Why this tool?

When building software with AI it is preferable to limit the context of a given agent to the project it is working on. That eliminates the risk of hallucinations and context drift.

But sometimes you need an agent from another project to give you contextual feedback about a specific feature, API, configuration, or strategy. An example might be asking another agent how OAuth was implemented for end-to-end on-behalf-of flow, so that a new agent can integrate similar or identical functionality in its own project without needing to relearn or rediscover the same bugs — burning tokens on something another agent has already solved.

codependence.ai lets the client session prompt a peer and learn from that peer’s memory, context, and repository. You do not need to leave the confines of a project directory to explore another project, or clone another repo to scan its codebase, both of which burn tokens rapidly. Instead, just ask the peer.

Is this safe?

A question is text written on someone else’s machine, and it is read by a Claude Code session that has tools. That is a real risk, so here is exactly what is done about it, and what is not.

• The peer answers read-only. A forked session runs with Read, Grep and Glob and nothing else. It cannot run a command, edit a file, or reach the network, whatever the question asks for.
• Answering never touches your live session. The peer is a fork, so a question cannot interrupt, alter, or pollute the work you have open.
• The question arrives labelled as untrusted. It is wrapped in instructions telling the peer it is data, not direction, and not to act on anything inside it.
• Accounts cannot reach each other. Peers and tokens live in per-owner storage that is physically separate, not rows filtered by a query somebody could forget to write.
• A question and its answer are visible to two people: whoever asked, and whoever owns the peer that answered. Nobody else, ever.
• Tokens are per machine, stored only as hashes, and revocable one at a time. Browser sessions rotate, and a stolen cookie replayed after rotation kills the whole session lineage rather than working quietly until it expires.
• The bridge opens no ports. It dials out and holds one connection. Nothing can reach your machine from the internet because of it.

What this does not claim. Read-only tools are a real boundary; the untrusted-data wrapper is a mitigation, not a guarantee. A peer can quote what it reads, so only publish sessions whose repositories you are willing to have quoted back. Today you can only ask your own peers.

Contact

john@threatsr.us