The short answer
Move the record, not a retelling of it. With Zambo added to the AI doing the work, calls run through one endpoint, https://zambo.dev/mcp, and every step is saved outside the session with a verifiable receipt: an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right. At the switch point, package the session with the handover tool. The receiving AI opens the package, checks the receipts behind the steps it will build on, accepts, and continues. The model changes; the job does not restart.
The manual workaround, honestly
The standard switch is a brief: you ask the old AI to summarise itself, paste the summary into the new AI, and patch the gaps as they surface. For a two-step task, that is genuinely fine. The costs show up with size and time. The summary is written by the worker whose work it describes, so it rounds in the worker's favour. Detail decays on every hop: switch twice and the third AI works from a retelling of a retelling. And nothing in the brief can be rechecked, so the new AI either trusts it blindly or redoes steps to be safe, which is the starting over you were trying to avoid. The workaround fails precisely where the job is valuable: long, multi-stage, and worth checking.
The receipt-backed switch, step by step
- Work recorded from the start. Both AIs connect to the same endpoint, so the job's steps land in one record regardless of which model ran them. Starting is free: 20 free calls per day.
- Package at the seam. When the first AI reaches the natural hand-off point, or hits a wall, run the handover tool: the goal, the steps so far with their saved results and receipts, and the next steps, closed into one link.
- The receiver verifies before it builds. The new AI opens the handover, reads the steps as recorded, and checks receipts at the public verify endpoint. Steps that check are inherited; a step that does not check gets re-run, not trusted.
- Accept, then continue on the same trail. Acceptance lands on the record, and the receiver's own steps chain onto it. Anyone reading the finished job later sees both halves and the exact seam where one AI stopped and the other started.
Two directions get their own walk-throughs because the steps differ in each app: ChatGPT to Claude and Claude to ChatGPT. When the switch is forced on you by a limit rather than chosen, start at continuing after a Claude usage limit. When both sides are agents rather than people driving apps, the pattern is agent-to-agent handover, where the receiving agent signs for the package with a tool call.
Why the checks hold up
A handover between two AIs only works if the receiving side can test the record without asking the sender or the platform to vouch for it. The receipts follow AER-1, published as draft-zambo-aer1-26, and the format ships with a public conformance corpus of 126 vectors and 129 checks, runnable by anyone, catalogued on the corpus checks page. Independent implementations have passed those checks, which is what makes the receipt a piece of evidence instead of a badge: the math is open, the test set is open, and the verify endpoint recomputes rather than asserts.
Scope, stated once: a verifiable receipt proves the saved result was not changed since it was saved. It does not prove the result was right, and it does not identify which AI truly ran a step beyond what the producing session recorded. Judge the work from the record; the record's job is to make sure you are judging what actually ran.
Questions, answered
Can I switch from ChatGPT to Claude in the middle of a project?
Yes, and the middle is exactly where it matters. If the work so far ran through the Zambo endpoint, package the session with the handover tool: goal, steps, saved results, receipts, next steps, at one link. Claude opens the link, reads the steps as recorded, verifies the receipts behind the steps it will build on, accepts the handover, and continues. The same move works in the other direction, from Claude to ChatGPT, and between other AIs connected to the same endpoint.
What does the new AI actually receive when I switch?
The job itself, in recorded form: the goal, the sender's summary, the next steps, and every step of the work so far with the tool that ran, the saved result, and a verifiable receipt whose published copy points back to its source receipt. It does not receive a chat transcript to mine or a summary to trust. Anything the sender marked as not checked is marked as not checked in the package, so the receiver knows where its own verification work is.
Will the new AI trust work it did not do itself?
It does not have to trust it; it can check it. Each step's receipt commits to the exact saved bytes of that step's result. The receiving AI recomputes the check at the public verify endpoint and builds only on steps that verify. A receipt proves the saved result was not changed since it was saved; it does not prove the result was right, so the receiver still judges the work, but it judges the work that actually ran instead of a description of it.
What if the two AIs disagree about the state of the work?
Then the record settles the factual half of the disagreement. Both AIs read the same steps and the same receipts, so a claim like "that step already ran and returned this" is checkable instead of arguable. Judgment calls, what to do next, which draft is better, stay judgment calls. What the handover removes is the fog about what has already happened, which is where most mid-task switches actually go wrong.
What does switching AIs mid-task cost?
The handover tool is free, and the switch itself adds no charge: the work continues on the same terms as any call on the endpoint. You get 20 free calls per day, then a $0.99 Day Pass for 24 hours of higher limits, or the Zambo Pass at $29 for 30 days, one payment, $290 for 365 days, and $19 for 30 days for $ZAMBO holders. People pay by card through PayPal; agents pay in USDC on Base through x402.