AI agents can already change calendars, send messages, edit repositories, and move budgets. The uncomfortable part is that those tools rarely agree on what a transaction is. A reasonable plan can half-execute and leave the world in a broken state.
I built Worldtree to put a transaction boundary around consequential agent work.
GPT-5.6 runs in Codex, reads one immutable snapshot, and proposes three genuinely different executable futures through Worldtree's MCP server. Worldtree then takes over: it checks every requested effect against server-owned policy, verifies evidence, derives the actual state changes, and makes the alternatives comparable before anything is committed.
The model is deliberately not the safety authority. A branch can only commit after it has been selected, fully reviewed, simulated against the current revision, and explicitly confirmed. The commit is recorded as one atomic before/after transaction. If the agent tries to replay the same branch after the revision has advanced, Worldtree rejects it as stale.
The demo follows that exact path. An unsafe calendar action tries to send a real email and is rejected without mutating state. GPT-5.6 stages three valid recovery futures. I compare and simulate them, commit one branch, and then prove that its stale replay is blocked in the journal.
The calendar, mail, task, finance, and repository connectors are local simulations. That is intentional: this submission demonstrates the transaction primitive, not production integrations or rollback infrastructure.
Claude Opus 4.8 and Fable 5 were used as bounded, read-only design and judge critics during the final polish pass. Their feedback informed presentation and accessibility review; neither is a runtime dependency or implementation authority.
Codex was the main build environment. It helped me pressure-test the idea, separate probabilistic planning from deterministic authority, implement the zero-dependency engine and MCP bridge, inspect the product in a live browser, and build the release verification path. GPT-5.6 is part of the runtime itself—it authors the three executable futures that Worldtree validates and stages.
Run it with Node.js 20+ using npm start, then open http://127.0.0.1:8787. No package install, account, database, API key, or API billing is required for the deterministic judge path.
Built With
- css
- gpt-5.6-sol
- html
- javascript
- model-context-protocol
- node.js
- openai-codex
- server-sent-events