@usecarry/cli

Proof-carrying memory for AI agents — from your terminal.

carry gives you gated, receipted, Walrus-verified memory as a command-line tool. Every recall prints an Answer Receipt: what memory was used, whether the agent was allowed to use it, and whether each blob still resolves on Walrus. Revoke a namespace and the gate runs before retrieval — so your data is never fetched, and the receipt proves it.

It shares one on-disk vault (~/.carry/store.json) with the Carry MCP server, so a fact you remember in the terminal is recalled by an MCP agent in your IDE — same vault, same gate, same proof.

Install

npm i -g @usecarry/cli      # or: npx @usecarry/cli <command>

Use

carry seed                                    # load a demo vault (real Walrus testnet blobs)
carry recall "am I allergic to anything?"     # → Answer Receipt: authorized ✓  verified ✓ on Walrus
carry access revoke health                    # flip the retrieval gate
carry recall "am I allergic to anything?"     # → 1 namespace blocked · your data never reached the model
carry remember "Blood type is O negative" --ns health   # store a new fact on Walrus
carry anchor                                  # write the receipt itself to Walrus (resolvable blob)
carry list                                    # every memory + its Walrus ref
carry policy                                  # the allow/deny policy

Config

| Env | Default | | --- | --- | | WALRUS_PUBLISHER | https://publisher.walrus-testnet.walrus.space | | WALRUS_AGGREGATOR | https://aggregator.walrus-testnet.walrus.space | | CARRY_STORE | ~/.carry/store.json |

Works out of the box against Walrus testnet — no keys required.

How it works

The gate is enforced before retrieval: recall filters to allowed namespaces, then does a live aggregator GET on each used blob to prove it resolves. Blocked namespaces are never fetched — the receipt shows them as blocked, not omitted. This is the same engine behind the Carry app and its on-chain carry::access policy on Sui.

MIT © Carry