Packet28 — Your AI Agent Reads Too Much

What it does

READCompresses file reads

Your agent tries to read a 9,243-line file. Packet28 intercepts it and returns only the ~40 lines that matter for the current task — plus anchors back to the full file. No wasted tokens on irrelevant code.

without packet289,243 lines

The whole file enters the context window. Most of it is never referenced again.

with packet28~40 lines

Only the regions the task touches, with file and symbol anchors to fetch more on demand.

MAPMaps your repo

Builds a structural map of your codebase — files, symbols, dependencies — so the agent navigates directly to relevant code instead of exploring blindly.

MEMRemembers across sessions

A background daemon persists context between agent restarts. Pick up where you left off without the agent re-reading your entire project from scratch.

SETOne command setup

Auto-detects Claude Code, Cursor, Codex, Copilot, or Windsurf and configures everything. One command: packet28 setup --root .

How it works

A Rust workspace of 28 crates in five layers. Reducers compress artifacts, the runtime decides what the agent sees, and every output carries the same envelope.

Agent surface

packet28-agent wrapper · agent-prompt generator · MCP surface

CLI + daemon

Packet28 CLI · packet28d daemon · task / watch / stream protocol

Context runtime

kernel · scheduler · memory & recall · assembly · correlation · policy & guard · agent state

Reducers

diffy · covy · testy · stacky · buildy · mapy · proxy

Shared contracts

EnvelopeV1 · BudgetCost · FileRef / SymbolRef · Provenance

When it can’t compress safely, it gets out of the way

A command that policy excludes, or that can’t be rewritten safely, passes through raw.
Unsupported reducer families run raw, and the fallback reason is recorded in the tool result.
If pre-tool rewriting isn’t available for your runtime, activity is captured post-tool instead.
If the daemon or preferred search backend is down, reducers and search run locally.

The numbers

Real benchmark on a 12MB Java codebase (534 files).

WITHOUT PACKET28

139,000 tokens

WITH PACKET28

849

164x

fewer tokens consumed

Task: "Understand StringUtils.wrap and fix null delimiter handling"
Naive agent reads entire 9,243-line files. Packet28 serves ~40 relevant lines + anchors.
Bars are not to scale — at true ratio the Packet28 bar is 0.6% wide. Large-file reads are the best case; across a full session the typical result is ~90% less search overhead and about 3x more work per context window.