omni

Less noise. More signal. Cut your AI token consumption by up to 90%.

🇺🇸 English | 🇯🇵 日本語 | 🇨🇳 简体中文 | 🇸🇦 العربية | 🇮🇩 Bahasa Indonesia | 🇻🇳 Tiếng Việt | 🇰🇷 한국어

OMNI is a smart terminal layer that intelligently filters and prioritizes command output before it reaches your AI agent. By preventing your AI from getting confused by noisy output, you get accurate answers faster while saving massive amounts of token costs.

Fully transparent. You're always in control.

Table of Contents

• The Problem: Expensive Tokens & Noisy Outputs
• The Solution: Omni
• The Philosophy
• Performance & Use Cases
• Features Explained
• Architecture
• Quick Start & Installation
• How to Use It
  • Multi-Agent Support & Integrations
  • Documentation Index
• Works Even Better with Heimsense
• Contributing & License

The Problem: Expensive Tokens & Noisy Outputs

When you use autonomous AI agents (like Claude Code) in your terminal, they read everything. A simple git diff, npm install, or cargo test command can easily dump 10,000 to 25,000 tokens of useless terminal noise into your AI's context.

This causes three huge problems:

1. It's extremely expensive: You pay real money for every single token of that junk output.
2. It makes the AI "dumb": Critical errors get buried under megabytes of warning logs and loading bars, confusing the AI and diluting its reasoning.
3. Model Lock-in: Advanced agent frameworks force you to use their most expensive flagship models just to have a context window big enough to handle all that noise.

The Solution: Omni

I built Omni because I wanted to run AI agents efficiently and cheaply every single day in my own workflow.

Omni acts as the perfect filter between your terminal and your AI.

The result? You can run your AI agent on a super-advanced framework and feed it zero noise. Because the AI is only fed highly focused, straight-to-the-point context, even affordable or ordinary models will perform on-par with expensive flagship models, since they are never distracted by junk data.

My ultimate passion isn't to monetize this—it's to build the ultimate open-source toolbelt for the Agentic AI era. By aggressively saving token costs, I can develop software robustly and cost-effectively today, and you can too.

The Philosophy

OMNI wasn't built just to "cut context" or "save tokens"—those are simply the happy side effects. The true philosophy behind OMNI is Context Quality.

AI agents like Claude are only as smart as the context you feed them. When you flood them with megabytes of dependency logs or loading bars, you force them to sift through garbage to find the actual problem. This dilutes their reasoning and leads to degraded or unhelpful responses.

OMNI's goal is to feed your AI pure, highly-dense signal. This means only grabbing the context that is actually important and meaningful for Claude. We clean up the noise the AI doesn't need, which means:

1. Automatically, the tokens you use are drastically fewer.
2. The AI's response is of significantly higher quality because its context window is laser-focused on the real problem.

Try it for a week. Feel the difference in the quality and speed of your AI's reasoning when it's fed on a diet of pure signal instead of raw terminal noise.

Performance & Use Cases

OMNI is built in Rust for zero-overhead execution and ruthless efficiency. Here are the actual benchmarks measured on the release binary:

| Command / Context | Input Size | Output Size | Token Savings | Impact on AI | |-------------------|------------|-------------|---------------|--------------| | docker build (multi-stage) | 9.2 KB | 49 bytes | 99.5% | Eliminates caching noise; AI instantly sees the real build error. | | cargo test (large suite) | 16.5 KB | 4.3 KB | 78.0% | Strips hundreds of "ok" tests; AI focuses only on the failures and stack traces. | | git status (dirty) | 496 bytes | 113 bytes | 77.2% | Removes clean files and hints; keeps only modified/untracked files. | | kubectl get pods | 840 bytes | 762 bytes | 10.0% | Selectively surfaces CrashLoopBackOff/Error pods, skipping healthy ones. | | git diff (multi-file) | 397 bytes | 220 bytes | 50.0% | Preserves hunks with changes, dropping excessive context lines. |

• Pipeline Latency: < 100ms (end-to-end, including binary startup)
• All-Time Savings: 97.3% token reduction across average development sessions.
• ROI: $35+ USD saved per developer/month (measured against flagship models).

To see your own actual token savings, just run omni stats after a few days of usage.

Features Explained

• No More AI Confusion: Omni acts like a smart sieve. If a test fails, it shows the AI only the specific error line and stack trace. Your AI stops getting distracted by loading spinners or noisy dependency logs, allowing it to focus directly on the real problem.
• 90% Token Reduction: By completely eliminating useless terminal noise, you drastically cut your agentic API bills instantly.
• Zero Information Loss: Worried Omni filtered something important? Don't be. Omni saves the raw output in a local archive (RewindStore). If the AI actually needs the full log, it can just automatically ask for it using omni_retrieve.
• Session Intelligence: Omni remembers what you are doing. It knows which files you are actively editing and stops feeding the AI context it already knows. Cross-session memory is now capable of preserving specific fixes permanently via omni_knowledge.
• Multi-Agent Collaboration: Omni is fully aware of its environment via omni_agents. If you have Cursor running alongside Claude CLI, they can seamlessly share the same filtered memory streams, active errors, and execution environments without clashing.
• Distill Monitor: Track your token savings and costs over time. Use omni_budget and omni_history right inside your LLM, or run omni stats locally to visualize your money saved.
• Visual Impact (omni diff): See exactly how much money and space you are saving. Just run omni diff to see the bulky raw output compared side-by-side to Omni's sleek, filtered version.
• Lightweight Dependency Graph: OMNI builds a fast local file relationship graph at hook time (no daemon, no LSP). When your AI reads a heavily-imported file, OMNI warns it: "this file has 12 dependents — call omni_context for full impact map.".
• Adaptive Compression: OMNI tracks when agents retrieve omitted output. If a command family is frequently retrieved, OMNI automatically softens compression next time — self-tuning without configuration.
• Smart High-Speed Bypass: To ensure zero latency for small tasks, OMNI automatically bypasses distillation for outputs under a 2000-token threshold. This prioritizes speed while still capturing the big data when it matters.
• Omission Visibility: OMNI now explicitly labels removed content (e.g., [OMNI: omitted X lines of noise]) in the output, giving your AI agent better situational awareness of what was filtered.
• Debug Passthrough: Need to see the raw output for a moment? Just set OMNI_PASSTHROUGH=1 in your environment to completely bypass the engine and see every single character of the original output.
• Structured ReadFile + Grep: Instead of raw file dumps or flat grep output, OMNI returns structured outlines (imports, public API, risk markers) and grouped grep summaries (top files by match count, priority lines first).
• Factual Anti-Hallucination Guards: OMNI emits warnings only when it has hard facts — no speculation. If output is heavily compressed and no rewind exists: it says so. If a file has many dependents: it says so. Keeping your AI grounded in reality.

Architecture

flowchart TB
    Agent["Claude Code / OpenClaw / Hermes Agent / MCP Agent"]

    subgraph Hooks["Native Hook Layer (Transparent)"]
        Pre["Pre-Hook\n(Rewriter)"]
        Post["Post-Hook\n(Distiller)"]
        Sess["Session-Start\n(Context)"]
        Comp["Pre-Compact\n(Summary)"]
    end

    Agent --> Pre
    Pre -->|"omni exec"| Output["Raw Stream"]
    Output --> Post
    Post --> Agent

    subgraph OMNI_Engine["OMNI — Semantic Signal Engine"]
        direction LR
        R["Registry\n(Filters)"]
        S["Scorer\n(Context Boost)"]
        D["Distiller\n(Semantic Magic)"]
        R --> S --> D
    end

    Post --> OMNI_Engine
    Pre --> OMNI_Engine

    subgraph Persistence["Persistence Store (SQLite)"]
        ST["Sessio