With capsul, your €20/month Claude and ChatGPT plan feels a lot bigger.
capsul sends Claude & Codex only what your task needs, never your whole repo. Fewer tokens per request means the plan you already pay for stretches much further.
$ npm i -g @penra/capsulYour repo goes in. Only the lines your task needs come out.
context sent to the model
13,837/ 13,837 tokens
against a naive full-file dump · the real alternative, `claude -p`, is measured in the benchmark below
/benchmark · September 2026
Measured with Claude and Codex.
What actually goes over the wire, model by model.
The same five short questions, to the same model, with and without capsul. Bars show the reduction in cache-weighted input. These are separate campaigns; savings vary with the task and cache.
share of the input capsul does not send
saved by capsulstill sent= claude -p
Haiku 4.5
−73 %
- claude -p
- 396 725
- capsul
- 106 491
Sonnet 5
−50 %
- claude -p
- 244 486
- capsul
- 123 370
Sonnet 4.6
−51 %
- claude -p
- 147 746
- capsul
- 72 692
Opus 4.6
−59 %
- claude -p
- 171 370
- capsul
- 70 482
Opus 4.7
−32 %
- claude -p
- 185 278
- capsul
- 126 413
Opus 4.8
−80 %
- claude -p
- 358 993
- capsul
- 70 478
Opus 5
−58 %
- claude -p
- 189 342
- capsul
- 79 166
Fable 5
−45 %
- claude -p
- 160 682
- capsul
- 88 530
Fable 5.1
−58 %
- claude -p
- 224 278
- capsul
- 95 064
/agent
One CLI. Two agents. The subscription you already have.
capsul doesn't ask for an API key to get started. It drives the agent CLI you're already signed into.
Claude
your Pro or Max plan
capsul drives the Claude CLI you're already signed into, so the subscription you pay for every month is what answers. No second bill, no API key to paste.
Codex
your ChatGPT plan
The same deal on the other side: your existing Codex sign-in does the work, with continuation, usage and reasoning effort reported back into capsul.
Same budget, same approval prompts, same ledger, whichever one answers.
$ capsul doctor
Your code stays yours. Here is exactly what leaves.
The work happens on your machine. What crosses the wire is short enough that you can read all of it before it goes, so here it is, both sides of the line.
Sent to the model
- Your task, word for word.
- The files you pointed at.
- Short extracts of the repo you are in: what the task needs, not the repo.
- Whatever the agent then asks to open. Every request shows up in your terminal.
Never sent
- Whole files nobody asked for.
- Anything outside the repo you ran capsul in.
- Your past sessions.
- Your keys, remotes and environment variables.
Bring your own key
Your key, or a CLI you are already signed into. No capsul model on the path, no relay in the middle.
Our server counts, it never reads
It is told that a prompt ran, never the prompt, the repo or the reply. On the free tier that is a counter; once you subscribe there is nothing left to count.
Keys where your system keeps secrets
The OS keychain first. Where there is none, a file only your account can open. Never in your repo.
Other CLIs stay read-only
capsul checks they are signed in. It never reads, copies or forwards their token.
The ledger is local
Your savings history is a file on your own disk. Nothing about it is uploaded.
capsul doctor, and see for yourself
One offline command names where your keys live, what protects them and which CLIs it found. It says whether a key is set, never what it is.
Ten prompts free. Then €5.99 a month.
Try it on a real task before you decide anything. Solo or a whole team.
Free
- 10 prompts, no card
- The full engine: map, matching, budget, ledger
- Claude and Codex
Pro
unlimitedLaunch offer
- Unlimited prompts
- Everything in Free
- Savings history across machines
Team
15 seats included, then €3.50 each
- Unlimited prompts for every seat
- Invite by email, each seat is a separate account
- One bill; seats you add, remove or revoke at any time
$ capsul --help
The questions people actually ask.
What does capsul actually send to the model?
Your task, the files you point at, and short extracts of the repository you are in, never the whole thing. The full inventory, both what leaves and what never does, is in the section above.
Does this break prompt caching?
No, and that is a design constraint rather than an afterthought: what capsul sends is built so a cache stays useful. Caching itself is on the roadmap as a lever in its own right, and it will be measured before it is claimed.
Does my code leave my machine?
Only what the request needs. Not whole files nobody asked for, not your past sessions, not your keys: those stay in your OS keychain, and the savings ledger is a plain file on your own disk. capsul doctor shows you where they live without ever showing what they are.
Why not just send everything and use a big context window?
Because you pay for every token of it, on every single request, and a model given your whole repository has to find the corner that matters before it can answer. A smaller, better-chosen context is cheaper and easier to answer from.
Which languages does it understand?
JavaScript and TypeScript are the most precise. Python, Go, Rust, Java and Ruby are well covered. Past those, capsul still works on any text file at all. What changes is how sharp its reading is, not whether it runs.
What if the budget is too low for my task?
capsul stops at the ceiling and tells you what it left out: nothing spills over quietly. Raise it with --budget, or name the file you care about with --open. It is a dial, not a guess.
How does Team work?
One subscription, held by the person who pays, with 15 seats included (up to 30 at €3.50 each; more on quote). From the Team page of their account, the owner invites people by email: each one gets a link valid for 7 days and takes a seat by signing in with that address. A seat is a full capsul account, with its own terminal, unlimited prompts and its own savings ledger. The owner can remove a member, a member can leave, and both see the seat number and the subscription status. When the subscription is cancelled, access lasts until the end of the paid period, then every seat falls back to the free tier at once. Team does not include API access or BYOK: the CLI's API-key mode is a per-session choice, available on every plan.
How do I cancel?
From your account, in one click, and it stops at the end of the period you already paid for. The free tier gives you ten prompts on your own repository before any of that becomes a question.
$ reading
Where the tokens actually go.
Guides and a glossary on token usage, context and cost in CLI coding agents. Written to be useful whether or not you ever install anything.
- How to reduce Claude Code token usageSeven concrete ways to cut the tokens Claude Code sends on every turn, from scoping the request to capping the context, with what each one is actually worth.
- Claude Code usage limits, and how to stop hitting themWhat the Claude Pro and Max session and weekly limits actually count, why they arrive mid-task, and the habits that keep a subscription lasting the whole week.
- Codex CLI token usage: reading it, then lowering itWhere the tokens go in an OpenAI Codex CLI session, how to see the real input count before spending a request, and what reduces it without changing the answer.
Get more out of the plan you already pay for.
$ npm i -g @penra/capsul