fajarhide/omni

What it solves

OMNI reduces the amount of redundant and noisy data sent to AI agents, lowering token costs and preventing context window saturation. It specifically targets the problem of agents re-reading the same files or seeing repetitive build logs and progress bars that provide little value to the model.

How it works

OMNI acts as a set of hooks around an agent's tool calls. It processes the output of these commands before the agent reads it using two primary mechanisms:

1. Distillation: It removes "noise" such as Docker layer hashes, ANSI colors, and progress bars from build logs.
2. Ledger-based Deduplication: If a piece of output (like a file's contents) has been seen by the agent earlier in the session, OMNI replaces the repeated bytes with a small marker and a retrieval handle.

All removed or folded content is archived in a local SQLite database (RewindStore), allowing the agent to retrieve the full, byte-for-byte original output using the omni retrieve command or an MCP tool.

Who it’s for

Developers using AI agents (such as Claude Code, Cursor, Windsurf, or Aider) who want to optimize their context window and reduce costs associated with repetitive tool output.

Highlights

• Lossless Compression: Nothing is permanently deleted; all original data is stored locally and can be retrieved via handles.
• Context-Aware Folding: When a file changes, OMNI folds the parts already seen while keeping the line numbers accurate for the agent.
• Zero Fabrication: It never summarizes or invents results; if it cannot safely distill a output, it passes the raw output through verbatim.
• Broad Compatibility: Supports various hosts via direct hooks, handoff-first configurations, or the Model Context Protocol (MCP).