Vexa-ai/vexa v0.12.0

Vexa 0.12 — one species, two kinds, one runtime

on GitHub

0.12 is a complete rebuild of the whole system on two principles — SoC (separation of concerns) and SSoT (single source of truth). Every service is a modular monolith: one-concern modules joined only through sealed contracts CI won't let drift.

🤖 Bots — rebuilt as capture bricks (join, per-platform capture, STT, recording…) that compose three ways: the meeting bot, a desktop host, and an eval harness that replays real captured meetings through the stack. The API is unchanged — api.v1 is sealed hash-equal to the 0.10.x line, machine-checked in CI. Nothing you built breaks.

🧠 Agents — the new kind. Open the Terminal to feel them. Same runtime as the bots — tenant-isolated, designed to scale on Kubernetes.

📚 Workspaces — the wiki your LLM builds: plain markdown, git-versioned, open format (OKF). Personal and shared — a shared graph lives between the people on your daily, updating in real time while you speak.

Try it

git clone https://github.com/Vexa-ai/vexa && cd vexa && make all

(make bot after, to send bots to calls. Prereqs: Docker + make; the first build is the heavy step — 8+ vCPU / 16 GB makes it comfortable.)

Official images: Docker Hub vexaai/*:0.12.0 — the quickstart pulls them; make build still builds everything from source.

What we guarantee — and how each guarantee was verified

Deploy paths:

Claims:

• API continuity — api.v1 sealed (sha256 in contracts.seal.json); CI fails on any drift; the running app is conformance-checked against the seal. Known edge, stated below: two sealed endpoints are not yet served in the open stack.
• Concurrent bots (#478 fixed) — proven by a control/fix A/B on the exact failing image: 2 concurrent bots, the verbatim failure reproduced on the old code and eliminated on this one; per-bot profile dirs; leak sweep with unit proofs.
• Pipeline determinism — a real captured meeting's signal replays byte-identical through the stack (gate:replay); the fixture ships in the repo.
• Sovereign egress — measured on instrumented real calls on Meet and Teams: the bot's only network egress was the meeting platform itself — zero calls home. (Scoped to the bot with self-hosted STT; a measurement you can rerun, not a blanket audit.)
• The eval harness ships (core/meetings/eval) — contract-driven, self-oracling: score your own deployment; don't take our word for it.

Our rule: every claim is proven at the altitude it's made — contracts by hash seals, pipelines by replay of real signal, behavior by live A/B, sovereignty by measured egress. Anything that can't meet its bar goes on the list below instead of in a guarantee. Apache-2.0 + this frozen tag + pinned images make all of it re-checkable by anyone.

Not yet in 0.12 (the honest list — explicitly NOT guaranteed)

• Webhook + scheduling bricks exist but are not yet wired onto the unified HTTP path.
• The old dashboard was replaced by the Terminal; Terminal user docs are thin and coming with the alpha.
• Live bot-control endpoints (config/speak) are sealed in the API but not yet served in the open stack.
• Helm/Kubernetes deploy path: in progress (Compose is the shipped path).
• Teams 1:1 calls: live transcription can gate on the two-party layout (#481) — being reworked; audio capture itself is unaffected.
• Scale beyond "designed to", transcription-quality numbers, and agent counts beyond what we've demoed: not claimed.

Hosted service note

vexa.ai hosted stays on the 0.10.x line until our staged cutover completes its regression gates — hosted users will switch with zero API change. This week's hosted incidents are being handled in the open: root-cause analysis is posted in #483, #490, #491, and each stays open until its fix is verified live.

Community

0.12's module layout is built for contribution: pick one brick, test it in isolation, ship. First-contribution issues are being cut from this week's triage — watch the tracker. Thanks to contributors with PRs in flight — #488, #489, #476, #485, #487 — reviews in progress, <24h response during alpha.