skill

token-saviour

Agent skill that routes each coding task to the most token-efficient tool per layer: serena for code reads, rtk for command output, caveman for prose, Ponytail for generated code — −70% tokens measure

vagkaratzas/token-saviourcommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/vagkaratzas/token-saviour itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 26, 2026

token-saviour

Spend tokens where they matter. An agent skill that routes every coding task to the most token-efficient tool for the layer it stresses — instead of reflexively reading whole files, dumping raw command output into context, or writing more code than asked.

Token cost has four independent layers, and a different tool owns each. token-saviour picks one winner per layer — the best measured combination from a 9-tool benchmark on a realistic Python codebase:

Layer	Tool	Measured
Code-read input (symbols, callers, call paths, architecture)	serena	−66% alone — the dominant cost
Command-output input (tests, builds, git, grep, listings)	rtk	−65% on a test run, −36% on command tasks
Generated prose output (chatty replies, write-ups)	caveman	−6% on prose answers (grows with chattiness)
Generated code output (implementations you write)	Ponytail	~−40% on a verbose implementation (illustrative)

Stacked (serena + rtk + caveman): −69.6% total tokens on the comprehension suite. For code-generation work, swap caveman → Ponytail (≈ −64%).

The picks were re-validated (2026-07) against newer entrants — archex, Pare, lazy-cat — and every incumbent held its layer. The runner-ups, the niches where they flip (e.g. Pare's structured pytest for test-dominated loops), and the full evidence live in skills/token-saviour/references/.

Install

Claude Code (plugin)

/plugin marketplace add vagkaratzas/token-saviour
/plugin install token-saviour@token-saviour

Or as a plain personal skill, no plugin system involved:

git clone https://github.com/vagkaratzas/token-saviour
mkdir -p ~/.claude/skills
cp -r token-saviour/skills/token-saviour ~/.claude/skills/

Codex

codex plugin marketplace add vagkaratzas/token-saviour
codex plugin add token-saviour@token-saviour

In Codex the skill is invoked with @token-saviour. The VS Code Codex extension and the Codex app read AGENTS.md, which this repo ships — so it also works from the repo root with no setup, and cp AGENTS.md ~/.codex/AGENTS.md makes the always-on rules global.

Any other agent

Copy AGENTS.md (compact, always-on ruleset) or skills/token-saviour/SKILL.md (full playbook) into whatever instruction file your agent reads.

What the skill does

Classifies the task by the layer it stresses: reading code, reading command output, writing prose, or writing code.
Routes to that layer's tool with concrete commands (serena MCP calls, rtk proxies, caveman terse mode, Ponytail rules) — and degrades gracefully to plain Read/Grep/Bash when a tool isn't installed (install/verify commands in references/tool_links.md).
Refuses anti-patterns: two code-read tools at once, rtk for comprehension, caveman on code, Ponytail on prose, tooling-up trivial one-line lookups.
Announces what it used: 🪙 token-saviour: serena + rtk + caveman.

The tools it routes to (install the smallest set you need)

serena — uv tool install -p 3.13 serena-agent — the one code-read tool (LSP symbols; also does semantic edits/renames).
rtk — brew install rtk or cargo install --git https://github.com/rtk-ai/rtk — add only for noisy command loops.
caveman — see repo — add only when prose brevity is the bottleneck.
Ponytail — /plugin marketplace add DietrichGebert/ponytail — add only when code-generation work is the bottleneck.

Evidence

All numbers come from token-consumption-benchmark: real tools, real artifacts, one tokenizer (tiktoken o200k_base), per-task tables, and honest limitations (modeled-not-metered, small clean codebase, single run). Summary: references/benchmark_results.md.

License

MIT — see LICENSE.