Open source · Free forever

Your agent rereads the same files.
You are paying for it.

DRIP sits between your AI agent and the filesystem — replacing every re-read with a delta or an unchanged sentinel. Same workflow, 34–88 % fewer tokens.

Install DRIP See the math

Single Rust binary. SQLite locally. No daemon, no network call, no telemetry.

fewer tokens: 0–0 %
agents wired: 0
telemetry sent: 100 bytes

~ /you/projectdrip 0.1.0

→ 400 lines (2,403 tok, semantic-compressed)

→ [DRIP: unchanged]

→ +3 lines

→ [DRIP: delta | +3 −0 | 280 B]

DRIP Meter (Since Install)

──────────────────────────────────

Tokens saved:102.8K

Reduction:87 %

$ saved:$0.31

████████████████████▒▒▒▒ 87 %

Drop-in for the agents you already use

Claude Code

Codex CLI

Gemini CLI

MCP

Apache 2.0

Rust

Open source

Claude Code

Codex CLI

Gemini CLI

MCP

Apache 2.0

Rust

Open source

Claude Code

Codex CLI

Gemini CLI

MCP

Apache 2.0

Rust

Open source

Claude Code

Codex CLI

Gemini CLI

MCP

Apache 2.0

Rust

Open source

The problem

Coding agents bleed tokens on file reads

Your model rereads the same file five times in a session. You pay for it five times.

5×

Same file, again and again

A 400-line app.py read five times = 2 000 lines sent to the model — even though 397 of them never changed.

No visibility on what costs

Pricing dashboards show monthly invoices. They don't tell you which file your agent re-reads twenty times an hour.

↺

Context limits hit faster

The more redundant input, the sooner Claude / Codex compact. Compactions throw away DRIP's baselines and force re-sends.

Before / after

Same workflow. 13 % of the input.

DRIP intercepts file-read calls and substitutes a delta or an `unchanged` sentinel. Same agent, same tools, dramatically smaller payloads.

Without DRIP

30,025 tok

read app.py     → 400 lines  (6,005 tok)
read app.py     → 400 lines  (6,005 tok)
read app.py     → 400 lines  (6,005 tok)
read app.py     → 400 lines  (6,005 tok)
read app.py     → 400 lines  (6,005 tok)
                              ─────────
                              30,025 tok

With DRIP

2,403 tok92 % saved

read app.py     → 400 lines  (2,403 tok, semantic-compressed)
read app.py     → [DRIP: unchanged]
read app.py     → [DRIP: unchanged]
read app.py     → [DRIP: unchanged]
read app.py     → [DRIP: unchanged]
                              ─────────
                              2,403 tok

See how DRIP intercepts this →

Under the hood

A single Rust binary that pays for itself in a session

Delta on every re-read

First read sends the file. Every subsequent read becomes a unified diff via `similar`, an `[unchanged]` sentinel, or — for big diffs — a fallback to full content.

Semantic compression

13 languages recognised. Function bodies elided on first reads, signatures + imports + class declarations preserved. Verified zero structural loss on the 8-language benchmark fixtures.

Partial-read awareness

`Read(file, offset, limit)` is window-scoped: DRIP diffs only the requested range, returns an `unchanged (lines X-Y)` sentinel when the window is byte-identical, and never mutates the file's full-content baseline.

Edit certificates

Read a file right after editing it and DRIP returns a 390-byte attestation (file hash + touched lines + symbol names) instead of letting the agent re-read 600 lines it just wrote.

Context-compaction ledger

Schema v9 tracks how often Claude / Codex compact, how many tokens were re-sent, and decorates first reads with `↺ context was compacted` so the agent knows why.

DripMeter for macOS (optional)

Companion menu-bar app surfaces your savings live: per-agent breakdown, top files, sparkline, real-time activity feed. Reads DRIP's local SQLite, no network.

Read the architecture →

One init per agent

Three agents. One CLI.

DRIP wires into Claude Code via Pre/PostToolUse hooks and into Codex/Gemini via MCP. One command per agent — idempotent, non-destructive.

Claude Code

PreToolUse / PostToolUse hooks

$drip init -g

Global (~/.claude/)

Codex CLI

MCP server (read_file)

$drip init --agent codex

~/.codex/config.toml

Gemini CLI

MCP server (read_file)

$drip init -g --agent gemini

Global (~/.gemini/)

Companion app · macOS

Watch your savings build up — live, in your menu bar.

DripMeter is the macOS companion app for DRIP. It reads the local SQLite (no telemetry) and surfaces your savings in real time: per agent, per file, per cost model. One click and it's running.

100% on-device
macOS 14+
Swift 6.2
Zero network
Open source

DripMeterUpdated 2m ago

108.2Ktokens saved

Saved87 %

Money$0.31

CO₂28 g

Per agent

Claude Code62 K

12 sessions · 5m ago94%

Codex28 K

3 sessions · 1h ago71%

Gemini0

Hooks installed · waiting0%

Cost projection

$249/ month at Opus 4.7

Live menu-bar meter

A tiny green droplet icon that fills proportionally to your % saved. Click to expand.

Per-agent breakdown

Claude Code, Codex, Gemini — see where your savings come from, with their official logos and install-detection badges.

Cost projection

Pick Sonnet 4.6, Opus 4.7, GPT-5 or Gemini 2.5 — DripMeter extrapolates your savings per day, week, month, year.

Markdown usage report

One click exports a portable .md with lifetime totals, streak, 7d/30d tables and per-agent breakdown.

Compaction tracker

Surfaces DRIP's v9 ledger: how often your agent compacts, how many tokens were re-sent, with a 'quality' score.

Real-time activity feed

Watch DRIP intercept reads as they happen, FSEvents-driven. Click any row to open the file in your editor.

$brew install --cask drip-cli/drip/dripmeter

Download for macOS DripMeter on GitHub

And there's more

Three pieces of the Overview tab — each one's a live SwiftUI view in the real app.

12-day streak

Streak panel

Pulsing flame, week strip, target progress.

Today12-day streak

18.2Ksaved · target 25K

31.5K avg / day

Rollup tiles

7d / 30d toggle, daily average, peak day.

Rollup

7d30d

Saved

220K

Avg/d

31.5K

Peak

80K

Active

6/7

73 active days

Activity heatmap

90 days, log-bucketed intensity, today ringed.

Activity~3 months

lessmore

Install

One command per platform.

DRIP is a single Rust binary. macOS via Homebrew, Linux via the install script (musl-static auto-detected), Windows via the prebuilt zip, or `cargo install drip-cli` anywhere.

1
Install the binary
Single 4 MB Rust binary. macOS Apple Silicon + Intel, Linux x86_64 + arm64, Windows x86_64.
$brew install drip-cli/drip/drip
Want the menu-bar app too? brew install --cask drip-cli/drip/dripmeter
2
Wire it into your agent
Idempotent and non-destructive — drip init writes hooks via tmp-file + rename(2), never overwrites your existing CLAUDE.md or settings.json.
$drip init -g
3
Watch the savings
drip meter prints the human-readable savings since install. Or pipe --json into your own dashboard.
$drip meter

FAQ

Questions, honest answers

No. Single Rust binary, single SQLite database (~/Library/Application Support/drip/sessions.db on macOS). No daemon, no network call, no telemetry. Verified on every release with `cargo machete` plus a manual audit of the dependency graph.

Measured 34 % to 88 % on an 8-language benchmark fixture (8 languages × 4 workflow styles), with verified zero structural loss in semantic compression. See BENCHMARKS.md in the repo for the full table.

Claude Code (PreToolUse / PostToolUse hooks), Codex CLI (MCP server), Gemini CLI (MCP server). Cursor's built-in read tool bypasses MCP and is not on the roadmap — DRIP would only see calls the model deliberately routed through it, which isn't the contract the other agents get.

No. drip init writes via tmp-file + rename(2), preserves Unix mode bits, never touches hand-edited content in CLAUDE.md or settings.json. drip uninstall reverses every byte DRIP wrote — pre-existing hooks and unrelated MCP servers stay intact.

macOS arm64 + x86_64, Linux x86_64 + arm64, WSL, Windows x86_64. The Rust toolchain ≥ 1.95 builds it from source on any platform that has cargo.

drip meter, drip sessions, drip replay are the CLI surfaces. For macOS users, DripMeter is a companion menu-bar app that reads the same SQLite and surfaces savings live, with per-agent breakdown and click-through to your IDE.

Ask on GitHub →

The team

Built by Perform Code, in Lille.

Three engineers shipping the boring infrastructure that AI agents need to stop bleeding tokens. Open source by default, no VC ask, no upsell.

Maxence Bombeeck

Founder · Designer & Swift

Designed the brand and built DripMeter. Owns the macOS Swift surface end-to-end.

@MaxenceB59

Hugo Barbosa

Founder · Core DRIP

Wrote the DRIP Rust core: session model, semantic compression, delta engine, agent hooks.

@Hugobrbs

Hugo Ponthieux

Founder · DevOps & Infra

Owns release pipelines, benchmarks, infra and distribution channels.

@Hugoy8

Operated by Perform Code SAS

Stop paying for the same file twice.

Single Rust binary, zero telemetry. DripMeter keeps the score in your menu bar. Save 34–88 % of input tokens from session one.

Install DRIP Star on GitHub

Your agent rereads the same files.You are paying for it.

Coding agents bleed tokens on file reads

Same file, again and again

No visibility on what costs

Context limits hit faster

Same workflow. 13 % of the input.

A single Rust binary that pays for itself in a session

Delta on every re-read

Semantic compression

Partial-read awareness

Edit certificates

Context-compaction ledger

DripMeter for macOS (optional)

Three agents. One CLI.

Watch your savings build up — live, in your menu bar.

Live menu-bar meter

Per-agent breakdown

Cost projection

Markdown usage report

Compaction tracker

Real-time activity feed

Streak panel

Rollup tiles

Activity heatmap

One command per platform.

Install the binary

Wire it into your agent

Watch the savings

Questions, honest answers

Built by Perform Code, in Lille.

Maxence Bombeeck

Hugo Barbosa

Hugo Ponthieux

Stop paying for the same file twice.

Your agent rereads the same files.
You are paying for it.