What is Chisle?
Chisle cuts a coding agent's token use on three fronts: a ruleset for terse prose, a YAGNI-first rule for code, and hooks that compress command output before the model reads it. It began as a merge of the Caveman and Ponytail rulesets, and its command filters follow the design of RTK. It installs with `npx chisle`, as a Claude Code plugin, or by script on macOS, Linux and Windows, and supports Claude Code, Pi, Cursor, Codex, Gemini and four other agents.
A `chisle tools trim` command also denies tools a coding loop never calls and shortens descriptions of unused skills, to shrink every request. The repository was created on 30 June 2026; v3.14.2 shipped on 19 November 2026. It is MIT-licensed with zero dependencies.
Is Chisle worth it?
The headline numbers come from chat-style answers with tools switched off. In the 3.14 run, 13 prompts with 2 seeds on Haiku 4.5, Chisle billed 69% of the bare model's tokens (95% interval 59 to 81%), level with Caveman at 69% and ahead of Ponytail at 80%, while writing the shortest answers and the only set with no non-answers. On Sonnet 5.5 it billed 53%. Maths and coding accuracy held at 156 of 156 problems.
The agentic table is the honest one. Across 7 fixtures and 5 seeds, Chisle 3.14 passed 31 of 35 tasks, the same as the bare agent, and cost 105% as much per run, within noise; version 3.13 had cost 9% more, outside it. What moved is tool output: on a fixture whose test run fails noisily, the agent read 76,111 characters bare and 6,862 with 3.14's filters. On a corpus of 37 commands the filters left 28% of the characters, against 84% for RTK alone.
The catch: rulesets make Haiku think more, not less. Hidden reasoning ran at 172% of the bare model, which is why billed savings trail the shorter answers. The maintainer's own replay of 33,449 real tool results saved little, because those sessions already ran under RTK, and the same notes on where spend goes put cache reads at 75% and model output at 7.5%. Writing less is the smaller lever.
Our take: of the terse-answer packs, this is the one to trust, because its tables can be rerun from committed transcripts. Expect the output filters, not the clipped prose, to pay for the install.
Where it earns its place
A contractor billing Claude Code API costs to clients
Shrink test and build logs that flood the context window
Install with `npx chisle` and keep the PreToolUse filters on; in the maintainer's noisy-log fixture tool output fell from 76,111 to 6,862 characters. Free; savings depend on how much of the session is command output.
An engineering lead at a 25-person team choosing between token-saving rulesets
Pick one on evidence rather than README banners
Rerun the committed side-by-side suite with the published commands against Caveman and Ponytail on the team's own model, and compare billed tokens and backfires within one run, as the docs advise.
Before you commit
- Setup
- `npx chisle`, or `claude plugin marketplace add JayPokale/Chisle` then `claude plugin install chisle@chisle`; curl and PowerShell install scripts for other setups.
- Integrations
- Claude Code, Pi, Cursor, Codex, Gemini and four more agents; works alongside RTK; `chisle tools trim` edits Claude Code permissions and skill overrides.
- Data and privacy
- Rules and hooks run locally; the filters rewrite command output before the model sees it, so nothing extra leaves the machine. MIT licence.
- Cost at scale
- Free. In the maintainer's agentic benchmark cost per run was 105% of the bare agent, within noise; savings show up mainly in output-heavy chat and noisy tool output.
- Limits
- Most benchmarks use Haiku 4.5; separate runs of the same rivals differed by 10 to 30 points, so only within-run comparisons hold.
What’s good
- Every benchmark transcript committed, losses included
- No accuracy loss on 156 maths and coding problems
- Output filters cut noisy tool output sharply
- Zero dependencies; installs into nine agents
What’s not
- No cost saving in the agentic benchmark
- Raises hidden reasoning on Haiku
- Benchmarks run by the maintainer, mostly on Haiku
- Overlaps with Caveman, Ponytail and RTK
Best for
Developers paying per token for agent sessions who want shorter replies and smaller tool output with published evidence.
Chisle alternatives
Chisle: common questions
- Is Chisle worth it?
- Chisle publishes the runs it loses, which makes it the token-saving ruleset whose claims you can actually check, and the checks say the savings are real but smaller in agent work. We rate it 3.8 out of 5.
- How much does Chisle cost?
- Free and open source (MIT). You pay for your agent's model usage.
- What are the downsides of Chisle?
- No cost saving in the agentic benchmark; raises hidden reasoning on Haiku; benchmarks run by the maintainer, mostly on Haiku; overlaps with Caveman, Ponytail and RTK.
- Who is Chisle best for?
- Developers paying per token for agent sessions who want shorter replies and smaller tool output with published evidence.
- How much does Chisle cost for a team or in production?
- Free. In the maintainer's agentic benchmark cost per run was 105% of the bare agent, within noise; savings show up mainly in output-heavy chat and noisy tool output.
- Is Chisle safe for company data?
- Rules and hooks run locally; the filters rewrite command output before the model sees it, so nothing extra leaves the machine. MIT licence.
- What are the best Chisle alternatives?
- The closest alternatives we have reviewed are Caveman, Ponytail and rtk (Rust Token Killer).
Answers checked Nov 21, 2026.