# SoftGlaze AI (Beta): an MCP Server for Your Browser Profiles

> Connect Claude, Cursor, VS Code, Gemini CLI, Codex or n8n to SoftGlaze Browser through a local MCP server. 40 tools, approvals in the app, Stop all AI.

- Canonical URL: https://browser.softglaze.com/features/ai-agents
- Last updated 2026-11-09

---

SoftGlaze AI (Beta)

## Let your AI app work inside your browser profiles, and approve every change

SoftGlaze AI is an MCP server inside SoftGlaze Browser. Connect Claude, Cursor, VS Code, Gemini CLI, Codex, Windsurf, Cline or n8n, and the AI can read your profiles and proxies, explain why a profile gets CAPTCHAs, start profiles, run macros and work inside pages. Each profile it touches keeps its own proxy and fingerprint.

Anything that changes something waits for your Allow in SoftGlaze, whatever the AI app’s own settings say. One button stops it all.

[Read the connection guide](https://browser.softglaze.com/docs/ai-agents) [See all 40 tools](https://browser.softglaze.com/docs/ai-tools)

Beta. Off until you turn it on. Runs on your computer only, at 127.0.0.1.

Beta, from version 1.0.22

SoftGlaze AI ships as a Beta from SoftGlaze Browser 1.0.22, so the version on the download page has it. If the SoftGlaze AI page isn’t in your sidebar, update the app. The Microsoft Store copy can trail the download by a few days. Tools, prompts and setups may still change before the full release.

### Copy and paste between a chat and the app, or let the AI look for itself

Without SoftGlaze AI

You copy profile names and proxy errors into a chat, then carry the AI’s advice back by hand, one setting at a time. The AI never sees the real state of your workspace, so its answers are guesses.

With SoftGlaze AI

The AI reads your profiles, proxies and activity itself, explains what is wrong and proposes the fix. You see each change in a SoftGlaze dialog, with the profiles it touches named, and press Allow or Deny.

01What it is

### A local MCP server inside the app

The Model Context Protocol (MCP) is the open standard AI apps use to call outside tools. SoftGlaze AI is an MCP server that runs inside SoftGlaze Browser on your own computer, at `http://127.0.0.1:<PORT>/mcp`. It answers only there, and only to requests that carry a token you created.

Apps that speak Streamable HTTP connect straight to it. n8n can use the older SSE transport at `/mcp/sse`, and Claude Desktop starts a small bridge that ships with SoftGlaze. The same 40 tools are on the local REST API too, for scripts.

Terminal

```
claude mcp add --transport http softglaze http://127.0.0.1:<PORT>/mcp --header "Authorization: Bearer <YOUR_TOKEN>"
```

_The Claude Code setup the app gives you. The port is 8080 unless you change it; the token is yours._

02Real profiles

### The AI works inside real profiles, each with its own proxy and fingerprint

Page tools act in a running profile’s main tab, through that profile’s proxy and fingerprint, over the connection SoftGlaze already has to the browser. No second debugger is attached. Clicks move the mouse like a person and typing goes one key at a time with human-like gaps, using the same steps as SoftGlaze’s own macros.

Page text comes back marked as untrusted, and AI apps are told never to follow instructions found in it. Page tools work on Chromium profiles, and you can limit them to the sites you list.

Terminal

```
page_navigate      page_snapshot      page_click
page_type          page_press_key     page_scroll
page_screenshot    page_extract_text  page_wait_for
```

_The nine page tools, by the names AI apps see._

03Approvals

### Every change asks you inside SoftGlaze

When an AI app wants to create, start, change or run something, SoftGlaze shows a dialog on any page of the app: which app, which token, which member it acts as, what it wants to do and which profiles it touches. Allow once, Allow for 1 hour, or Deny. No answer in 2 minutes counts as Deny.

SoftGlaze decides this, not the AI app’s own settings. Read-only tools don’t ask by default. Starting runs on many profiles and stopping all profiles ask every time.

Terminal

```
Claude Code is asking for permission
Claude Code wants to: Start profile #42
[ Deny ]   [ Allow for 1 hour ]   [ Allow once ]
With no answer, the request is denied.
```

_The shape of a request. The profile id is a placeholder._

04Tokens and Stop all AI

### Scoped, expiring tokens, and one button that stops everything

Each AI app connects with its own token. A token acts as one team member, with that member’s permissions and profiles, and can be limited to some groups. Choose Read only, Operate or Full, and let it expire after 7, 30 or 90 days. SoftGlaze shows a token once and stores only its hash. Revoke one and that app is cut off at once.

Stop all AI stops every run and profile an AI app started, disconnects every AI app, denies every waiting request and switches SoftGlaze AI off.

Terminal

```
Token:     sgai_ + 40 characters, shown once
Scope:     Read only | Operate | Full
Expires:   7, 30 or 90 days, or never
Profiles:  all the member's, or only some groups
```

_What you choose when you create a token._

05Memories and prompts

### Your rules, read before every task, and 12 ready-made prompts

The workspace memory is a page of your own rules in Markdown, up to 20 KB: how you name profiles, what each group is for, proxy rules, working hours and a never-do list. SoftGlaze tells every AI app to read it first, with two built-in guides: how SoftGlaze works, and the acceptable use policy.

Twelve prompts come ready: a daily health check, why a profile gets CAPTCHAs, create profiles for a country, fix timezone and language mismatches, replace failing proxies, write a macro from a description, test a site from several countries, a weekly report, and more.

Terminal

```
## Proxy rules
- Match the proxy country to the profile's country.
- One proxy per profile for logged-in accounts.
- Check a proxy before use if its last check is older than a day.
```

_Lines from the workspace memory template. Replace them with your own rules._

06n8n and automation tokens

### n8n workflows, with automation tokens for the runs nobody watches

Self-hosted n8n on the same computer connects with the MCP Client Tool node, so an n8n AI Agent can use SoftGlaze, or calls tools directly with the HTTP Request node. A scheduled workflow has nobody to press Allow, so the Owner can give it an automation token: it pre-approves only the tools you tick, only in the groups you tick, with at most 5 profiles per call and 60 calls an hour unless you set other limits. Everything else still asks.

Four workflow templates (Beta) are in the guide, from a daily health report to profiles created from a Google Sheet.

Terminal

```
SSE Endpoint:    http://127.0.0.1:<PORT>/mcp/sse
Authentication:  Bearer, with your token
Tools to Include: All, or only the ones you need
```

_The MCP Client Tool node in n8n. n8n Cloud can't reach your computer yet._

### The AI apps you can connect today

Each has a step-by-step section in the guide, with the setup the app’s Connect tab gives you. Pick an app to go straight to it.

**Supported AI apps**

| AI app | How it connects |
| --- | --- |
| [Claude Desktop](https://browser.softglaze.com/docs/ai-agents#claude-desktop) | The SoftGlaze bridge, started by Claude Desktop |
| [Claude Code](https://browser.softglaze.com/docs/ai-agents#claude-code) | Streamable HTTP, one terminal command |
| [Cursor](https://browser.softglaze.com/docs/ai-agents#cursor) | Streamable HTTP, in mcp.json |
| [VS Code (GitHub Copilot)](https://browser.softglaze.com/docs/ai-agents#vscode) | Streamable HTTP; VS Code asks for the token and stores it securely |
| [Gemini CLI](https://browser.softglaze.com/docs/ai-agents#gemini-cli) | Streamable HTTP, in settings.json |
| [ChatGPT desktop and Codex](https://browser.softglaze.com/docs/ai-agents#codex) | Streamable HTTP, in config.toml, token from an environment variable |
| [Windsurf (Devin Desktop)](https://browser.softglaze.com/docs/ai-agents#windsurf) | Streamable HTTP, in mcp\_config.json |
| [Cline](https://browser.softglaze.com/docs/ai-agents#cline) | Streamable HTTP, in cline\_mcp\_settings.json |
| [n8n (self-hosted)](https://browser.softglaze.com/docs/ai-agents#n8n) | SSE with the MCP Client Tool node, or the local API |
| [Other MCP apps](https://browser.softglaze.com/docs/ai-agents#other) | Streamable HTTP, SSE or the bridge |

Not yet: ChatGPT on the web, Grok and n8n Cloud. They run on someone else’s servers and can’t reach a server on your computer. They need the Cloud Relay, which comes after the Beta.

### How SoftGlaze keeps the AI on a short lead

#### Approvals inside SoftGlaze

Changes wait for Allow once, Allow for 1 hour or Deny in the app. No answer in 2 minutes is a Deny. The AI app’s own auto-approve settings don’t count.

#### Stop all AI

One button stops what AI apps started, disconnects them, denies every waiting request and switches SoftGlaze AI off.

#### Scoped, expiring tokens

Read only, Operate or Full; acts as one member; can be limited to groups; expires after 7, 30 or 90 days; stored as a hash; revocable at once.

#### Everything is logged

Each call is listed under Activity: app, token, tool, what happened, how it was approved, the result and how long it took. Secrets are hidden.

#### Never exposed

No tool reads cookies, 2FA codes, proxy or saved-login passwords, deletes anything, touches billing, changes members or roles, weakens security settings or runs arbitrary JavaScript.

#### Plan limits apply

A macro run handles at most 3 profiles at once on Pro and 10 on Enterprise, whatever the AI asks for. Each token is held to 120 calls a minute.

### Connect your first AI app in three steps

1.  Turn it on

    Open SoftGlaze AI in the sidebar and switch on AI server. The Owner or an Admin can do this.

2.  Create a token

    On the Tokens tab, name it after the app, choose who it acts as and what it may do, and copy it. It is shown once.

3.  Paste the setup

    The Connect tab has the setup for your app with your port filled in. Paste it with your token, then ask the AI to read your workspace summary.

[Read the full guide](https://browser.softglaze.com/docs/ai-agents)

Limits and caveats

### What to know before you rely on it

#### It is a Beta

It is in version 1.0.22 and later, and tools, prompts and setups may change before the full release. Tell us when something looks wrong.

#### Local only, for now

The AI app must run on the same Windows computer as SoftGlaze. Web-only AI apps and n8n Cloud need the Cloud Relay, which comes later. Want to use it from claude.ai, ChatGPT or your phone? [Vote for it](https://browser.softglaze.com/feature-requests#request-1): your votes decide when we build it.

#### Sites can notice automation

Page tools act like a person, through each profile’s own proxy and fingerprint, but some sites can still detect automation of any kind. SoftGlaze makes profiles harder to link, not impossible.

#### AI apps make mistakes

They can misread a page or propose the wrong fix. That is what the approval dialog is for: read each request before you allow it.

#### Your AI provider sees what the AI reads

SoftGlaze sends nothing to SoftGlaze. Your AI app sends what it reads, such as profile names, proxy countries and page text, to its own AI provider under that provider’s terms.

#### Page tools need Chromium

They work in running Chromium profiles only, not Firefox ones, and only on the sites you allow when you set a list.

### SoftGlaze AI questions

[Email support](mailto:support@softglaze.com)

One business day, Monday to Friday for a first reply.

#### What is MCP?

The Model Context Protocol is an open standard that lets AI apps call outside tools. SoftGlaze AI is an MCP server, so any app that supports MCP over HTTP, or that can start a local program, can connect to it.

#### Which AI apps work with SoftGlaze AI?

Claude Desktop, Claude Code, Cursor, VS Code with GitHub Copilot, Gemini CLI, the ChatGPT desktop app and Codex, Windsurf, Cline, self-hosted n8n, and other MCP apps that run on the same computer. The guide has the setup for each.

[Connection guide](https://browser.softglaze.com/docs/ai-agents)

#### Can the AI change things without asking me?

Not unless you set it up that way. Changes ask you in SoftGlaze first, whatever the AI app’s own settings say. Read-only tools don’t ask by default, and an automation token can pre-approve a few tools you pick, for groups you pick, within limits. Stop all AI ends everything at once.

#### Can the AI read my passwords, cookies or 2FA codes?

No. No tool returns cookies, 2FA codes, proxy usernames or passwords, or saved-login and Data Vault passwords, and none can delete anything, touch billing or change members and roles.

[What is never exposed](https://browser.softglaze.com/docs/ai-tools#never-exposed)

#### Does my data leave my computer?

SoftGlaze AI sends nothing to SoftGlaze. The AI app you connect sends what it reads to its own AI provider, as it does with anything else you show it. Limit a token to the groups the AI needs, and choose Read only when it only has to look.

#### Does it work with ChatGPT on the web or n8n Cloud?

Not yet. They run on remote servers and can’t reach a server on your computer. They need the Cloud Relay, which comes after the Beta. Self-hosted n8n on the same computer works now.

#### Will sites detect a profile the AI drives?

They might. Page tools act like a person, through the profile’s own proxy and fingerprint, without attaching a second debugger, but some sites can detect automation of any kind. The optional Expert mode, which hands a profile to Playwright or Puppeteer, is easier to detect and is off by default.

#### Which plans include SoftGlaze AI?

It is part of the app, on Pro, Enterprise and the trial. Plan limits still apply: a macro run handles up to 3 profiles at once on Pro and 10 on Enterprise and during the trial.

### Related features

#### [Automation](https://browser.softglaze.com/features/automation)

The macros and parallel runs the AI can start, and the local API it shares.

#### [Proxy pool](https://browser.softglaze.com/features/proxy-pool)

The proxy checks and health history the AI reads when it explains a CAPTCHA.

#### [Team access](https://browser.softglaze.com/features/team-access)

Every token acts as one member, with that member’s role, limits and profiles.

Read it in the source

### The files this page is checked against

-   [src/main/ai/toolRegistry.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/toolRegistry.js) the checks every call goes through, the never-exposed list and the rate limit
-   [src/main/ai/tools.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/tools.js) the tools, their scopes, approval rules and what they return
-   [src/main/ai/pageTools.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/pageTools.js) the page tools and how page text is marked untrusted
-   [src/main/ai/approvals.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/approvals.js) the approval dialog, Allow for 1 hour and the 2-minute timeout
-   [src/main/ai/aiTokens.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/aiTokens.js) token format, scopes, expiry, hashing and automation limits
-   [src/main/ai/memories.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/memories.js) the workspace memory template and the built-in guides
-   [src/main/ai/prompts.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/prompts.js) the 12 prompts, word for word
-   [src/main/ai/snippets.js](https://github.com/softglazee/softglaze-browser/blob/main/src/main/ai/snippets.js) the setup each AI app gets

[The repository](https://github.com/softglazee/softglaze-browser)

### Connect your first AI app

The guide has the exact setup for each app, the prompt library and every tool. SoftGlaze AI is in version 1.0.22 as a Beta, off until you turn it on.

[Read the connection guide](https://browser.softglaze.com/docs/ai-agents) [Download for Windows](https://browser.softglaze.com/download)

Windows 10 and 11, 64-bit. 7-day trial, no card needed.
