Redcon

Deterministic context budgeting for AI coding agents

Stop sending agents 200k tokens of irrelevant code. Redcon scores, compresses, and packs repo context so your agent gets what it actually needs.

Install - Quick Start - How It Works - Docs

The Problem

AI coding agents burn tokens on irrelevant context. You either:

• Dump the whole repo and pay for 200k input tokens per request, or

• Let the agent grep blindly and waste tool calls figuring out where to look

Redcon solves both. It ranks files by task relevance, compresses them with language-aware strategies (full, snippet, symbol extraction, summary), and packs the result under your token budget. Deterministic, local-first, no embeddings. One MCP server covers Claude Code, Cursor, Windsurf, Cline and Zed; a plain CLI covers CI. Measured on this repository it cuts input tokens by more than 83% at the same task coverage (methodology).

Related MCP server: project-graph-mcp

Punch above your plan

On a flat subscription - Claude Pro/Max, Cursor, GitHub Copilot - the token bill isn't what stings, the usage limit is. Redcon cuts the tokens each task needs, so the same plan covers far more work before you hit the weekly wall. Same subscription, more runway.

It stretches the budget you already pay for, and reports exactly how much it saved.

Install

Option 1: VS Code Extension (easiest)

1. Install Redcon - Context Budget from the marketplace

2. Open the Redcon sidebar, click Install & Set Up

3. Reload window. Done.

The extension installs the CLI via pip, registers the MCP server for Claude Code, Cursor, and Windsurf, and gives you a sidebar with budget analytics, file rankings, and compression dashboards.

Option 2: CLI + MCP Server

pip install "redcon[mcp]"
redcon init                      # creates redcon.toml + registers MCP

The init command auto-configures MCP for Claude Code, Cursor and Windsurf, plus VS Code, Codex CLI and Gemini CLI when they are detected, so your AI agent can call redcon_rank, redcon_search, redcon_compress, and redcon_budget as native tools. It also writes a short AGENTS.md section that tells agents to prefer these tools for context selection.

Option 3: CLI only

pip install redcon
redcon init --no-mcp

Quick Start

# Rank files relevant to a task
redcon plan "add rate limiting to auth API" --repo .

# Pack context under a token budget
redcon pack "refactor payment flow" --repo . --max-tokens 30000

# Compare compression strategies
redcon benchmark "add caching" --repo .

# Audit a PR for context growth
redcon pr-audit --repo . --base origin/main --head HEAD

Output goes to run.json (machine-readable) and run.md (human-readable). Use them in CI, or feed the compressed context directly into your agent.

How It Works

task: "add rate limiting to auth"
       |
       v
  [1] scan    - incremental scan of repo files (cached)
       |
       v
  [2] rank    - score each file: keyword match, imports, file role, git history
       |
       v
  [3] compress - per-file strategy: full / snippet / symbol extraction / summary
       |
       v
  [4] pack    - fit top-N compressed files under token budget, drop the rest
       |
       v
  run.json + run.md + compressed_context ready for your agent

Every step is deterministic. Same input, same output. No embeddings, no random chunking.

Benchmark: context-eval

Selection quality is measured, not claimed. context-eval/ is an open benchmark for context-selection tools: tasks come from real git commits, ground truth is the files each commit actually modified, and every tool packs the same token budget. Current results (33 tasks, 24k budget):

Rerun it on any repo: python context-eval/run.py --repo /path/to/repo. Methodology, limitations, and how to add your own tool: context-eval/README.md.

MCP Integration (Pull Model)

mcp-name: io.github.natiixnt/redcon

Instead of pushing a 30k-token blob to your agent, Redcon exposes 9 MCP tools the agent calls on demand:

Typical agent flow uses ~5k tokens for exploration instead of 30k for a blob. The agent itself decides what to read in full.

Config gets written automatically to:

• .mcp.json (Claude Code)

• .cursor/mcp.json (Cursor)

• ~/.codeium/windsurf/mcp_config.json (Windsurf)

Command Output Compression

Source files are only half the bloat. The other half is command output: git diff, pytest, cargo test, grep, ls -R. Redcon's redcon_run MCP tool (and redcon run CLI) wraps the call, parses the output, and returns a budget-aware compressed view that preserves every fact the agent actually needs.

Headline reductions on representative inputs:

Quality is enforced separately. Every compressor declares must_preserve_patterns (file paths in a diff, failing test names in pytest, branch name in git status, slowest node operator in EXPLAIN); the M8 quality harness rejects any compressor whose compact output drops a fact present in the raw input. Run it as a CI step:

redcon cmd-quality   # exits non-zero if any compressor regressed
redcon cmd-bench     # markdown table; --json for CI baselines
redcon run "git diff" --quality-floor compact --max-output-tokens 4000

Twenty compressors ship today: git_diff, git_status, git_log, pytest, cargo_test, npm_test (vitest+jest), go_test, grep, ls, tree, find, lint (ruff+mypy), docker, pkg_install (pip+npm+yarn), kubectl_get/kubectl_events, profiler (py-spy+perf), json_log, coverage, sql_explain (Postgres+MySQL TREE), bundle_stats (webpack + esbuild metafiles). Full per-schema benchmarks: docs/benchmarks/cmd/.

Cross-call dimension

Beyond per-call compression, four layers compose across an agent session:

• Path aliases (V41): repeated paths like redcon/cmd/pipeline.py collapse to f001 on later mentions. Lazy first-use, never net-negative.

• Content reference ledger (V43): paragraph-shaped blocks above 6 cl100k tokens get session-stable {ref:001} aliases on second-and-later occurrences. Empirically 23% of session output had block-level overlap.

• Symbol aliases (V49): CamelCase types / multi-word snake_case identifiers (>=8 chars) collapse to c001 aliases the same way paths do. Empirically 72% of distinct symbols recur >=2 times per session.

• Snapshot delta vs prior call (V47): when the same argv runs twice, ship only the delta. Schema-aware renderers for pytest (set-diff over failure names), git_diff (file-set with per-file +/- counts), and coverage (per-file pp moves) win meaningfully over generic line-diff. Always picks min(cost_delta, cost_abs) so non-regressive by construction.

• Invariant cert (V93): every COMPACT/VERBOSE output stamps mp_sha=<16hex> over the sorted multiset of (pattern, capture) extracted from raw. Auditors recompute the cert against the compressed text to detect spurious additions or capture thinning - upgrades the existing must-preserve boolean to set-equality.

Empirical measurement on 5 simulated agent sessions (benchmarks/measure_sessions.py): the cross-call layers add +8.3% session-level saving on top of the per-call compressors, with +15% on heavy-overlap sessions (debugging, search-and-edit) and near-zero on distinct-content sessions. V85 adversarial GA fuzzer ratchets all 20 schemas as a hard CI gate (REDCON_V85_ENFORCE=1).

VS Code Extension

Once installed you get:

• Sidebar chat: type a task, send, watch the pack run live

• Dashboard: donut/pie/bar charts for budget, strategies, token impact per file

• Status bar: current budget usage with risk indicator

• CodeLens: compression strategy and token count shown above each file

• File decorations: relevance score badges on files in the explorer

• History: browse past runs, diff them, export to clipboard

Branding: red->navy gradient with triple chevron mark, glass-style UI.

Workspaces (Multi-Repo)

One task can span multiple repos. Place the workspace TOML in a folder that contains all the repos (a monorepo root or a common parent directory) - repo paths must resolve inside that folder, so a workspace file cannot reach above itself:

name = "backend-services"

[scan]
include_globs = ["**/*.py", "**/*.ts"]

[budget]
max_tokens = 28000
top_files = 24

[[repos]]
label = "auth-service"
path = "auth-service"

[[repos]]
label = "billing-service"
path = "billing-service"
ignore_globs = ["tests/fixtures/**"]

Artifacts include workspace, scanned_repos, selected_repos, and repo-qualified paths like auth-service:src/auth.py.

See docs/workspace.md.

Python API

from redcon import RedconEngine

engine = RedconEngine()

# Rank files
plan = engine.plan(task="add user auth", repo=".", top_files=15)

# Pack context
result = engine.pack(
    task="add user auth",
    repo=".",
    max_tokens=30000,
    top_files=25,
)
print(f"Used {result['budget']['estimated_input_tokens']} of {result['max_tokens']} tokens")
print(f"Risk: {result['budget']['quality_risk_estimate']}")

for file in result["compressed_context"]:
    print(f"{file['path']}: {file['strategy']} ({file['compressed_tokens']} tokens)")

Full reference: docs/python-api.md.

Features

• Deterministic scoring: keyword match, import graph, file role (test/docs/prod), git history

• Language-aware compression: Python, TypeScript, JavaScript, Go, Rust, Java, and more

• Command output compression: 20 compressors covering git, test runners, grep/rg, listings, lint, docker, pkg-install, kubectl, profilers, JSON logs, coverage, SQL EXPLAIN, and bundle stats - 70-99% reduction at compact level

• Incremental scanning: cached file metadata with git-aware change detection

• Multi-repo workspaces: single task, multiple repos, shared config

• Budget policies: enforce max tokens, quality risk levels, file counts in CI

• Quality harness: must-preserve regex assertions per compressor, deterministic, robust to truncated/binary input

• Streaming runner: chunked Popen reader with bounded memory and early SIGTERM/SIGKILL when output cap is hit

• Run history: SQLite-backed artifact store for both file packs and command runs, diff/heatmap/drift analysis

• Cost analysis: estimate token costs across GPT-4o, Claude, and other models

• PR auditing: detect context growth in pull requests

• Plugin system: custom scorers, compressors, token estimators, summarizers

• Cache backends: in-memory, local file, Redis

• Doctor command: diagnose environment, Python version, disk space, git availability

Documentation

• Getting Started - first pack in 60 seconds

• CLI Reference - all commands and flags

• MCP and Hooks - agent integration, tool list, deterministic context injection

• Configuration - redcon.toml fields

• Workspaces - multi-repo setup

• Python API - programmatic usage

• Agent Integration - middleware layer

• Plugins - custom extensions

• Architecture - how it all fits together

• Migration Notes - upgrading between versions

License

Everything in this repository is licensed under the Functional Source License, FSL-1.1-MIT: use it freely for any purpose except building a competing commercial product or service. Each release automatically becomes MIT-licensed two years after it ships.

The hosted Redcon Cloud control plane (orgs, quotas, usage metering, billing, team dashboards) is separate commercial software maintained in a private repository.

Commercial licensing: natjiks@gmail.com