01
History on repeat
Every turn re-sends the entire growing transcript. You pay for the same tokens again, and again, and again.
Your Claude, minus the wasted tokens.
Tidbit sits between your client and the model, and trims the waste out of what you send. Point your client at it once. No SDK, no code changes, and the replies come back untouched.*
tokens per task
without Tidbitfull
with Tidbitreduced
outcome depends on workload
Works with✓ Claude Code✓ Claude Desktop✓ the Anthropic API
Works with Claude Code · Desktop · APINo SDK · no code changesCut the waste · keep the qualityPrivate by designSet up in minutesWorks with Claude Code · Desktop · APINo SDK · no code changesCut the waste · keep the qualityPrivate by designSet up in minutes
Where your tokens go
You're paying for the same tokens, over and over.
The same context gets re-billed turn after turn, and it adds up quietly, in more ways than one. The longer and more complex the task, the more there is to save.
02
Cache left on the table
Repeated context can bill at a deep discount, but only when it's flagged correctly and reused in time. Most setups miss it, so every turn pays full price.
03
Tool-output bloat
One giant file read or log dump rides along inside every single request that follows it.
The usual fixes don't fit: SDK rewrites mean code changes in every app, quota walls stall the team mid-task, and nobody can see where money actually went.
How it works
One setting, three steps.
Your clientCLI · desktop app · SDK
Tidbit gateway
- 1Point your client at Tidbit. One setting, whichever client you use. Your key stays with you.
- 2Tidbit trims what goes upstream. It reworks the input side of your traffic before it reaches the model. The reply comes back untouched.
- 3Watch it in the Console. Today and historical views of how much was sent, what it cost, and what was saved.
Anthropicapi.anthropic.com
The savings
Savings depend on the workload.
Two different things get smaller: what you send, and what you pay when context repeats. We measure them separately, and every mark below names the tracked run it came from.
14%Fewer billed input tokens
One large tool-heavy request, 197,521 tokens. The recorded answer check passed. Evidence: large-trace-lossless.
61%Lower billed cost, four sends
The same stable context sent four times, with the cache-filling send inside this measurement. Evidence: cache-reuse-4.
88%Lower billed cost, 24 turns
A pre-warmed 24-turn window, with the cache-filling send outside this measurement. Evidence: cache-session-24.
Every number above is regenerated from a committed result file. These runs used our most conservative setting. See all four tracked runs →
The dashboard
Proof you can audit.
The product view reports observed workload outcomes.
usetidbit.ai/app
Cost of tokens saved$1.20on $4.78 spent
Tokens saved15.0%1.94M of 5.09M tokens
Spend efficiency25.0%saved 35.4% of pre-Tidbit spend
Requests served80this period
Saved over time
this weekIllustrative trend; measured results vary by workload.
Cache attribution
Tidbit-added vs client20%80%
Tidbit-added Client cache
Cost of tokens saved over time
- Illustrative sample data
- Results vary by workload
- Live + historical views
Drop-in, reversible
Start for free →See it work on your very next request.
Works where you already work
One engine. Three ways to run it.
The same optimization core, wherever you already work: the CLI, the desktop app, or any tool built on the API.
01
Setup scriptClaude Code
Run one script and the CLI routes through Tidbit. Same commands, same workflow, calls optimized from then on.
02
One-click installerClaude Desktop
A one-click Mac installer wires the desktop app through Tidbit: no terminal, no config. It is signed, verifies each of its own updates before the swap, and refuses downgrades. The coding work you do in the app is optimized on the way out.
03Any tool
Endpoint swapAnthropic API
Already building on the API? Point any tool at the Tidbit gateway URL with your key. SDKs, scripts, and agents all keep working.
Pricing
Pricing that follows the plan you already have.
On a Claude subscription, Tidbit is free. On the API, Tidbit is free for the duration of early access. Your Console shows the savings we measure and attribute to Tidbit's own work.
Subscription
Free
On any Claude plan, the full gateway and Console, no charge.
API
Freeduring early access
Start free →
No card to start.
Questions
The honest FAQ.
What does Tidbit cost?
On a Claude subscription, Tidbit is free, including the full gateway and Console. On the Anthropic API, Tidbit is free for the duration of early access. Your Console shows the savings we measure and attribute to our own work. Savings produced by your own client caching will never be billed.
Do I have to change my code?
No. Tidbit sits in front of your existing setup as a gateway. Point your client at Tidbit and keep your current workflow, with no SDK swap and no library migration.
What did you actually test for quality?
We ran a small number of live workloads twice over: once straight to Anthropic, once through Tidbit on our lossless profile, the setting where nothing is dropped from your request. On those runs the answer check passed both ways, and the input and output token counts matched across the repeated-context turns we measured. That is a handful of tracked runs on particular workloads, so read it as evidence about those runs and not as a general quality or latency guarantee. The method and the raw numbers are on the benchmarks page.
How do you prove the savings are real?
Every number we publish comes from a recorded benchmark run rather than a projection. Each one carries an evidence ID, a fingerprint of the source data, and a derivation checked into the repository, so it can be reproduced from the recorded run. The savings in your Console are a separate thing, measured from your own account activity.
How do you keep my data and key secure?
Tidbit does not use request content to train a model and does not sell data. On the API, you keep your own key and Tidbit uses it only to forward your request. On a Claude subscription, your client keeps signing in the way it already does and we never hold a key for it. Retention windows and processor terms are in the privacy policy.
Are you pooling my keys or doing rate-limit tricks?
No. Your traffic always rides your own credential, never a shared or pooled account, and our own key pays only for our benchmark runs. There is no rate-limit circumvention and no terms-of-service workaround. The efficiency comes from sending fewer tokens, not from evading limits.
Which clients and models does it support?
Claude Code, Claude Desktop, and services that call the Anthropic API. Tidbit forwards your request to Anthropic rather than choosing a model for you, so it works with the Claude models your client already uses.
What happens if Tidbit goes down?
You can point your client back at Anthropic at any time and carry on, because removal is immediate in every mode. If Tidbit cannot resolve a credential it refuses the request outright rather than quietly sending it somewhere you did not choose. We do not have a live status page yet, so until that exists the trust page and direct email are where we report incidents.
How easy is it to remove Tidbit?
Removal is one step. Point your client back at the direct endpoint and the path is exactly as it was before, with nothing left behind in your application code.
What we don't do
No surprises, by design.
Doesn't alter your prompts
We add cache hints and trim re-sent history. What you actually wrote reaches Claude untouched.
Doesn't send your data to another AI provider
Your traffic goes to Anthropic and nowhere else. No third-party models, no detours, and nothing you send is collected or used for training.
Doesn't touch your output
We only optimize the input you send; the model's response is never altered.
Doesn't store your key
Your key is used to forward the request to Anthropic, and nothing more.
Start cutting tokens today.
Keep every workflow, and watch the savings add up live. Free to start, no card required.
No card. No SDK. Switch it off whenever you want.