Less noise. More signal. Right signal. Reduce AI token consumption by up to 90%.
The Semantic Signal Engine that cuts AI token consumption by up to 90%.
OMNI acts as a context-aware terminal interceptor—distilling noisy command outputs in real-time into high-density intelligence, ensuring your LLM agents work with meaning, not text waste.
When Claude runs git diff, cargo test, or kubectl get pods, it receives everything — every "Compiling..." line, every cached Docker layer, every passing test. This noise consumes tokens, dilutes reasoning, and slows your agent down.
Most filters work by rules: strip lines matching patterns. OMNI works by meaning:
- It understands that
error[E0432]matters more thanCompiling foo v0.1.0 - It knows your session — boosting files you're actively editing
- It never drops — compressed content is always retrievable via
omni_retrieve()
Reduce up to 90% AI Token Usage
Zero Information Loss. Maximum Agent Context. <2ms Overhead.
- Classify — OMNI identifies the content type (git diff, build output, test results, logs) by analyzing structure, not filenames
- Score — Each line gets a semantic relevance score based on signal tier (critical → noise) and your current session context
- Compose — High-signal content is kept, noise is removed, and anything dropped is stored in RewindStore for retrieval
All of this happens transparently — your AI agent doesn't know OMNI exists. It just gets better signal.
# Install
brew tap fajarhide/omni
brew install omni
# Setup Claude Code hooks (one-time)
omni init --hook
# Verify
omni doctor
# View token savings after your first session
omni statsOr install via script:
curl -fsSL https://raw.githubusercontent.com/fajarhide/omni/main/scripts/install.sh | sh
omni init --hook| Content Type | Example Command | Typical Reduction | What's Kept |
|---|---|---|---|
| Git Diff | git diff HEAD~3 |
56-80% | File tree, changed hunks, +/- lines |
| Git Status | git status |
70-80% | Branch, staged/modified/untracked summary |
| Build Output | cargo build |
65-85% | Error count, error messages, warnings |
| Test Results | pytest, cargo test |
70-90% | Pass/fail count, failure details |
| Infrastructure | kubectl get pods |
60-75% | Pod status summary, non-running details |
| Docker | docker build . |
70-85% | Step count, cache hits, image ID |
| Logs | access logs, app logs | 0-40% | Errors, warnings, anomalies |
docker build . — 15 lines → 1 line (81% reduction)
# Before (what Claude sees without OMNI):
Step 1/15 : FROM node:18-alpine
---> a1b2c3d4e5f6
Step 2/15 : RUN npm install
---> Using cache
Step 3/15 : COPY . .
---> Running in 9876543210
...11 more steps...
Successfully built 9b9c9d09a123
# After (what Claude sees with OMNI):
docker: 5 steps | 1 cached | Successfully built 9b9c9d09a123
OMNI doesn't just compress — it understands your session context.
When you're debugging src/auth/mod.rs, OMNI:
- Boosts any output mentioning
auth/mod.rs(because it's a hot file) - Prioritizes errors matching patterns you've seen before
- Infers your task domain ("auth module") for smarter scoring
- Persists across compaction events, so Claude never loses context
This is powered by the SessionState engine that tracks hot files, recent commands, active errors, and domain hints — all stored in local SQLite.
When OMNI compresses aggressively, the original content isn't deleted — it's stored in the RewindStore with a SHA-256 hash:
[omni: 1,247 chars stored → omni_retrieve("a1b2c3d4")]
If Claude needs the full content, it simply calls omni_retrieve("a1b2c3d4") via MCP and gets everything back. Zero information loss, guaranteed.
Extend OMNI for your company's internal tools without writing code:
# ~/.omni/filters/deploy.toml
schema_version = 1
[filters.deploy]
description = "Company deploy tool"
match_command = "^deploy\\b"
strip_ansi = true
[[filters.deploy.match_output]]
pattern = "Deployment successful"
message = "deploy: ✓ success"
strip_lines_matching = ["^\\[DEBUG\\]", "^Waiting"]
max_lines = 30
[[tests.deploy]]
name = "strips debug lines"
input = """
[DEBUG] Connecting...
Deployment successful
"""
expected = "deploy: ✓ success"Test your filters: omni learn --verify
See docs/FILTERS.md for the complete filter writing guide.
$ omni stats
─────────────────────────────────────────────────
OMNI Signal Report — last 30 days
─────────────────────────────────────────────────
Commands processed: 1,247
Data Distilled: 18.4 MB → 3.2 MB
Signal Ratio: 82.6% reduction
Estimated Savings: $0.046 USD
Average Latency: 2.1ms
RewindStore: 23 archived / 8 retrieved
By Filter:
1. git 203x 89% ████████████████████
2. build 89x 82% ████████████████
3. test 44x 79% ███████████████
4. infra 31x 76% █████████████
Route Distribution:
Distill: 1247 (97%)
Keep: 25 ( 2%)
Drop: 12 ( 1%)
Passthrough: 0 ( 0%)
Session Insights:
Hot files: src/auth/mod.rs (12), tests/auth_test.rs (8)
───────────────────────────────────────────────── | Agent | Integration | Status |
|---|---|---|
| Claude Code | PostToolUse hook (automatic) | ✅ Full support |
| Any MCP client | MCP server (omni --mcp) |
✅ Full support |
| Shell pipe | command | omni |
✅ Works now |
| Command | Description |
|---|---|
omni init --hook |
Setup Claude Code hooks |
omni stats |
Token savings analytics |
omni session |
Session state inspection |
omni learn |
Auto-generate filters from passthrough |
omni doctor |
Diagnose installation |
omni version |
Print version |
omni help |
Show help |
cmd | omni |
Pipe mode — distil any command output |
omni --mcp |
Start MCP server |
omni --hook |
Hook mode (used by Claude Code) |
See docs/CLI_REFERENCE.md for full usage details.
flowchart TB
Agent["Claude Code / MCP Client"]
Agent -->|"PostToolUse hook / MCP"| OMNI
subgraph OMNI["OMNI — Semantic Signal Engine"]
direction LR
C["🔍 Classifier\n(10 content types)"]
S["📊 Scorer\n(signal tiers + context boost)"]
R["✂️ Composer\n(threshold + RewindStore)"]
C --> S --> R
end
subgraph Persistence["SessionState + SQLite"]
direction LR
HF["Hot Files"]
AE["Active Errors"]
DH["Domain Hints"]
RW["RewindStore"]
end
OMNI <--> Persistence
R -->|"distilled output"| Agent
R -->|"dropped content"| RW
To ensure your code meets all quality standards before pushing to the repository, run the comprehensive CI pipeline locally:
make ci # Run fmt, clippy, tests, security audit, and binary size checksFor individual checks during development:
cargo build # Build the binary
cargo test # Run all 147 tests
cargo insta review # Review and accept snapshot changesSee CLAUDE.md and DEVELOPER.md for the full contributor guide.
MIT