Feature
Every assistant can search the web now. With Pinchy, you control which agents get it, see every search in the audit trail, and keep the query inside the company.
Searching the web stopped being the hard part. Who's holding the controls is.
When your team uses a public assistant to "look something up," three questions get answered by accident, not on purpose. Who can reach the open web? Anyone using it. What was searched? No one can say afterward. And does the query leave the company? Yes — sometimes carrying internal context — on its way to a third-party SaaS.
Pinchy gives the same capability and lets you answer those three questions on purpose. Web search is a power your agents can have: granted deliberately, logged completely, and run from your own infrastructure.
Most of your agents shouldn't touch the internet. A few are far more useful when they can. Web search is how you draw that line cleanly, one agent at a time. Here's what a small team actually builds with it.
Point it at your rivals' sites and ask "what changed on their pricing page this week?" It pulls the live pages, summarizes the diff, and cites every source. No more manual tab-checking on Monday mornings.
"Find the three most-cited sources on X and pull the headline numbers." The agent searches, opens the results, and brings back an answer with links — so the next person can verify it, not just trust it.
Before drafting a reply, it checks the current version of your public docs or a vendor's changelog — so it answers from today's reality, not a model trained last year. A human still approves the send.
The pattern is always the same: the agent that needs the web gets it, and the invoice bot stays offline. You decide, per agent.
Web search is a Power Tool in Pinchy's allow-list. A new agent has no web access at all. It can't reach the internet until an admin turns it on — for that one agent.
So the line is explicit, not accidental. The research agent gets the web. The invoice bot reading your ledger does not — no prompt can talk it into a tool it was never granted. A prompt that says "please don't browse the web" isn't a boundary; the agent simply has no web tool until you grant it.
It's one toggle in the agent's permissions, next to its other tools and its directory scope. Enable it, save, done. Made a mistake? Toggling it back off is the same one click.
"Which agent searched for what, and when?" should have an answer. With Pinchy it always does.
Every web search and every page fetch writes an entry to the tamper-evident audit trail — the same HMAC-signed, append-only log that records every other action on the platform. Web access is not a blind spot.
Each row is signed independently, so altering an entry fails its own signature and shows up when the trail is verified — and the log is append-only, so rows can't be quietly deleted. Export it to your SIEM like any other event. When a CISO asks what your agents do on the open web, the answer is a record you can pull up, not a shrug.
"ChatGPT can search too" is true. But to do it, your prompt and the context around it leave the company to be processed by someone else's assistant. For a lot of teams, that's the whole problem.
Pinchy is self-hosted. Your instance performs the search and reads the results. The query and the page content don't flow through a third-party assistant SaaS on the way. Pair it with a local model and the prompt that drove the search never leaves your network either.
An enabled agent decides it needs current information.
Your Pinchy instance runs the search and fetches the pages.
The agent reasons over the results, and the action lands in the audit log.
30-minute demo. We'll enable web search on one agent, leave the others offline, and show you the audit entry the moment it searches.
FAQ
Web search is a Power Tool in Pinchy's allow-list, granted per agent and default-deny. A newly created agent has no web access at all — it cannot reach the internet until an admin enables web search for that specific agent. The research agent can be given the web while the invoice bot stays offline. There is no global on switch; the decision is made one agent at a time, and there's no prompt to bypass because the agent simply has no web tool until it's been granted.
Yes. Every web search and every page fetch writes an entry to Pinchy's tamper-evident, HMAC-signed, append-only audit trail — the same log that records every other action on the platform. Each entry records which agent searched, what it searched for, and when. Because each entry is independently signed, an altered record fails signature verification and is detectable, and you can export the events to your SIEM like any other audit row.
No third-party assistant sees it. Pinchy is self-hosted, so your own instance performs the search and reads the results rather than routing the query through an external assistant SaaS. That's the difference from using a public assistant, where your prompt and its surrounding context leave the company to be processed. If you pair Pinchy with a local model via Ollama, the prompt that drove the search stays inside your network too.