Bring Your Own Agent: Real Logins with Notte's Hosted Browser
Use any agent to run a real browser task behind login: install the Notte browser skill, authenticate the CLI, keep secrets out of the prompt, replay the session, and shut it down cleanly.
Most agent demos run against public forms. Fill in some fields, submit, done. That proves the agent can click buttons, not that it can do your actual job.
The work people want automated lives behind a login. Login is where things get complicated: the agent needs credentials it shouldn't see in plaintext, a browser that persists cookies across steps, some way to recover when it misclicks, and a record of what happened. Most agents skip all of this. It's browser infrastructure, not reasoning, and agents are only good at one of those.
This post shows how to plug Hermes, Nanobot, OpenClaw, PicoClaw, or your own agent into Notte so it gets a hosted browser, credential vault, and session replay. No changes to the agent.
Setup
The whole setup is one prompt. Pick the tab for your agent, copy it, and run it.
Demo: Vault login, authenticated scrape
The agent starts a Notte hosted browser session, navigates to a site that requires login, uses a saved Vault credential to authenticate, then scrapes data that only appears after login, like the authenticated user's applications.
What happened
The agent starts a hosted browser session, goes to the website, uses the Vault to log in, and uses notte page scrape to extract the logged-in user's applications.
The Vault credential never enters the agent's context. The agent tells Notte to start a session, Notte handles the login with the stored password, and the agent only sees the page after authentication. Swap the agent and the flow stays the same. Credential isolation and the browser session are on Notte's side.
Every session is replayable from the Console. Every page load, every click, every retry. If something breaks midway, scrub to the failure and see what the browser actually saw.
Live watching and replay
The session URL shows up in the agent's response. Open it and watch the run live. If the agent misclicks, you see it happen. Tell the agent in the same chat and it corrects.
For after-the-fact debugging, you replay. Notte records DOM snapshots, network calls, and every interaction. Post-mortems turn into a timeline you scrub through instead of logs you grep.
Why this works
Running a browser behind login is a lot of moving parts: the browser itself, proxy and fingerprinting, credential storage, session cookies, replay recording. Most agents don't want to own any of that. They're built to reason and call tools.
Notte handles the browser side. The agent sends CLI commands and gets JSON back. Hermes in Telegram, PicoClaw on a Raspberry Pi, your own script. They all use the same commands, same vault, same replay. None of them need to run their own browser.
Try it
Pick a site you have credentials for. Add the password to a Notte vault, start a session with the vault attached, and watch the run.
Free API key at console.notte.cc.