A
Agiliton
Browser agents for production teams · made in EU

Every major AI model, running in your real Chrome. No third-party subscriptions.

A browser agent that drives the Chrome you’re already logged into — and lets you choose Claude, GPT, Gemini, Grok, Llama, or Mistral on a per-task basis. We deliver the model, the agent runtime, and the MCP context layer through one LiteLLM gateway. Hosted in Helsinki, built on the open stack — every layer on GitHub under MIT or Apache‑2, every byte in EU-only data residency. No Anthropic Pro / Max sub per user. No eight-vendor inference bill. No US data plane.

See it run

You and the agent, watching the same Chrome.

Other agents work in a sandbox you can't see into. SiteBridge runs in your real Chrome — every screenshot, every click, every keystroke shows up in a live trace you can read, intervene on, or hand off to a colleague. Pick a workflow below and press play.

app.hubspot.com / contacts / deal / 84321
CONTACT
John Marsden
VP Eng · Northwind GmbH
john@northwind.eu
ACTIVITY
· May 8 · email · sent quote
· May 9 · call · 18 min
DEAL · #84321
Northwind · annual tier B
€42 000 · close 2026-06
STAGE
Decision Maker Bought-In
PROBABILITY
75%
NEXT STEP
Send revised quote · Tue
Save changes
What the agent saw0 frames
screenshots land here as the agent works
What the agent didundo points: 0

Press ▶ run agent and you'll see every step in plain English — what it read, what it clicked, what it screenshotted. You and the agent watch the same trace.

Safety guard active · 0 clicks verified
Sovereignty by default

Built and hosted in EU. On the open stack.

Every visible competitor has a US opaque-box runtime at the bottom. We built the opposite — every layer open-source, every byte in EU data residency, your security review can reach all of it. Anthropic can spin up a Frankfurt subsidiary; their data plane still terminates in us-east-1.

GDPR / BDSG by construction
EU-only data residency · no SCCs
Self-hostable end-to-end
Open source · auditable on GitHub
Optional on-prem inference
Audit log per tenant
1 vendor
We deliver models + context + MCP through LiteLLM. No Claude Pro/Max sub per user, no eight-vendor inference bill, no “go set up an Anthropic API key” onboarding step.
All models
Anthropic, OpenAI, Google, xAI, Meta, Mistral, MiniMax, Z.ai — every major frontier and open-weights model in one product. Use the right model for each task; switch any time without changing tools.
0.0%
Anthropic’s own post-mitigation prompt-injection rate. One in nine attacks still succeeds. Our pre-click guards + undo snapshot ring are defense-in-depth for exactly that case.
All major AI providers

One product — every model that matters.

Claude beats GPT at one thing. GPT beats Claude at another. Gemini wins on long context. Grok wins on speed. Llama is the only thing that runs in your VPC. With SiteBridge through LiteLLM, you don’t pick one — you get them all, route per task, and try the new release the day it ships.

1. pick a task
2. LiteLLM routes it
prompt
LiteLLM
+ MCP context
MiniMax
low-cost router

Browser navigation: find-button-and-click. Doesn't need reasoning. Route to the cheapest fast model — that's most of agent traffic.

cost / 1M tok$0.30
latency p500.8s
3. force any model

Plus any new model added to LiteLLM the day it lands. Customers don’t sign a new contract or wait for procurement — we add the deployment, they benefit silently.

The wedge

Primitives are commodity. The operations layer isn’t.

Five field leaders, same five verbs: navigate, click, fill, read, screenshot. The primitive layer plateaus in twelve months. Value moves up the stack — to the layer that takes eighteen months of incident data to build.

What none of them ship:

  • Pre-click expectation guards — refuse if dialog text doesn’t match the agent’s intent.
  • Undo snapshot ring buffer — answer “what did the agent do?” after the fact.
  • Cross-session selector learning — don’t rediscover the same button on day forty-seven.
  • Compound wait-aware tools — one round-trip instead of three.
  • Model-agnostic via LiteLLM — route to any provider, route by complexity, route to self-hosted.
  • Real domain integrations — Smartling, n8n, Confluence, Google Docs, Jira.
Try the guard

Pre-click expectation guard, live in your hands.

Type what the agent intends. Type (or pick) the dialog the page shows. Watch the guard accept, refuse, or escalate. Same logic that gates every click in production.

The pre-click guard compares the agent's stated intent to the dialog actually shown. Hidden swaps, prompt-injected destructive verbs, and ambiguous confirms get caught here — before the click goes through.

simulated browser dialog
example.com says

Save changes to comment? Your draft will be preserved.

CancelOK
click allowed
intent and dialog share 3 key terms — consistent
term overlap
75%
matched terms
save, draft, comment
Pre-trained helpers

For the tools you actually use.

Generic browser agents see a page of pixels. SiteBridge ships domain modules that know each tool’s shape, vocabulary, and quirks — and surface only when you’re on that tool. No toolbar clutter, no wrong suggestions, no “please describe your workflow” onboarding.

Atlassian — Jira, Confluence, Bitbucket

Reads boards and tickets, drafts ADF-formatted comments, navigates Confluence whiteboards, recognizes ATCTT/ATATT API tokens so generated-token dialogs round-trip cleanly.

Smartling — translator’s cockpit

Glossary-aware proposals, bulk-apply with TM diff, autofill empty segments, persona & style enforcement, numeric-format checks, lint pass before submit. Built with a real translation team.

n8n — workflow co-pilot

Knows the node catalog. Adds and configures nodes through the n8n API instead of clicking through the canvas. Understands credentials and connection shapes.

Google Docs — long-form fluency

Read structure, replace content blocks, restructure sections — without the copy-paste-back-to-the-doc round-trips that break formatting.

WordPress — content, not curl

Reads and writes posts through the REST API, picks the right Gutenberg block, respects taxonomies and scheduled-publish state.

Chrome — the hardened foundation

Every helper rides on typed selectors, an action log with undo, network capture, screenshot, clipboard, and dialog handling. New tool tomorrow? The foundation is already there.

Roadmap is open: customer asks for tool X, we ship a domain module for tool X. The same week, usually.

Trust by default

Gets smarter every day. Keeps your data off the wire.

The three questions every serious team asks of an agent: does it learn? where does my data go? can other tenants see anything? SiteBridge has an opinionated answer to all three — wired into the runtime, not bolted on.

Self-improving with every conversation

Every successful click, every accepted correction, every glossary fix feeds an on-device learning store. Selectors that worked on day one stay remembered on day forty-seven. Translator preferences stick. Patterns the agent saw once become first-try-correct. None of it round-trips through a vendor — the memory is yours.

Patterns up · data down

The n8n / WordPress / Stripe helpers get sharper for everyone — selectors, tool sequences, recovery strategies pool across tenants with k-anonymity ≥ 5. Form values, IDs, secrets, customer names? Never. A scrubbing pipeline gates every contribution; a per-tenant audit log shows exactly what crossed the wall. We learn the shape of n8n. We don't learn the shape of your workflows.

Secrets stay on your machine

API keys, customer credentials, sensitive values never enter the LLM transcript. They live in an on-device vault with personal end-to-end encryption — the agent sees a placeholder, the browser injects the real value at execution time. We physically can't leak what we can't see, and the audit trail proves it.
The bigger win for modern teams

Your users don’t need a Claude Code sub. We deliver the whole stack.

Claude for Chrome ships the agent — and the bill. Every user needs Anthropic Pro, Max, Team, or Enterprise. Every team needs Anthropic billing in their procurement. SiteBridge through Agiliton’s LiteLLM gateway collapses that into one vendor relationship and one inference bill — and gives you cost, privacy, and failover knobs Anthropic-direct doesn’t.

One bill, one vendor

Procurement signs one contract — with us. Onboarding adds users via SSO, not “create an Anthropic account, generate an API key, paste it here.” Per-user spend, per-team budgets, audit trail in one place.

Cost: complexity-routed

Browser tasks are mostly cheap (find button, click button). LiteLLM auto-routes those to MiniMax M2.5 / GLM 5.1 at fractional cost. Reserve Claude / GPT-5 / Opus for hard reasoning. We capture the spread; customer sees a flat per-seat or metered price.

Failover built in

Anthropic outage? LiteLLM falls back to OpenAI / xAI / Gemini in the same call. The browser fleet doesn’t notice. Claude-for-Chrome customers get a 502 and a tweet from Anthropic Status.

Privacy: route to on-prem

Regulated workload? Route to self-hosted Llama / Qwen via the same MCP. Inference never leaves the VPC. One agent codebase, two compliance postures.
At a glance

The capabilities modern teams actually ask about.

Six rows from the full matrix. Tap through to see all seventeen.

CapabilitySiteBridgeClaudeforChromePlaywrightMCPBrowserMCPDevToolsMCP
Bundled with model + agent runtime (no third-party sub)
Runs inside your real logged-in Chrome1
Works with any LLM provider (no model lock-in)2
Local-only data path possible3
Pre-click safety guards (expectedTarget, viewport check)
Undo log + DOM snapshot ring buffer
Go deeper

Pick the angle that closes your room.