Oathra v0.1 launch candidate — give AI agents a phone, and proof of what happened

[FORIFOR]

Text:

Oathra is an open-source (Apache-2.0) runtime for AI agents that make phone calls: call a restaurant, negotiate, book, and return a result you can trust.

The part we care most about is the "trust" part. The result is not an LLM summary. Every field (date, time, party size, price, "confirmed") is anchored to an utterance from the other party, and completion is decided by code — never by asking the model whether it succeeded. Refusals ("19:00 is full, but 19:30 works") never become offers; corrections supersede; hedges ("probably fine") do not confirm; a confirmation goes stale if the deal changes after it was spoken. We fuzz this with 10,000 mutated simulated callees (never-confirm, wrong restate, negate-then-offer, silent hangup, echo trap) and the count of false completions is 0.

npx oathra demo opens a small "arena" where two agents negotiate in your browser with no API key (there is a play mode where you answer the phone yourself). The same runtime drives real calls: Twilio direct media streams, or any SIP trunk (Plivo, custom) through a LiveKit gateway. The voice engine is separate: GPT-Live (full-duplex) was the most natural on our test calls, so it is the default; there is also an OpenAI Realtime engine and a Deepgram + LLM + TTS pipeline.

What is verified today: Twilio calls to a Japanese mobile, a 6-minute GPT-Live call, the evidence engine (0/10,000). What is not: Plivo/custom SIP over PSTN (implemented from provider docs), Telnyx/others, MCP.

Repo: https://github.com/FORIFOR/oathra · Site: https://forifor.github.io/oathra/

Happy to answer anything about the evidence rules, the phone layer, or the latency numbers.