AI Token Optimizer
An MCP server that cuts AI coding-agent token usage by 60–80% via framework-aware context optimization.
AI coding agents read whole files when they need 15 lines. They load *.g.dart,
R.java, node_modules/, and dist/ into context — then do it again next turn.
AI Token Optimizer is a Model Context Protocol
server that intercepts those tool calls and strips framework noise before it
reaches the model.
Works with anything that speaks MCP: Claude Code, Cursor, Cline, Continue, Zed, JetBrains AI, GitHub Copilot agent mode, Antigravity, Windsurf, OpenAI Codex CLI.
Quickstart
npm install -g @ai-optimizer/core
Then point your agent at the ai-optimizer stdio server. For Claude Code, add to
~/.claude/settings.json:
{
"mcpServers": {
"ai-optimizer": {
"command": "ai-optimizer"
}
}
}
Or scaffold the config and preview your savings automatically:
npx @ai-optimizer/init
# "we'd skip 47 files, ~14K tokens/session, ~$0.42/session at Sonnet pricing"
Full per-agent setup (Cursor, others), LSP install, and verification steps live in
packages/core/README.md.
What it does
It registers tools the agent uses instead of its built-in file/dir/error reads:
| Tool | What it saves |
|---|---|
optimized_read_file | Framework-aware skip + optional line slice |
optimized_list_files | Walks the tree applying ignore globs (+ .gitignore) |
optimized_grep | Content search scoped by ignore filters |
optimized_diagnostics | Spawns the right LSP, returns errors + a narrow code window |
read_symbol | LSP workspace/symbol slice instead of a full-file read (Pro/LSP) |
strip_bash_noise | Strips ANSI, npm/maven/gradle progress, JVM warnings |
optimizer_status | Detected frameworks, active packs, cumulative tokens saved |
When a project type is detected (e.g. Python via pyproject.toml), the matching
pack's ignore rules activate automatically. .gitignore is respected on top.
Tiers
| Free (MIT core) | Pro ($9/mo) | Team ($29/seat) | |
|---|---|---|---|
| Framework packs | Python, JS/TS | + React, Flutter, Java, Kotlin, Go, Rust, C#, Swift, Ruby, Elixir, PHP, Solidity | all Pro packs |
| History compaction | — | ✓ | ✓ |
Scheduler (claude -p cron) | — | ✓ | + work-stealing across providers |
| Audit log + secret redaction + policy | — | — | ✓ |
See PRICING.md. Behavior-changing or data-sharing features
(scheduler, work-stealing, telemetry) are off by default and require explicit
opt-in. Display-only features (dollar-value savings) default on.
Monorepo layout
| Package | Purpose | Published |
|---|---|---|
packages/core | MCP server + free packs | npm (MIT) |
packages/init | npx @ai-optimizer/init config + savings preview | npm |
packages/sdk | definePack plugin SDK | npm |
packages/context-mode-import | Import .contextmode configs | npm |
packages/pro | Pro/Team packs + scheduler + audit | private |
apps/landing, apps/dashboard, apps/vscode-extension, apps/license-server | Site, local dashboard, IDE widget, license issuance | private |
tools/pack-bench | Pack quality benchmark harness | private |
Architecture details: ARCHITECTURE.md.
Development
npm install # at repo root (workspaces)
npm run build # build all packages
npm test # run all test suites
Security
Local-only by design — license validation is the only network call. Report
vulnerabilities per SECURITY.md.
License
Core engine is MIT — see LICENSE. Pro/Team packs are proprietary.