Works inside Claude Code, Cursor and Codex
Your AI can use websites.
It just cannot remember how.
Cairn is the part that remembers. It learns a site once, then walks it again in a single step.
Invoices
SeptemberAugustJulyIntroducing Cairn
Ask your AI to pull a report off a website and it starts from nothing. It opens the page, reads the whole thing, works out where to click, and slowly gets there. Tomorrow you ask for the same report. It does the same work again. Same reading, same guessing, same wait, same bill.
Cairn gives it a memory of the trail. The first walk is slow, because nobody has been there yet. Cairn leaves markers along the way. Every walk after that follows the markers instead of the map. And when the path moves, Cairn moves the marker and keeps going.
The boring work
Anything your AI does on a website,
it only has to learn once.
None of it is clever. All of it is repeated. That is exactly the work worth remembering.
- 1open the billing page
- 2sign in
- 3go to Invoices
- 4pick this month
- 5open the invoice
- 6download the PDF
Learns once. The first time, your AI walks the site through Cairn, and Cairn writes down the way.
cairn_run · billing.acme.com
7 steps replayed from memory
Runs in one step. Every time after, it follows what it wrote. No reading the page, no working it out again.
one step rewritten · six steps untouched
Notices when it breaks. A moved button stops the replay. Cairn hands your AI that one step, then saves the answer.
The second run is where you get your time back.
The same task, on the same site, five mornings in a row. Cairn walks it once on Monday and writes down the way. On Thursday the portal was redesigned and it repaired the one step that moved. These are measured numbers — package/benchmark.py prints them, and you can run it. See how it works
Memory that lasts
Close your editor. It still remembers.
Most AI memory disappears the moment the session ends. Cairn keeps the trail on your own machine, so tomorrow starts where today stopped.
It is a file, not a chat
What Cairn learns is written to your disk, not held inside a conversation.
Quit anything you like
Close the terminal, restart the machine, come back next week.
Ask again, it just goes
No warm-up, no re-explaining, no pasting yesterday's notes back in.
Sites change
When the site changes, it does not start over.
A saved playbook usually snaps the first time a button moves. Cairn checks every step, so it notices — then only the step that moved goes back to your AI, and the answer is saved for next time.
Every step is checked
Cairn does not click and hope. After each step it looks at whether the page actually changed the way it should have.
Only the broken step is redone
One button moved? Cairn hands your AI that one step to work out, writes the answer down, and the rest of the trail stays exactly as it was.
✓steps 1–4 replayed from memory
!step 5 did not land — the page did not change
cairn handed me step 5, and nothing else
✓“Download” is now “Get PDF”
✓cairn saved the fix to the trail
✓steps 6–7 replayed from memory
next run: 7 of 7 from memory, nothing to work out
Yours, on your terms
It runs on your machine.
memory
~/.sibyl-memory/memory.db
on this computerStays on your machine
Everything Cairn learns is written to one file on your own disk. Nothing is uploaded anywhere.
Uses the AI you already have
Claude Code, Cursor, Codex. Cairn is a tool they pick up, not another assistant to learn.
API key
No key, no account
Nothing to sign up for. A repeat run uses no model at all, so it costs you nothing.
$ cairn forget --site billing.acme.com
archived. 7 steps forgotten.
Forget on command
One command and Cairn forgets a site. The trail is archived, and replay has nothing left to follow.
Two agents, two machines
A trail one agent learned, another can buy.
A cairn is a pile of stones one hiker leaves for the next one. Learning a site costs real time and real calls — so the agent that walks it first can leave the route behind, and be paid by everyone who follows.
Left for the next agent, free
Two agents on one machine can hand a trail over for nothing. The second one inherits a working, self-checking route for a site it has never opened.
Or sold to one across the internet
Different machines share nothing but a network, and nobody publishes for strangers unless it is worth their while. So a shop can ask for a cent, over HTTP, in USDC on Base.
What travels is the route, never your account
Anything typed into a field is stripped before it leaves. A bought sign-in step arrives asking the buyer for their own credentials.
Browsing a shop is free — you have to see what you are buying. The trail itself sits behind x402, the HTTP standard for machine-to-machine payments. The payment goes on chain. The trail never does.
bob has never opened this site
GET /trails/posthog.com — free to look
✓1 trail · $0.01 · 4 clean runs · first walked by alice
GET /trails/posthog.com/weekly-active
!402 Payment Required
✓signed, settled on Base Sepolia
✓200 — the trail, and a receipt
cairn run --site posthog.com
✓1 call · 0 model calls · the answer
bob was taught nothing. He paid a cent for what alice already knew.
IN THE EDITOR YOU ALREADY USE
Leave a marker on the trail.
Your AI stops starting over on every site it visits.
Then playwright install chromium for the browser, and claude mcp add cairn -- cairn-mcp in the folder you work in. No API key, no account. The README has the same lines for Cursor and Codex.