Liaison calls customer support so you don’t have to.

Describe the problem in plain language. Liaison writes an inspectable plan, waits for your approval, makes exactly one call, and narrates the whole thing back to you as text. Nothing important ever exists only in audio.

It runs on your machine, against your own Twilio and OpenAI accounts. No vendor holds your account details, your recordings, or the authority to agree to things for you.
The model proposes; deterministic code decides. Purchases, passwords, one-time codes, SSNs, and card numbers are refused in every mode — there is no path that approves them.
One call per approval. Approving mints a short-lived code bound to that exact plan, number, and mode. Edit anything and the code stops working.
Every claim in the final report cites an exact transcript quote, or it doesn’t appear. “We’ll look into it” is never reported as resolved.

View the source on GitHub Try it in 60 seconds

$ git clone https://github.com/kaushika05/liaison && npm ci && npm run dev

Runs with zero API keys in mock mode · No project telemetry · Your deployment, your provider accounts

The whole workflow, no credentials required

Liaison handling a billing dispute: a plain-language message becomes an inspectable plan with an authority boundary, approval mints a one-time call code, the call runs while a hold is resolved with an A/B/C choice, a material offer is refused because it conflicts with an approved conditional rule, and the outcome cites an exact transcript quote. — A real run in simulator mode: one message becomes a plan you can read, approval mints a single-use code, and a retention offer is **refused because it conflicts with a rule you already approved** — the agent has no button to accept it. The final report quotes the transcript line it came from.