Skip to main content
Glama

JevRelay

Open-source local MCP runtime for fast browser automation.

An AI creates a declarative Jev Script. JevRelay validates it, runs Playwright locally, and calls the configured inference provider only for explicit bounded decisions.

Claude / ChatGPT / Roxy
-> local @jevrelay/mcp process
-> Playwright actions on the user's computer
-> optional TypeSafe, OpenRouter, or JevRelay decision call

Browser cookies, page state, and computer control stay local.

Status

This is an early MVP. It supports:

  • MCP over stdio.

  • Chromium through Playwright.

  • Declarative Jev Script validation.

  • Local actions, extraction, assertions, background runs, status, and cancellation.

  • Direct TypeSafe Jev inference.

  • OpenRouter structured-output inference.

  • The future api.jevrelay.com/v1/decide endpoint.

It does not execute arbitrary JavaScript, shell commands, or local files.

Related MCP server: laya-browser-mcp

Install

Node.js 20 or newer is required.

npm install -g @jevrelay/mcp
jevrelay-mcp install-browser
jevrelay-mcp doctor

Until the package is published, install and run it directly from GitHub:

npx github:roxy-gg/jevrelay install-browser
npx github:roxy-gg/jevrelay doctor

Or clone the repository:

npm install
npm run build
npx playwright install chromium
node dist/cli.js doctor

MCP Configuration

Claude Desktop

{
  "mcpServers": {
    "jevrelay": {
      "command": "npx",
      "args": ["-y", "@jevrelay/mcp"]
    }
  }
}

Roxy

Before npm publication, use the public GitHub repository:

{
  "mcpServers": {
    "jevrelay": {
      "command": "npx",
      "args": ["-y", "github:roxy-gg/jevrelay"]
    }
  }
}

After npm publication, replace the final argument with @jevrelay/mcp.

Scripts without decide steps need no provider or API key.

Inference Providers

Secrets are environment variables on the local MCP process. Never pass provider keys as MCP tool arguments because tool arguments can enter model transcripts and logs.

TypeSafe Jev

JEVRELAY_PROVIDER=typesafe
TYPESAFE_API_KEY=...

OpenRouter

JEVRELAY_PROVIDER=openrouter
OPENROUTER_API_KEY=...
OPENROUTER_MODEL=openai/gpt-4o-mini

Choose an OpenRouter model that supports structured outputs.

JevRelay Provider

JEVRELAY_PROVIDER=jevrelay
JEVRELAY_API_KEY=...
JEVRELAY_API_URL=https://api.jevrelay.com/v1/decide

The hosted JevRelay provider is optional and not implemented in this repository.

Example MCP configuration with a direct TypeSafe key:

{
  "mcpServers": {
    "jevrelay": {
      "command": "npx",
      "args": ["-y", "@jevrelay/mcp"],
      "env": {
        "JEVRELAY_PROVIDER": "typesafe",
        "TYPESAFE_API_KEY": "..."
      }
    }
  }
}

MCP Tools

jevrelay_validate

Validates a Jev Script without running it.

jevrelay_run

Runs a script. Arguments:

  • script: Jev Script object.

  • inputs: Optional input overrides.

  • headless: Defaults to true.

  • wait: Defaults to true. Set to false for a background run.

jevrelay_status

Returns progress and partial output for a run ID.

jevrelay_stop

Requests cancellation of a running automation.

Jev Script

{
  "version": 1,
  "name": "read-example",
  "permissions": {
    "origins": ["https://example.com"]
  },
  "steps": [
    {
      "action": "browser.goto",
      "url": "https://example.com"
    },
    {
      "action": "browser.assert",
      "target": { "role": "heading", "name": "Example Domain" },
      "state": "visible"
    },
    {
      "action": "browser.extract",
      "target": { "role": "heading", "name": "Example Domain" },
      "fields": ["text"],
      "saveAs": "heading"
    }
  ]
}

Run it through MCP, or use the exported RunManager from Node.

A decision step chooses only from script-supplied or extracted options:

{
  "decide": {
    "state": "${vars.videos}",
    "question": "Which video best matches the request?",
    "optionsFrom": "videos.href",
    "minimumConfidence": 0.5,
    "saveAs": "videoUrl"
  }
}

If confidence is below the threshold, the run returns needs_input and does not execute the next action.

Browser Actions

The MVP supports:

browser.goto
browser.fill
browser.press
browser.click
browser.wait
browser.extract
browser.assert

Targets can use:

{ "selector": "article a" }
{ "role": "button", "name": "Play" }
{ "text": "Learn more" }
{ "id": "video-123" }

browser.extract fields include text, id, href, value, ariaLabel, tagName, and attr:<name>.

Security Boundary

  • Scripts declare allowed top-level origins.

  • Navigation, redirects, and popups to undeclared top-level origins are blocked.

  • Subresources loaded by an allowed page are not origin-restricted in the MVP.

  • Unknown actions and fields fail schema validation.

  • No shell execution or arbitrary JavaScript exists in the script format.

  • Provider secrets come from the local process environment.

  • Inference runs only at explicit decide steps.

  • Low-confidence decisions stop instead of guessing.

This is not yet a complete sandbox. Review generated scripts before using them with sensitive accounts, purchases, messages, or destructive workflows.

Development

npm install
npm run format
npm run check

Install Chromium once for browser tests and real runs:

npx playwright install chromium

License

MIT

Available Tools

4 tools
jevrelay_runRun Jev ScriptA
Destructive

Validate and execute a Jev Script locally with Playwright. Inference is used only for explicit decide steps.

Inputs: script (required), wait, inputs, headless.

jevrelay_statusGet JevRelay Run StatusB
Read-only

Return the current status and partial output of a JevRelay run.

Inputs: runId (required).

jevrelay_stopStop JevRelay RunB
Destructive

Request cancellation of a running JevRelay automation.

Inputs: runId (required).

jevrelay_validateValidate Jev ScriptA
Read-only

Validate a declarative Jev Script without executing it.

Inputs: script (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct role in the script lifecycle: validate checks syntax, run executes, status reports progress, and stop cancels. Although run also validates, its execution focus makes it distinct from validate-only.

    Naming Consistency5/5

    All tools use the same jevrelay_ prefix and snake_case format with predictable verb/noun labels. The only minor variation is status being a noun, but it remains clear and consistent with the overall scheme.

    Tool Count5/5

    Four tools is well-scoped for a script validation and execution service. Each tool covers a necessary lifecycle operation without redundancy.

    Completeness4/5

    The surface covers the core lifecycle: validate, run, check status, and stop. Minor gaps exist, such as listing past runs or retrieving full logs/artifacts, but agents can work around them.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables web browser automation and inspection using structured data instead of screenshots, allowing AI agents to interact with web pages programmatically through the Playwright framework.
      ISC
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to perform Playwright-compatible browser automation through unchanged upstream tools, using a Chromium runtime with persistent profiles, proxy matching, humanized interactions, and Docker or stdio/Streamable HTTP deployment options.
      280 npm
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to perform Playwright-compatible browser automation through a Chromium-based runtime, with persistent profiles, proxy matching, extension loading, Docker deployment, and stdio or Streamable HTTP transports. It supports web research, daily automation, and testing workflows driven by natural-language prompts.
      280 npm
      MIT