Silent swaps
A new model, edited instructions, an extra tool. From the outside the agent looks exactly the same.
On-chain identity anchor for AI agents
The fixed point
for AI agents.
Agents change their models, instructions, tools and keys. pivot keeps who they are anchored, and every change on a verifiable record.
model swapped◆prompt edited◆key rotated◆tool added◆permissions raised◆operator changed◆every change, on the record◆
model swapped◆prompt edited◆key rotated◆tool added◆permissions raised◆operator changed◆every change, on the record◆
The problem
The name stays. Everything else moves.
An agent can be swapped, edited or impersonated without anyone noticing. Today there's no fixed point to check it against.
Borrowed names
Anyone can launch a copy with the same name and avatar. Users can't tell which one is real.
Actions without context
When an agent signs something, nobody can say which configuration was running at that moment.
How pivot works
One fixed point. Every turn recorded.
Anchor
Give the agent a permanent identity: an ID, the operator who controls it and the key it signs with.
Commit
Hash what the agent is: model, instructions, tools, permissions. The hashes are public; the details stay yours.
Turn
Each change becomes a new version linked to the one before. The history only grows; nothing is overwritten.
Verify
Anyone can check which version signed an action, and whether the chain behind it is intact.
Playground
Anchor an agent. Right here.
This runs in your browser with real cryptography: SHA-256 commitments, Ed25519 signatures and a hash-linked version chain. On-chain anchoring is coming; the record format is the same.
Version chain
- No versions yet. Anchor the agent to create v1.
Sign an action
Signed with the current version's keyVerify
Don't trust the agent. Check it.
Paste a signed action (sign one in the playground; it fills in automatically). Change a single character to see it fail.
Enter a configuration. pivot hashes it and finds the version it matches, without the config ever being published.
Recompute every version hash and every link. Then try tampering with an old version and run it again.
The record
Every version is a fixed point.
Together they draw the agent.
The constellation on the right is your playground agent. Publish more versions and it grows.
Append-onlyhistory that only grows
1 checkto see who signed an action
Hash-onlyconfigs proven, never exposed
Where it matters
Anywhere an agent acts for someone.
Trading agents
Prove which strategy version placed every trade, and show holders when the rules change.
DAOs and treasuries
Only let agents with an approved version touch funds or vote.
Agent marketplaces
List agents with a history buyers can check, not just a name and a promise.
Assistants people talk to
Let users confirm they're talking to the real agent, on the version they expect.
Agent-to-agent deals
Agents check each other's anchors before paying, trusting or delegating.
Audits
Reconstruct exactly what an agent was at any moment, from an append-only record.
Questions
Good to know
Is the playground on-chain?
Not yet. It uses the same record format and real cryptography, but everything stays in your browser. On-chain anchoring is the next step.
Do my agent's instructions become public?
No. Only hashes are anchored. You can keep the details private and still prove later that a version used exactly that config.
What if someone changes the agent without publishing a version?
They can, but it shows. Its actions won't verify against the current version, and anyone checking sees the mismatch.
What happens when a key leaks?
The operator rotates it. The rotation becomes a new version, so old signatures are judged by when they were made.
Can old versions be edited?
No. Each version includes the hash of the one before, so changing the past breaks every link after it. Try it in the Verify section.