Free, open-source Claude Code plugin
Claude Code reads whole files. Gio hands it the exact lines.
Gio is a plugin, not a proxy: a skill plus a few local Python scripts. It indexes your repo, turns each question into file:line spans, and tells Claude to read only those. Result on real repos: 6–14× fewer input tokens per task, same model, no API key, nothing leaves your machine.
/plugin marketplace add blkdynamite/Gio-prompt-optimizer
/plugin install gio@blkdynamite-plugins
Two commands, zero config. Paste them inside Claude Code; the index builds itself on first use. Manual install · v0.4.0
Built for developers on Pro and Max plans who keep hitting the session cap, and for anyone paying per token on a codebase bigger than a weekend project.
Index wrong? Gio reads every returned span to verify it before acting, and falls back to plain grep and whole-file reads when it doesn’t match. Say “read the whole file” and Claude will. Nothing is ever blocked.
Measured, not promised
Every number here comes from a script in the repo that you can run yourself, with the caveats stated next to it.
Fewer input tokens on Click
Eight generic questions (“where is the main entry point”, “where are errors handled”…) against pallets/click 8.1.7: Gio’s spans fed 21,858 tokens; reading the whole files would have been 296,388.
fastapi 0.115.0, 6.1× on Gio’s own small repo. bench/self_benchmark.py, 2026-09-06.
Retrieval hit@5, zero installs
Across 77 labeled queries on three repos, the right code is in Gio’s top five results 87% of the time (100% for identifiers and cross-file lookups) using only the built-in lexical ranker.
Deterministic, CI-gated at 80%. eval/RESULTS.mdContext tokens on real fixes
Two coding tasks on Click run with and without Gio, same model, patches applied and tested. Task B: 4 turns for 11¢ with Gio versus 28 turns and 87¢ without.
A demonstration (two tasks, one run each), not a statistical benchmark. DetailsTwo real tasks, with and without Gio
| Task on Click (patches verified) | Without Gio | With Gio | Reduction |
|---|---|---|---|
| A — add a “did you mean?” hint to an error both fixes correct | 1.15M ctx tokens · $0.52 · 27 turns | 636K · $0.33 · 16 turns | −45% tokens · −37% cost |
| B — guard an invalid option combination Gio matched the spec; the plain run reinterpreted it | 1.37M ctx tokens · $0.87 · 28 turns | 147K · $0.11 · 4 turns | −89% tokens · −88% cost |
“ctx tokens” is the input plus cached context the model processes each turn, read from Claude Code’s own logs. Savings grow with repo size; on tiny repos the gap nearly vanishes. For third-party-scored numbers across many issues there is a paired SWE-Bench Pro harness in bench/.
The dashboard /gio-benchmark writes for your repo
This is the actual file the Click run produced. Run it on your own code and screenshot yours.
How it works
Gio is a playbook Claude follows plus a few local Python scripts. No model calls of its own, no daemon, no account.
Locate before you read
A local index (BM25, optional on-device embeddings) turns “where is login handled” into exact file:line spans. Claude reads those lines, not the file.
Delegate the sweeps
Broad searches run in helper agents that return only the conclusion, so your main context stays small and every later turn is cheaper.
Keep the project sane
A DECISIONS.md log so choices aren’t re-argued, reuse-over-duplicate habits, and a six-phase review before anything merges.
Private by construction
Indexes live in your repo. The impact calculator reads the usage logs Claude Code already keeps on your disk. The only network access is an optional one-time embedding model download.
See what you saved
Ask “how much have I saved?” and Gio reports real spend from your logs plus an estimated saving versus a whole-file workflow, clearly labeled as an estimate.
When the index is wrong
Retrieval is a pointer, never the source of truth. Gio reads the span it found to confirm it before acting; if the code moved or the hit looks off, it falls back to grep and whole-file reads and rebuilds the index. You can override at any time by telling Claude to read the file. Gio is a playbook, not a gate.
MIT licensed
Fork it, ship it inside your team, read every line. Issues and pull requests are open on GitHub.
Gio for teams is next
One savings dashboard across every developer on your Claude bill, shared indexes for big monorepos, and a savings report your manager can read. Leave an email and you’ll get one message when it ships. Privacy.