tex-browse-use

What is TEX?

TEX gives Claude Code a hand on the mouse.

Most "agent" tooling stops where the API stops. But the vast majority of software in the real world — internal admin panels, legacy portals, dashboards, SaaS tools that gate everything behind a login — has no usable API, MCP connector, or CLI. For everything in that long tail, the only interface is the UI a human would click through.

TEX is the engine that clicks through it. You give it a task in plain language; it opens a real browser, navigates, reads, types, and reports back. It ships as a Claude Code plugin: a local engine (Node + Playwright) plus a thin MCP server that hands Claude two tools — browse_use and tex_health.

You:    "Log into the portal, find this month's invoices, and list the unpaid ones."
TEX:    *opens browser → logs in → navigates → reads the table → reports back*

Related MCP server: Cloudflare Playwright MCP

Why TEX?

• The long tail has no API. You can't integrate what nobody exposes. TEX turns "there's no API for that" into "done."

• Browse-use is becoming the default. As vision models get cheaper and sharper, driving the UI directly stops being a hack and starts being the general-purpose path. TEX is built for that future — and gets better the more you use it.

• It learns. Every successful run can be saved as a per-app skill, and a proven skill can be compiled to a $0, no-LLM replay. The tenth time you do a task, it can be deterministic and free.

• It's yours. Self-hosted (Node + a browser), your keys, your machine. No third party in the loop.

How it works — the fallback ladder

TEX only engages when there's no structured access (API / MCP / CLI / A2A). From there it descends a UI ladder, cheapest and most reliable first, escalating only when it has to:

   structured (API / MCP / CLI)        ← handled upstream; TEX doesn't run here
   ─────────────────────────────────────────────────────────────────────────
1. compiled-replay   →  learned, no-LLM script. deterministic, $0.
2. browser-use       →  DOM / CDP automation. fast, robust on normal pages.
3. computer-use      →  vision + stealth. the last resort for tough UIs.

A smart router picks the highest gear it can for the task, and downshifts on failure (self-healing). Completion is checked against verifiable postconditions, not the model's say-so — so "done" means done.

Quickstart

Requirements: Node.js ≥ 20 · playwright Chromium · an LLM key (Anthropic API or AWS Bedrock). Postgres is optional.

# 1. Engine deps + the browser
cd stack && npm install && npx playwright install chromium

# 2. MCP server deps
cd ../mcp && npm install

# 3. Config — copy the template (repo root) and set your provider key
cp .env.example .env        # set LLM_PROVIDER + ANTHROPIC_API_KEY (or AWS creds)

# 4. Start the engine (loads .env, waits for /health, prints status)
scripts/tex-up.sh

Register it with Claude Code — pick one:

# a) The whole plugin (tools + the skill that auto-triggers it)
claude --plugin-dir /ABS/PATH/TO/tex

# b) Just the MCP server, available in THIS project
claude mcp add tex-browse-use -- node /ABS/PATH/TO/tex/mcp/server.mjs

# c) Available in EVERY repo, persistently:
claude mcp add --scope user --env TEX_ENGINE_URL=http://127.0.0.1:18802 \
  tex-browse-use -- node /ABS/PATH/TO/tex/mcp/server.mjs

Then in Claude Code: run tex_health to confirm, and ask for a browser task in plain language.

Stop the engine with scripts/tex-down.sh. Full setup + troubleshooting in INSTALL.md.

The tools

Verified end-to-end: task "Report the main heading then say AUFGABE ERLEDIGT" on https://example.com → completed in ~7s (gear 1, ~3.4k tokens), returning The main heading is "Example Domain". AUFGABE ERLEDIGT. The full path Claude Code → MCP → engine → Playwright → LLM works.

LLM providers

Pick one in .env (LLM_PROVIDER; auto-detects if unset):

• anthropic — recommended for local. Set ANTHROPIC_API_KEY; model defaults to claude-sonnet-4-6 (computer-use capable). No AWS needed.

• bedrock — for EU / data-residency. Set AWS creds + BEDROCK_MODEL=eu.anthropic.claude-sonnet-4-6.

The MCP tools are repo-wide, but they all talk to one local engine on :18802 — start it once, it serves every project.

What's in this repo

tex/
├── stack/          the engine — Node/Hono + Playwright, vision/DOM/stealth tiers
│   └── src/        server, agent loops, smart-router, providers, vault, skills, verifier
├── mcp/            the MCP server (server.mjs) — browse_use + tex_health
├── skills/         the browse-use skill (tells Claude when to reach for it)
├── scripts/        tex-up.sh / tex-down.sh
├── .claude-plugin/ plugin manifest
├── .mcp.json       registers the MCP server (${CLAUDE_PLUGIN_ROOT})
└── examples/       example app "manifest" (advanced: declarative per-app flows)

Engine runs with no build step — tsx executes the TypeScript directly. On a Linux server the full stack adds stealth (:18803), a DOM gateway (:18804), an a11y tier (:18805) and Xvfb; on macOS the engine + headless Playwright is all you need.

Status

Verified: engine boots on macOS (Node 25 + tsx); Postgres-optional works; the full MCP → engine → browser → LLM path verified end-to-end against AWS Bedrock.

Not yet verified: the direct Anthropic-API provider is implemented and boots, but hasn't been exercised with a live key — smoke-test it with yours. The Linux-only stealth/gateway tiers weren't run on macOS.

Security

TEX drives a real browser and can hold logins — read SECURITY.md before pointing it at anything sensitive. Short version: the engine binds to 127.0.0.1 and is unauthenticated (keep it on localhost); credentials are stored AES-256-GCM encrypted and are sent to your LLM provider at run time; this repo ships a clean history with no secrets.

Credits & license

Released under the MIT License.