Local AI gateway that sits between your IDE, agent, or app and your model provider. It shrinks prompts and completion budgets before tokens are billed — then forwards a normal OpenAI-compatible request upstream.
Works with Cursor, Cline, Continue, Open WebUI, custom scripts, or anything that speaks /v1/chat/completions. Upstream can be OpenRouter, OpenAI, Ollama, LM Studio, or any compatible API.
| Version | 0.1.1 |
| License | MIT |
| Default address | http://127.0.0.1:20128 |
| Platforms | Windows desktop app · Node.js 20+ (macOS / Linux via server) |
- Run the installer:
Token Savers-Setup-0.1.1.exe(build one withnpm run dist, or use a release build if you have it). - Launch Token Savers from the Start Menu or desktop shortcut.
- The dashboard opens in the app window. Minimize or close (X) keeps it running in the system tray — use Quit in the tray menu to exit fully.
Uninstall from Windows Settings → Apps or the Start Menu uninstall entry.
npm install
npm run serverOpen http://127.0.0.1:20128/.
| Command | What it does |
|---|---|
npm run server |
Proxy + dashboard in the terminal |
npm run dev |
Server + Electron window (Windows) |
npm run app |
Build, then run Electron |
npm run dist |
Windows NSIS installer → release\ |
If Electron fails to install: node scripts/fix-electron.mjs
- Open Upstream in the dashboard.
- Add a provider (name, base URL, API key, models). Examples:
- OpenRouter →
https://openrouter.ai/api/v1 - OpenAI →
https://api.openai.com/v1 - LM Studio →
http://127.0.0.1:1234/v1 - Ollama →
http://127.0.0.1:11434/v1
- OpenRouter →
- Set the default provider and default model.
- Open Endpoint and copy:
- Base URL:
http://127.0.0.1:20128/v1 - API key: the local
ts-…key (this is not your OpenRouter/OpenAI key)
- Base URL:
Point your IDE or agent at that local base URL and key. Upstream secrets stay only in Token Savers.
- OpenAI-compatible
GET /v1/modelsandPOST /v1/chat/completions(streaming and non-streaming) - Multiple upstream providers with per-provider API keys and model lists
- Local-only by default (
127.0.0.1) — traffic stays on your machine - Live Log of client → post-saver upstream → response (in-memory; secrets redacted)
- Usage estimates comparing tokens with vs without savers
Every request runs through a local compression pipeline before it hits the provider. Toggle and tune each saver on the Token Savers page.
Prompt size (messages)
| Saver | What it does |
|---|---|
| Clip | Drops old chat turns; keeps system messages and the last N user turns |
| Prune | Shrinks chat and page-context scrapes: collapses repeated text, strips filler/stop words on large dumps, trims oversized context (including webpage text in the latest user message) |
| Blob | Replaces data-URIs, long base64, and hex blobs in tool results with short markers |
| Dedup | When the same tool payload appears again, keeps one copy and points later ones at it |
| Skeleton | Large code dumps → keep imports/signatures; drop function bodies |
| Diff | Unified diffs → keep headers and +/- lines; drop excess context |
| PathFold | Repeated absolute path prefixes → ./ |
| Whiteout | Collapse blank lines, trim trailing spaces, minify pretty-printed JSON in tools |
| HtmlLite | Strip scripts, styles, and HTML chrome from tool results; keep readable text |
| Crush | Content-aware shrink of large JSON arrays, logs, and search dumps (keeps errors and relevant hits) |
| RTK | Hard character ceiling on tool / tool_result content (head + tail) |
Completion cost (reply length / request fields)
| Saver | What it does |
|---|---|
| Ponytail | Injects brevity rules into the system prompt (lite / full / ultra) |
| ToolsTrim | Drops unused entries from a large tools[] schema list |
| Cap | Caps upstream max_tokens |
| Stop | Adds stop sequences when the client omitted them (cuts verbose sign-offs) |
Pipeline order:
Clip → Prune → Blob → Dedup → Skeleton → Diff → PathFold → Whiteout → HtmlLite
→ Crush → RTK → Ponytail → ToolsTrim → Cap → Stop → upstream
All savers are pure TypeScript in-process — no Python, no extra compressor installs.
Synthetic fixtures sized to trip each saver’s thresholds. Token estimates use chars÷4. Ponytail grows the prompt on purpose (brevity rules); Cap/Stop change request fields, not message size.
Re-run locally:
npm run test:report| Saver | OK | Before | After | Saved | % | What happened |
|---|---|---|---|---|---|---|
| Clip | ✓ | 55,858 | 7,227 | 48,631 | 87.1% | 34 → 7 msgs; kept last 3 user turns |
| Prune | ✓ | 2,735 | 2,378 | 357 | 13.1% | filler stripped on history turns |
| Blob | ✓ | 657 | 95 | 562 | 85.5% | data-URI + base64 → [Blob: …] markers |
| Dedup | ✓ | 5,762 | 2,937 | 2,825 | 49.0% | 2nd identical tool dump → pointer |
| Skeleton | ✓ | 5,437 | 219 | 5,218 | 96.0% | large TS dump → imports/signatures |
| Diff | ✓ | 3,148 | 672 | 2,476 | 78.7% | kept +/- lines; dropped context |
| PathFold | ✓ | 8,723 | 3,192 | 5,531 | 63.4% | absolute prefix → ./ |
| Whiteout | ✓ | 13,041 | 10,351 | 2,690 | 20.6% | pretty JSON minified |
| HtmlLite | ✓ | 4,095 | 3,363 | 732 | 17.9% | stripped script/style/chrome |
| Crush | ✓ | 27,041 | 3,489 | 23,552 | 87.1% | 400-item JSON → errors/edges kept |
| RTK | ✓ | 12,000 | 4,000 | 8,000 | 66.7% | hard tool-char ceiling |
| Ponytail | ✓ | 18 | 204 | — | — | injected [Ponytail lite] (+186 chars) |
| ToolsTrim | ✓ | 6,958 | 2,515 | 4,443 | 63.9% | tools 21 → 8 |
| Cap | ✓ | — | — | — | — | max_tokens 8192 → 2048 |
| Stop | ✓ | — | — | — | — | injected default stop sequences |
| Combined pipeline | ✓ | 104,658 | 13,727 | 90,931 | 86.9% | full stack; Clip + Blob + Dedup markers present |
All 16 checks passed. Estimated tokens saved on message savers (excluding Cap/Stop): ~48,987.
| Page | Use it for |
|---|---|
| Endpoint | Local URL, API key, host/port |
| Upstream | Providers, models, defaults, API keys |
| Token Savers | Enable and tune each saver |
| Usage | Estimated tokens saved over time |
| Log | Live traffic: raw client request, compressed upstream payload, response |
| Setting in your IDE / agent | Value |
|---|---|
| Base URL / API base | http://127.0.0.1:20128/v1 |
| API key | Local ts-… key from Endpoint |
| Model | Any model id your upstream provider lists (or leave blank to use the dashboard default) |
After a chat, open Log and compare the client request to the UPSTREAM entry — you should see smaller payloads and saver markers such as [Prune: …], [Clip: …], or [RTK compressed: …].
Settings persist on disk:
| OS | Folder |
|---|---|
| Windows | %APPDATA%\Token Savers\ |
| macOS | ~/Library/Application Support/Token Savers/ |
| Linux | ~/.config/token-savers/ |
Override with env var TOKEN_SAVERS_CONFIG_DIR. Main files: config.json, usage.json.
Created by Kei Sakai — @KeiSakaiX — kunani.dev
MIT — see LICENSE.