Put your existing web LLM accounts to work for your agents.
Quick start · What is Web Harness? · Providers · Capabilities
Local Dashboard captured on 2026-09-06. Values are local job records and output-token estimates, not benchmarks or billing savings.
Tokenless Harness manages tasks and tool calls; Tokenless API connects your web LLM accounts and returns model responses to the Harness.
Keep your Harness’s agent loop, tools, and sessions; connect its model interface to Tokenless API.
API setup and compatibility limits · Harness integration
Add the Tokenless skill to your Harness; invoke it for selected tasks and bring the results back into your usual workflow.
tokenless setup installs the skill into supported local agent skill directories. Setup · Skill instructions
AI-generated use-case illustrations. Available capabilities depend on the selected provider’s verified support.
19 in Browser mode, 35 in Direct API mode; 13 appear in both. Inventory audit.
Interact with provider websites in your signed-in browser.
ChatGPT Supported |
Claude Supported |
 Gemini Supported |
 Grok Supported |
 Qwen / 千问 Experimental |
 DeepSeek Experimental |
 Perplexity Experimental |
 Z.ai / GLM Experimental |
 Doubao / 豆包 Experimental |
Kimi Experimental |
 Dola Experimental |
Arena Supported |
Meta AI Experimental |
Microsoft Copilot Awaiting verification |
 GitHub Copilot Experimental |
 Lovable Experimental |
 HuggingChat Awaiting route |
 Monica Experimental |
 Agnes AI Experimental Chat + inputs |
Call provider endpoints through G4F; ChatGPT and Perplexity also have native backends. Some integrations are experimental; authentication and capabilities vary by provider.
| Provider | Provider | Provider | Provider |
|---|---|---|---|
| ChatGPT | Claude | Gemini | Grok |
| Qwen / 千问 | DeepSeek | Perplexity | Z.ai / GLM |
| Arena | Meta AI | Microsoft Copilot | GitHub Copilot |
| Black Forest Labs | Cerebras | Cloudflare AI | Cohere |
| DeepInfra | ElevenLabs | Groq | Hugging Face |
| MiniMax | NVIDIA | Ollama | OpenRouter |
| Opera Aria | Phind AI | Pi | Pollinations |
| Replicate | Sber GigaChat | Stability AI | Teach Anything |
| Together AI | WhiteRabbitNeo | YQCloud |
Direct API mode setup and limits · Verified provider capabilities
Requires Node.js 22.13+. Apple Silicon macOS is the current target; Windows x64 is prerelease.
npm install --global tokenless@latest
tokenless setup
tokenless run --provider chatgpt --prompt "Review this proposal."Setup opens the local dashboard. Reopen it anytime with tokenless dashboard.
Browser preparation and updates
Setup requires uv for the G4F runtime and synchronizes matching skills; upgrades sync them too. Use tokenless skills sync --json to refresh skills alone. The macOS menu app is a separate optional install and is not installed on Windows.
For native mode, use a current Chrome or Brave, enable remote debugging at chrome://inspect/#remote-debugging or brave://inspect/#remote-debugging, and approve the browser prompt. Setup also offers an Anti-Detect option.
Already installed? Run tokenless upgrade --check, then tokenless upgrade. See Updates for CLI and macOS app updates.
Developing on Windows? The Windows tray app opens Dashboard on left-click and provides a native right-click menu.
We call the layer that turns web LLMs into an agent’s working environment a Web Harness. Tokenless lets agents submit tasks through your existing web LLM accounts, use supported website capabilities, and bring results back into your workflow.
| What you want to do | What Tokenless handles |
|---|---|
| Put web LLMs to work for your agent | Submit prompts, read responses, and continue supported conversations. |
| Work with your own material | Use attachments, citations, and controls verified for the selected provider. |
| Connect an app and follow progress | Run tasks through the CLI or local compatible APIs; view history and usage in Dashboard. |
Web workflows need no separate provider API keys; sign-in stays in your selected browser. Tokenless API provides provider access; Tokenless Harness manages agent tasks and tool continuation.
Browser mode example: DeepSeek
Send tokenless/deepseek through the visible DeepSeek website using the local OpenAI-compatible interface, then return the response to the caller.
Local Spark X2.5-4B router engine
On Apple Silicon, the Dashboard can use the official Spark MLX server for the local Spark X2.5-4B model. Ollama is not required; the V1 integration uses the fixed OpenAI-compatible endpoint below.
git clone https://github.com/XHToken/Spark-MLX-LLM.git
cd Spark-MLX-LLM
python3 -m venv .venv
.venv/bin/python -m pip install -e '.[test]'
.venv/bin/spark-mlx-server --model XHToken/Spark-X2.5-4B --host 127.0.0.1 --port 8080 --allowed-origins http://127.0.0.1:7331In Dashboard → System → Semantic routing, select Spark X2.5-4B · local MLX and save. The health endpoint is http://127.0.0.1:8080/health; chat completions use http://127.0.0.1:8080/v1/chat/completions.
Codex integration
Install the optional Codex integration:
tokenless setup --install-codexRestart Codex, open /hooks, and trust Tokenless.
The experimental Tokenless Harness Browser Extension supports user-approved observation and text input on one selected Chrome tab.
- CLI commands
- API Proxy Integration
- Harness Integrations
- Capability Matrix
- Privacy boundaries
- Documentation index
Tokenless is in early access: it reduces agent-side token use, but does not eliminate token use or bypass provider account requirements.
