Let a small model do the reading.

Ganglio is a plugin for Claude Code. A cheap model searches your repository and hands Claude the files and the exact lines, so Claude reads less and costs less.

Sign in with GoogleSee the numbers →

$2.00 to start, no card. Claude Code today; other agents next.

Claude thinks. Ganglio reads.

  1. 01Claude asks where something is.
  2. 02A cheap model searches the repository.
  3. 03Claude gets the files and the exact lines.
  4. Nothing replaces Claude. It just reads less. And when you come back to a long conversation after a break, a handoff note lets you start fresh instead of paying for the whole conversation again.

Measured, not promised.

The same tasks in Claude Code with and without the plugin. Where it didn’t pay off, we say so.

Sonnet · the scout as a tool
−38%Claude’s cost
95% interval −45% to −29%
0.76 vs 0.76right files found
50 tasks × 2 runs
4 turns instead of 7
Opus · the scout runs first
−43%Claude’s cost
95% interval −59% to −18%
0.92 vs 0.99right files found: slightly fewer
10 tasks × 2 runs
2.9 turns instead of 5.5
Cutting tool outputs
−3%Claude’s cost: no saving
95% interval −9% to +3%
+4 son each output
50 tasks × 3 runs, Sonnet
so it is switched off

Baseline: Claude Code as it is, on the same find-the-files tasks from open-source repositories. Cost is Claude’s at API prices, before Ganglio’s own price: on these short tasks a scout call saved about $0.036 on Sonnet and costs $0.01. How it was measured →

Know the limits

Four steps.

  1. 1Sign in with Google
  2. 2Make a key
  3. 3Install the plugin in Claude Code
  4. 4Work as usual

Works in Claude Code today; Codex CLI, Gemini CLI and Cursor are next. What leaves your machine is answered before you connect.

In Claude Code
/plugin marketplace add https://ganglio.network/marketplace
/plugin install ganglio@ganglio
! ganglio check
→ balance $2.00, the small models answer

Pay for what it reads.

Prepaid: you top up a balance and calls are paid from it. We never charge more than you put in, and nothing renews by itself. What you paid and did not spend can be refunded.

To start$2.00on your balance when you sign in. That is 200 scout calls or 8 summaries. No card.Sign in with Google
Then, per call
Scout call
$0.01
Summary on return
$0.25
On your machine
free
A scout call is one question from Claude, however many rounds of searching it takes. Someone in Claude Code all day spends roughly up to $15.00 a month — an estimate, not a limit.
EnterpriseLet’s talk.
The small model runs in your VPC or on-prem, so your code stays inside.Contact us

Your code. Your data.

What leaves my machine?
  • A scout call. Claude’s question and the parts of the repository the small model reads to answer it.
  • A summary. The conversation of a session that went idle, so the note can be written.
  • Tool outputs. Nothing. Repeats and folded logs are handled on your machine.

It goes to Ganglio, which passes it to the provider that runs the small model. We keep: no request or answer text. For each call: the account, the kind, the model, the token counts and the time.

Who pays for Claude?

You, exactly as before. Claude Code talks to Anthropic with your own subscription or key, and Ganglio never sees either. Your Ganglio balance pays only for the small model’s work.

What happens when the balance runs out?

The scout tells Claude to search by itself and no summary is written. Claude Code works as it did without the plugin. Nothing is charged later.

How was it measured?

In real Claude Code sessions: the same tasks with and without the plugin, two or three runs each, Claude’s cost at API prices, 95% intervals. The summary’s saving is different: it is computed by replaying one heavy user’s recorded sessions, not by a live test. On your own work, run /ganglio:stats in Claude Code.