vezir 0.22.0

Self-hosted team intelligence. Record a meeting on any device; Vezir gives you back a diarized transcript, an AI summary, and a PDF — processed on your own GPU server and synced into a private team archive you control. Sign in with Nostr or Google.

Vezir wraps millet (the transcription/diarization/summarization pipeline) and turns it into a multi-user, multi-team service: a scribe records on their laptop or phone, the audio uploads to a central GPU box, and the team gets back labeled transcripts and summaries — with speakers resolved to GitHub handles.

Status

Alpha (0.21.0). Built for small teams that want meeting audio to stay inside their own infrastructure: one GPU server (Linux/CUDA or Apple Silicon) reachable over ordinary HTTPS.

Recent highlights (full history in CHANGELOG.md):

• Screen recordings are summarized from the screen (0.21.0). For a video session using the iteration-plan template, the summary is deferred until after cue frames are extracted, then generated once with the frames attached — so the plan reports what is visibly wrong, not just what the narrator said. Needs millet-pipeline >= 0.20.1.
• Summary attestation (0.20.0). Every job records <backend>/<model> in summary_provenance; the TUI states it in the session detail and badges only the exception. The preset axis is deprecated.
• Video upload + templates (0.18.0/0.19.0). vezir upload demo.mp4 --template iteration-plan, TUI video import (ctrl+u), and cue frames extracted from the recording.
• AI harness integration (0.15.0+). vezir mcp — a read-only MCP server (optional [mcp] extra) — plus vezir ctx <id-or-title> for a one-shot context doc on stdout. See AI harness integration (MCP).
• Identity sign-in. Members sign in with Nostr (a remote signer like Amber via NIP-46, or the NIP-55 Android intent flow) or with Google (@workspace-domain accounts via the OAuth device grant). No key or password touches the client. vzr_ bearer tokens are retained for machine/CI use.
• Multi-team by membership (0.7.0). A token/identity is a person, not a team; scope is supplied per-request via X-Team-Id and validated against a memberships table. The TUI/Android auto-discover your teams.
• Public-access front. A small VPS terminates nothing — it WireGuard-forwards TLS to the server, which keeps the cert. Works from CGNAT / IPv6-only links with no per-client VPN.

The JSON-only API (no web dashboard since 0.7.0) is consumed by the TUI, the Android app, and the CLI. Speaker labeling happens in the TUI (open vezir tui → Sessions → press l on the row) or in the Android app.

Linux and macOS (Apple Silicon) laptop clients and an Android client are supported.

Requires millet-pipeline >= 0.20.1 (pinned by the [server] extra; also probed at runtime).

Sign-in & access

Members authenticate with their own identity; an admin authorizes it once.

# Admin, on the server — authorize an identity (one of):
vezir npub add   --npub npub1…             --github <handle> --label "<who>"
vezir google add --email them@blinkbtc.com --github <handle> --label "<who>"
# …and grant team scope (one handle covers every team they join):
vezir team add-member --team <slug> --role scribe --github <handle>

# Member, on their laptop:
export VEZIR_URL=https://your-vezir-host
vezir login --team <slug>                  # Nostr (remote signer / Amber)
vezir login --method google --team <slug>  # Google (@workspace-domain)

vezir login stores a rotating session (a short access JWT + a refresh token) in ~/.config/vezir/teams.json; the client uses the access JWT as Authorization: Bearer and refreshes transparently on expiry — no re-login needed. vezir logout revokes the session.

Production note: set VEZIR_PUBLIC_URL=https://your-host on the server so NIP-98 login-URL verification is pinned to a fixed base (not reconstructed from request headers).

Architecture

[Scribe laptop / phone]                 [GPU server]
  vezir tui / scribe   ──HTTPS──▶   vezir serve (FastAPI, 127.0.0.1)
   (record, list,                     │  fronted by Caddy (TLS terminates here)
    label, view,                      │
    pull artifacts)                   │   identity sign-in:
                                      │     NIP-46 / NIP-55 (nostr)  ─┐
  vezir-android        ──HTTPS──▶     │     Google device grant      ─┴─▶ session JWT
   (record + sign in)                 │     vzr_ tokens (machine/CI)
                                      │
                            ┌─ public VPS front (optional) ─┐
   any network ────────────┤  WireGuard + nftables TLS-     │
   (incl. CGNAT/IPv6)       │  passthrough → server :443    │
                            └───────────────────────────────┘
                                      │
                                      ├── sqlite job queue (per-team)
                                      ▼
                                    worker  ── HOME-shim ──▶ millet
                                      │   (per-team voiceprints + sync)
                                      ▼
                                    millet transcribe / label --auto / sync
                                      └──▶ private git repo (per-team)

Millet runs as an unmodified subprocess via a per-job HOME shim that exposes per-team voiceprints + sync config. Vezir owns the job queue, per-team voiceprint DBs, team roster/memberships, and auth.

Clients

All desktop clients resolve credentials from a vezir login session (stored per-team in ~/.config/vezir/teams.json), or from VEZIR_URL+VEZIR_TOKEN for machine/CI. The TUI/Android auto-discover every team you belong to from /api/me.

AI harness integration (MCP)

Pull meeting context — session lists, summaries, transcripts — straight into an AI coding harness (opencode, Claude Code, …) instead of manually running vezir pull and pasting paths. Both entry points are read-only, reuse the same credentials as vezir pull (teams.json / VEZIR_URL / VEZIR_TOKEN / VEZIR_TEAM_ID), and need no server-side changes.

vezir mcp — a stdio Model Context Protocol server (optional [mcp] extra):

pip install 'vezir[mcp]'

Wire it into opencode (~/.config/opencode/opencode.json):

{ "mcp": { "vezir": { "type": "local", "command": ["vezir", "mcp"] } } }

It exposes six read-only tools:

vezir ctx <id-or-title> — a one-shot alternative for any harness (no extra needed; base install). It pulls the session (unless --no-pull) and prints a single context document (header + summary + transcript) to stdout:

vezir ctx "brainstorm phoenix" | opencode run "summarize the decisions"
opencode run "$(vezir ctx 01M0TS2SWWD15JT0VQHNDREFKH)"   # inline as prompt
vezir ctx 01M0TS2SWWD15JT0VQHNDREFKH --path              # just the artifacts dir

Summarization: private by default

Every summary is produced by a private backend — a hardware-attested Tinfoil TEE, or fully local Ollama. millet-pipeline 0.19.0 removed the cloud backends (Claude Max, OpenRouter, generic OpenAI) after a blind evaluation found the TEE model beat Sonnet 4.6 on both precision and recall in every language tested, so routing meeting content through a provider that can read it bought nothing. Method and numbers: the evaluation (10 meetings, blind, two independent in-TEE judges, 80 verdicts).

Attestation

The worker records <backend>/<model> on every job in jobs.summary_provenance, read from millet's .summary.meta.json sidecar. The TUI shows it in the session detail:

  summary: tinfoil/glm-5-3-flash (hardware-attested TEE)

The session list badges only the exception — a yellow · unattested when a summary demonstrably did not come from a TEE (a pre-0.19.0 session, or a local Ollama fallback). A positive badge on every row would appear everywhere and stop being read. Sessions with unknown provenance (predating the column) are not badged: absence of evidence isn't evidence of absence.

Screen recordings are summarized from the screen

Upload an MP4/MOV with the iteration-plan template and Vezir samples one cue frame per transcript timestamp, then summarizes with the frames attached — so the plan can report a misrendered value or a mislabeled control, not only what the narrator said aloud.

vezir upload ./walkthrough.mp4 --title "glow 1.1 walkthrough" \
  --template iteration-plan

Ordering matters and is why this is not a single pass: the frames are sampled from transcript timestamps, so they cannot exist while transcription is running — which is when the summary used to be produced. For these sessions the worker passes --no-summarize, extracts frames, then generates the summary once. Every other session keeps the single-pass flow.

Only glm-5-3-flash is vision-capable in millet's allowlist; a summary that falls back to a sibling model degrades to text-only rather than failing. See the case study for what this does and does not buy — it is an n=1 case study, not an evaluation.

Presets (deprecated)

Presets used to select between backends with different privacy/quality tradeoffs. With only private backends left there is nothing to choose, so high-quality, confidential and alternative are now aliases for the same default and will be removed in 0.22.0. They are still accepted — stored jobs and shipped Android builds send them.

A requested preset still pins the backend: the server does not silently fall back, so it either succeeds or fails loudly. When millet does serve a summary from a fallback, the session records summary_fallback and the TUI shows a · fallback badge.

Privacy toggles (per upload)

CLI: --auto-label/--no-auto-label, --sync/--no-sync, --personal on scribe and upload. Server-side kill switches: VEZIR_SKIP_SYNC=1 (global sync off), VEZIR_DELETE_AUDIO=1 (drop audio after artifacts).

Install profiles

The base install uses millet-record (capture only); [server] adds millet-pipeline for transcription/diarization/summarization (plus mlx-whisper on Apple Silicon via a PEP 508 marker for the MLX ASR backend).

Quick start — server

git clone https://github.com/pretyflaco/vezir.git && cd vezir
pip install --user -e '.[server]'

mkdir -p ~/vezir-data
vezir voiceprints seed --from ~/.config/meet/speaker_profiles.json   # optional

vezir team create --id myteam --name "My Team"
vezir team set-sync --id myteam --remote https://github.com/yourorg/meetings.git  # optional

# Authorize yourself + grant scope (identity sign-in):
vezir npub add --npub npub1… --github you --admin --label "laptop"
vezir team add-member --team myteam --role admin --github you

export VEZIR_PUBLIC_URL=https://your-vezir-host   # recommended in prod
vezir serve                                       # binds 127.0.0.1:8000; front with Caddy

TLS via Caddy

cd infra/caddy && ./install-caddy.sh
# edit the Caddyfile for your hostnames, then:
sudo systemctl enable --now caddy

For a public-access deployment (clients on any network, incl. CGNAT), see infra/vps/ — a VPS WireGuard-forwards :443 to the server, which terminates TLS (the VPS sees only ciphertext).

Quick start — scribe client

pip install --user 'vezir[tui]'
export VEZIR_URL=https://your-vezir-host

vezir login --team myteam                 # Nostr / Amber
# or: vezir login --method google --team myteam

vezir tui                                 # record + browse + label
vezir scribe --title "team sync"          # CLI record (p = pause; Ctrl+C = stop)
vezir upload ./recording.ogg --title "…"  # existing file (resumable)
vezir pull                                # artifacts for meetings others recorded
vezir doctor                              # diagnose creds / connectivity / certs

After upload, artifacts (summary, transcript, PDF) auto-download into ~/vezir-meetings/<team>/meeting-…/. Standalone uploads accept .wav/.ogg/.mp3 and .mp4/.mov.

Meeting attachments (0.13.0)

vezir scribe prints a fixed staging folder — ~/vezir-attachments/ — when recording starts. Drop slides, agendas, screenshots or PDFs in there while the meeting runs; when recording stops, scribe lists what it found and waits for Enter as a last chance to add more (skipped without a TTY, or with --no-pause). The files upload with the meeting, then move into that recording's own attachments/ folder so the staging folder is empty for the next meeting.

vezir tui does the same: the record screen shows the folder and how many files are staged, and prompts with the list when recording stops.

Attachments show up in the TUI detail screen alongside the artifacts, are fetched by vezir pull into <meeting>/attachments/, and sync into the team's git archive under the meeting folder, names intact. They are not fed to summarization.

macOS (Apple Silicon) scribe

pip install vezir pulls millet-record, whose macOS wheel ships a Swift sidecar that captures mic + system audio via native APIs (no virtual drivers). Grant both Microphone and System Audio Recording to your terminal app; verify with millet check. The server does the heavy lifting.

Environment variables

Performance (rough, 1h audio)

Runtime dirs are created 0700, sensitive files 0600; the systemd unit sets UMask=0077.

License

MIT — see LICENSE.