VIC-E TokenSaver
Local, cross-client token savings with built-in evidence
3 followers
Local, cross-client token savings with built-in evidence
3 followers
TokenSaver is a local context optimizer for Claude Code, Codex, Gemini, Grok, and other AI coding tools. It reduces repetitive logs, searches, file reads, and tool output before they consume paid model context. Unlike command-specific reducers, it provides one cross-client proxy, a native Rust engine, fail-open processing, adjustable profiles, and evidence reports showing measured savings, overhead, failures, and per-client results. Windows, macOS, and Linux; free for personal home use.
Free Options
Launch Team / Built With
Hi Product Hunt! We built TokenSaver after seeing AI coding agents repeatedly fill their context with long logs, searches, file reads, and test output instead of the information needed to finish the task.
Context reduction already exists in tools such as RTK and Headroom, as well as built-in client compaction. TokenSaver is different in its combination of one cross-client local proxy, a native Rust engine, bounded fail-open processing, adjustable profiles, and evidence generated from your own workflow.
Instead of showing only a savings percentage, TokenSaver reports measured tokens before and after optimization, local processing overhead, failures, and per-client results. You can connect Claude Code, Codex, Gemini, Grok, or another supported client, work normally, and generate a self-contained QA report to judge the tradeoff yourself.
In our published 32-scenario benchmark, VIC-E reduced 282,323 benchmark tokens to 60,917 while passing all 32 fidelity checks. We publish the methodology and limitations because this is maintainer-run evidence and should be open to scrutiny.
TokenSaver runs on Windows, macOS, and Linux. Personal, non-commercial home use is free.
I’d especially value difficult feedback: output that should remain exact, missing context, integration problems, or workloads where the overhead is not worthwhile. Thanks for trying it, I’ll be here answering questions throughout the launch.
Nick.
vic-e.com
HARD TRUTH: 80% compression can still cost 31% more. We measured the whole agent task.
A huge compression percentage looks impressive. But did the agent actually spend less to finish the job correctly?
We checked. Our compression-only configuration removed 80.34% of eligible context tokens and still produced a 30.99% higher token-priced task cost than its raw-agent control. Provider input tokens increased 60.39%. Both sides answered every checked question correctly.
That is the trap: the number removed locally is not the number saved at the provider. If an agent has to retrieve missing context, repeat work, or generate more output, those tokens count too. We observed repeated source reads in the transcripts; we cannot attribute every extra token to a particular compression step.
That is why TokenSaver does more than compression.
TokenSaver optimizes the agent’s workflow, not just the size of its messages. We measure success by the cost of a correctly completed task.
The target is a cheaper, correct completed task. We measure provider usage through the finish line.
Here is what TokenSaver achieved against a matched raw-agent baseline:
Five runs on each side
Raw agent
VIC-E TokenSaver
Provider input tokens, including cache
209,089
152,967
Cached input tokens, included above
134,912
100,096
Output tokens
6,356
5,866
Cost at identical token rates
$0.02516084
$0.01961532
Tool calls
20
5
Correct checked answers
50/50
50/50
Median task time
31.29 s
30.45 s
26.84% fewer input tokens. 22.04% lower token-priced cost. 75% fewer tool calls. Every checked answer correct.
Optimization also kept pace: the median was slightly faster, and TokenSaver finished faster in 3 of 5 pairs. Two pairs were slower. We show all five rather than promise that every run will speed up.
Provider input was lower in all five pairs. The completed task, checked answers and provider usage are more useful evidence than a compression counter on its own.
Disclosure: we build TokenSaver, and this is a small, reproducible benchmark. We ran five alternating pairs on one generated Rust investigation using the same model, gpt-5.6-luna, on Linux under WSL. Each trial had fresh local state and one agent at a time. We checked ten answers and their citations per trial. All runs are retained.
The compression-only comparison was a separate cohort with its own matched raw control. We do not combine those cohorts to claim a precise feature-level uplift. We tested our own configurations; this is not evidence that most competing products increase costs.
Both sides use the same published rates per million tokens: $0.20 uncached input, $0.02 cached input, and $1.20 output. Reasoning is included in output and is not charged twice in our calculation. These are normalized task costs, not measured invoice charges. Provider caching was uncontrolled, and one task is not a universal savings average.
We extended the investigation to Claude and Grok. The completed confirmation cohorts also reduced cost, with every checked answer correct.
Model / completed pairs
Raw token cost
TokenSaver token cost
Cost reduction
Correct checks per arm
Claude Haiku 4.5 / 3
$0.16911780
$0.15863180
6.20%
30/30
Claude Opus 5 / 2
$0.49710750
$0.37861700
23.84%
20/20
Grok 4.6, streaming / 2
$0.06990200
$0.04856400
30.53%
20/20
Same cohorts
Provider input, raw → TokenSaver
Output, raw → TokenSaver
Median task seconds, raw → TokenSaver
Haiku 4.5
235,372 → 138,401
14,703 → 13,358
41.72 → 41.35
Opus 5
130,644 → 77,458
4,401 → 3,838
27.77 → 23.25
Grok 4.6
35,029 → 16,242
4,582 → 3,896
31.47 → 26.99
Input includes provider cache reads and, where applicable, cache writes. Output includes thinking once. Each model uses its own published rates, including the correct cache prices and TTLs; the Luna rates above apply only to Luna. Grok's reported token charges also matched our independent price calculation on every metered response.
TokenSaver can finish faster as well as use fewer tokens. Opus and Grok had 16.30% and 14.23% lower median task time in these small cohorts. Haiku's median was essentially unchanged. The goal is a correct completed task with less total work, not a promise that every latency measure improves.
Claude was tested in Claude Code; Grok used our native Rust API investigation agent. Each comparison uses its own matched raw control. Timings across those clients are not directly comparable, and these separate cohorts are not a product-wide average.
Antigravity also completed two live CLI pairs, with 20/20 answer-and-citation checks on each side. Its CLI reported 121,741 → 58,116 input tokens and 16,164 → 2,323 output tokens including thinking. Median process wall time was 25.56 s raw and 11.70 s with TokenSaver.
Before trusting a savings claim, ask for three things: the completed task, the provider token counts, and the checked answers.
That is the standard we are building TokenSaver around.
Explore TokenSaver: https://vic-e.com/products/dev-tools/tokensaver
We tested VIC-E TokenSaver with 3 concurrent Codex agents: 94.25% less eligible log context, 6/6 clean runs.
We ran a straightforward performance test using real Codex CLI agents:
3 concurrent direct agents
3 concurrent agents through VIC-E HTTP
6/6 clean runs
3/3 eligible contexts compressed
Eligible context reduced from 4,905 to 282 tokens. 94.25%
Every agent recovered its exact verification sentinel
Zero failures, timeouts, authentication errors, or fallbacks
Important limitation: this is a local synthetic Linux test of eligible context, not a claim of 94.25% lower billing or universal savings.
We’d appreciate technical criticism. What real-world workloads should we test next?
https://vic-e.com/products/dev-tools/tokensaver