Squeeze your agent's token bill.
The #1 tool to cut what your AI agent sends to the model.
One install, zero config. Tokenade compacts command output, MCP results, web searches and stale context before the model reads a single line.
100% local · Tokenade never uploads your code or your conversations.
Free & open source (MIT/Apache-2.0) · Linux, macOS, Windows
Running in 30 seconds
2 commandsnpm install -g @tokenade/clitokenade install- Detects every agent on the machine and hooks into each one
- Works on every prompt from then on — nothing to configure
- `tokenade dashboard` shows every token saved, per feature — in a local report
What gets squeezed
10 layers · automaticNoisy command output, compacted
Builds, tests, logs — every noisy output trimmed to what matters before the model reads a line. Failures, numbers and decisive lines always survive.
Your MCP servers, on a token diet
One command wraps every MCP server. What they return is compacted before it reaches the model.
Web search, rebuilt for agents
Same answers as compact result lists instead of walls of rendered text — with automatic fallback to the native tool.
A local search that understands your code
Semantic search, repo map, symbol query, file skeletons, impact analysis — your agent stops reading whole files to find one function.
Models that skip the small talk
No pleasantries, recaps or filler — every technical detail kept. Output tokens are the most expensive ones.
A context window managed like a budget
Old context compacted before the prompt cache expires and re-bills it. Your agent nudged to batch independent commands.
Never pay for the same bytes twice
Identical re-reads become a tiny reference — or just the diff when the content changed.
Your agent reads the whole repo, not just the code
PDF specs, spreadsheets, recorded meetings — extracted properly instead of improvised at full price.
Your code never leaves your machine
100% local: your code, prompts and file paths never leave your machine, and the source is open for you to audit.
Tokenade stops secret leaks
Secrets are redacted before they ever reach your model — or its provider.
Works where you work
measured · 3 OSes100% compatible
80%+ compatible
Weakly compatible
Measured, not claimed
THOLon the Token-Harness Optimizer Leaderboard
An end-to-end benchmark of agent token optimizers, run on real coding sessions. Set Tokenade up once — it works on every prompt.
