliberately held back. Recall adds a round trip.
| Experiment | Workload and date | Result | Limits | Report |
|---|---|---|---|---|
| Reversible-context study | 5 synthetic agent sessions of 13-18 file reads each; Sonnet 5.5 Run date 2026-10-26 | -63% Hono transcript, first turn after an idle gap with a cold cache: $0.1740 to $0.0644 -58% Undici transcript, first turn after an idle gap: $0.2275 to $0.0960 -56% Ajv transcript, first turn after an idle gap: $0.1206 to $0.0536 -51% MCP SDK transcript, first turn after an idle gap: $0.1625 to $0.0789 -38% Anthropic SDK transcript, first turn after an idle gap: $0.1353 to $0.0842 $0.00 Saved on a warm, busy session in all 5 transcripts: the lever held back by design 30 of 30 Line-number answers correct with the lever (30 of 30 without it) 18 of 20 Sessions where the model asked for the original back when the answer sat in a shortened result | Cost of recalls: 38 model calls for the 20 early-question sessions instead of 20. Scripted agent turns, one model, API keys through the relay only. |
|
Lab results at Anthropic API list prices on fixed tasks show what a lever did on those tasks, not what it will do on your bill. For plugin levers on Pro or Max, list-price values describe usage economics, not the subscription invoice. Relay levers apply to API keys only.
Read the measurement method and evidence limitsQuestions, answered
Is this a production savings guarantee?
No. Lab figures describe fixed tasks and settings. Read the quality outcomes and limits next to each result.
Where can I inspect the evidence?
The workload, date, uncertainty and report name are in the table above. The repository is private, so there are no source links, and customer data is never included.
Keep exploring
A little more care.
A clearer picture.
Add Tokensitter to Claude Code and see where your tokens go.
Add to Claude CodeOpenAI Codex: coming soon