Working in a browser the user is already signed into

Most browser automation starts from an empty profile: no cookies, no session, no access to anything behind a login. For public pages that is fine and you should keep using it.

This skill is for the other case — where the authentication is the hard part. A scripted login against SSO, MFA, a CAPTCHA or bot detection is slow to build, brittle to keep, and often against the site's terms. Attaching to a session the user already has skips all of it.

The tools come from the T3rnel Session Bridge MCP server (session_*). It is a local stdio server; nothing is hosted and no page content leaves the machine.

Decide first: does this task actually need a real session?

Situation	What to use
Public page, no login	An ordinary headless browser or fetch. Don't reach for this.
Anything behind a login	This skill.
A scripted login that keeps breaking	This skill — stop maintaining the login.
Reading a page's rendered CSS, or its React tree	This skill; static HTML won't have it.
Bulk scraping of a site you have no account on	Neither. That's a different problem.

If the page is public, say so and use the simpler tool. Reaching into someone's live browser for a page you could have fetched is a cost with no benefit.

The rule that matters most

This is the user's real browser, logged into their real accounts. A wrong click here is not a failed test — it can send an email, cancel a subscription, post publicly, or move money.

1. Read before you write. session_list_tabs, then session_snapshot or session_read_page, before any click or fill.
2. Confirm anything irreversible, in your own words, before doing it: purchases, payments, sending or publishing, deleting, permission and sharing changes, anything in an account-settings or billing page. Describe the specific action and wait.
3. Never enter credentials you generated. If a field is a password, an OTP, a card number or a security answer, stop and ask. session_store_login exists so the user supplies these once, deliberately.
4. Stay on the tab you were given. Don't wander into other tabs of a signed-in browser looking for context.

Workflow

session_list_tabs          → find the tab, or confirm which one the user means
session_navigate           → go somewhere (reuses the active tab unless newTab: true)
session_snapshot           → see the accessibility tree: what is actually clickable
session_click / _fill      → act, using a selector you saw in the snapshot
session_snapshot again     → verify the outcome changed the way you expected

Verify, don't assume. A click that resolves is not a click that worked. Re-snapshot and assert the thing you expected: the row is gone, the banner says saved, the URL moved. Half of all silent automation failures are a selector that matched nothing and a caller that never looked.

When a form does not submit, read the form. A page that rejects your input says why, and session_snapshot carries it: a refused field comes back with invalid: true and a problem holding the message — "Description must be at most 400 characters" — taken from the page's own error text or the browser's validation message. Fields also report maxLength and required, so the limits are visible before you fill rather than after you are refused.

So the order after a failed submit is: re-snapshot, look for problem on any field, fix that, submit again. Do not conclude "it probably needs a login" from a form that stayed put — that guess is wrong more often than it is right, and the answer was on the page. If the snapshot shows nothing and the form still will not go, session_screenshot and read it.

Prefer selectors you have just seen in a snapshot over ones you guessed from memory of how the site used to look.

When something isn't there

• No browser session / relay not found. The bridge falls back to a standalone automation browser, which works but is signed into nothing. If the task needs the user's session, the free T3rnel Browser extension is what supplies the signed-in half — Setup below has the three steps, and session_install returns the same instructions at runtime.
• A tool says it needs a licence. The bridge itself is free and holds no licence. Some of the deeper extension tools (CSS-as-component export, the React inspector, session recording, the credential vault) are part of the extension's paid tier; the refusal comes from the extension and names what it needs. Report it and move on; don't retry it in a loop.
• A page won't load or a selector never appears. Use session_wait on a condition rather than sleeping, then re-snapshot. If it still isn't there, say so — do not invent a selector.

What the extension adds

The bridge on its own serves 14 tools against a standalone browser: session_install, session_health, session_list_tabs, session_navigate, session_snapshot, session_read_page, session_click, session_fill, session_type, session_select, session_press, session_evaluate, session_screenshot and session_wait. That is enough to browse, read and fill in almost anything.

The other eight refuse without the extension, so don't plan a standalone run around them: session_close_tab, session_login, session_store_login, session_record_start, session_record_stop, session_record_list, session_record_events and session_record_replay. session_login in particular has no standalone equivalent — the credential vault lives in the extension, so sign in by hand once in the standalone profile instead and the session persists.

With the free T3rnel Browser extension connected, the same server serves 103 — the extra ones run inside the user's real tabs:

• CSS extraction — copy an element's computed styles out as a component (styled-components, Tailwind, CSS Modules, JSX and others), which static HTML cannot give you.
• Full-page capture and annotation, beyond the viewport.
• React inspector — component tree, props and state on a running app.
• Network and console capture for what the page actually did.
• Form filling that respects each field's real constraints, and never invents passwords, OTPs or card numbers.
• Record and replay a UI flow, and export it as a Playwright, Puppeteer, Cypress or Selenium spec.

session_install explains how to add it. The extension is free; the tools listed above that are marked Pro will say so themselves when called.

Setup

Add the bridge to any MCP client as a stdio server:

{
  "mcpServers": {
    "t3rnel-session": {
      "command": "npx",
      "args": ["-y", "@t3ratech/mcp-session-bridge"]
    }
  }
}

That alone gives you the 14 standalone tools against a dedicated automation browser — free, no extension, no account. If the task is a public page, stop here.

Adding the signed-in half

The signed-in tools come from the free T3rnel Browser extension, and connecting it takes three steps rather than one. The middle step is the one that gets missed.

npm install -g @t3ratech/mcp-session-bridge
mcp-session-bridge --install     # registers the native messaging host

Then install the extension for the user's browser:

Browser	Where
Chrome, Brave, Chromium	https://chromewebstore.google.com/detail/egpckhdpkoeimoekciejbmbbcackhdmd
Microsoft Edge	https://microsoftedge.microsoft.com/addons/detail/t3rnel-browser/dnplmolfblplbeclnglmjamppekpbcjo
Firefox	https://addons.mozilla.org/en-US/firefox/addon/t3rnel-browser/

Restart the browser, then session_health to confirm the transport is extension.

Why --install matters. The extension and the bridge talk over a native messaging host. Without that registration both halves are installed, both look healthy, and neither can reach the other — session_health reports the extension as missing after what looked like a clean install. If you skip one line, skip a different one.

session_install returns these same instructions at runtime, so an agent that already has the bridge can fetch them without this file.

On first run the extension asks one question

Onboarding offers "Allow MCP automation", ticked by default. Left ticked, connected clients act without a prompt each time. Unticked, the approval gate stays on: high-risk actions stop and ask, you get a refusal naming the exact action, and nothing happens until a human approves it. It is changeable later in Settings.

A refusal that names an action is that gate, not a failure. Report what it wanted and wait. Never retry it in a loop, and never suggest turning the gate off to get past one action.

Optional environment

T3RNEL_SESSION_MODE (auto | extension | standalone), T3RNEL_SESSION_HEADLESS=1, T3RNEL_SESSION_TIMEOUT_MS, T3RNEL_SESSION_BROWSER, T3RNEL_SESSION_PROFILE.

Firefox has no debugger API, so network capture, console capture and PDF export are absent there. Say so rather than letting the user find out.

Verifying a run actually happened

Before reporting success, the last session_snapshot or session_read_page shows the state you claim to have produced. A click that returned OK is a delivered command, not a changed page — the confirmation screen is the evidence, not the tool's own result.

Docs: https://t3ratech.github.io/t3rnel-browser-plugin/tools.html