Skip to content

Harness

The harness lets the Playground accept a task in natural language, for example "Go to CNN and get the page title". The harness sends the task to an LLM, which writes the script. The server verifies that script before it runs.

Together with the task, the harness sends the accessibility tree of the page that is open in the browser. Thus, the LLM can write a script for that page, for example "Click the Sign in button". Each request has only the system prompt, the page, and the task, not the earlier tasks.

Without an LLM, the playground still works, but it accepts only scripts.

Configuration

The harness works with any service that has an OpenAI-compatible chat completions endpoint (/chat/completions). This includes a managed service or a local LLM server. Set these environment variables:

Env name Description
LLM_API_BASE_URL Base URL of the API, without /chat/completions
LLM_API_KEY API key, if the service needs one. A local server often does not.
LLM_CHAT_MODEL Model name

Both LLM_API_BASE_URL and LLM_CHAT_MODEL are required to turn on the harness.

At startup, the server asks the LLM to write a simple script. If the script is correct, the playground offers natural language. The server checks again every 5 minutes, so the playground follows the status of the LLM without a restart. If the test fails, the server logs HARNESS LLM check failed, and the playground accepts only scripts.

The quality of the scripts depends on the model. A fast model that writes good JavaScript works best, because the user waits for the answer before the script runs.

The text that a user types, and the content of the page that is open, go to the LLM service. If that is a problem, use a local LLM server.

The server does not trust the LLM. A task can contain a prompt injection, for example text that the user pasted from a different site. The page can also contain one. Thus, a script from the LLM must also obey stricter rules, as described in Verification.

Verification

The server verifies each script before it runs. It verifies the user's own scripts in the same way, except where noted below. If verification fails, the playground shows the reason.

Each script must obey these rules:

  • It is JavaScript only. It declares variables with const or let.
  • It uses only the page methods that the Playground documents, and only console.log, info, warn, and error.
  • Apart from its own variables, it uses only JSON, Math, Number, String, Boolean, Array, undefined, and NaN.
  • It does not use fetch, network objects, cookies, storage, or location.
  • It does not change a string into code, for example with eval or Function.
  • The functions in page.evaluate, page.$eval, and page.$$eval only read the page. They do not change elements.

A task can contain text that tries to control the LLM (a prompt injection), for example text that the user pasted from a different site. Thus, the server does not trust a script from the LLM. That script must also obey these rules:

  • It does not use page.cdp.
  • It uses an index only with a literal value, for example items[0]. A variable index could make a name that is not permitted.

If the server refuses the script from the LLM, the task fails. The LLM does not get another attempt.

Verification is not a sandbox. A script that passes verification still controls the browser. Give the LLM only tasks that you also run as a script.

Examples

OpenAI

LLM_API_BASE_URL=https://api.openai.com/v1
LLM_API_KEY=yourownapikey
LLM_CHAT_MODEL=gpt-6-luna

OpenRouter

LLM_API_BASE_URL=https://openrouter.ai/api/v1
LLM_API_KEY=yourownapikey
LLM_CHAT_MODEL=deepseek/deepseek-v4.1-flash

DeepInfra

LLM_API_BASE_URL=https://api.deepinfra.com/v1/openai
LLM_API_KEY=yourownapikey
LLM_CHAT_MODEL=deepseek-ai/DeepSeek-V4.1-Flash

Local LLM

A local LLM server also works, as long as it has an OpenAI-compatible endpoint. Such a server often needs no API key.

With llama.cpp, start llama-server with a model. It serves only that model, so LLM_CHAT_MODEL can be any name, but it must be set.

llama-server -m /path/to/model.gguf

LLM_API_BASE_URL=http://127.0.0.1:8080/v1
LLM_CHAT_MODEL=local

With LM Studio, load a model and start the server in the Developer tab. Set LLM_CHAT_MODEL to the model name that LM Studio shows there, for example:

LLM_API_BASE_URL=http://127.0.0.1:1234/v1
LLM_CHAT_MODEL=openai/gpt-oss-20b

Other local servers work in the same way.

A small local model may fail the test at startup, or write scripts that often fail. In that case, try a larger model.