vexaVexa会议助手

Chat-oriented interactions

Speak directly to the user as in a natural chat. Do not output internal reasoning, plan summaries, or procedural notes ("I need to...", "According to the skill...", "I will inform..."). Reply with only what you would say to the user — conversational, warm, and to the point.

Plain text formatting for meeting chat

When sending messages to meeting chat (Google Meet, Teams, etc.), always use plain text — no markdown. These chat surfaces do not render markdown, so asterisks, hash headers, backticks, and bullet dashes appear as ugly raw characters. Use line breaks and spacing for structure instead.

Use the bundled CLI:

• - INLINECODE0
• INLINECODE1
• INLINECODE2
• INLINECODE3

Environment

Set for the OpenClaw process (or shell):

• - VEXA_API_KEY (required)
• INLINECODE5 (optional env override — takes priority over endpoint config)

Endpoint switching (prod / local dev)

The CLI supports named endpoints stored in secrets/vexa-endpoints.json. Default endpoints: prod (https://api.cloud.vexa.ai) and local (http://localhost:8000).

• - List endpoints & active env: INLINECODE9
• Switch to local dev: INLINECODE10
• Switch to prod: INLINECODE11
• Add custom endpoint: INLINECODE12
• Remove endpoint: INLINECODE13

Priority: VEXA_BASE_URL env var > vexa-endpoints.json active endpoint > default prod URL.

Secrets-first check (before any onboarding)

Always check first: INLINECODE16

• - If output is {"secrets_ok": true} → secrets exist. Skip onboarding. Use normal flows.
• If output is {"secrets_ok": false} → run onboarding. Load references/onboarding-flow.md.

Webhook — proactive setup: Whenever secrets exist and you're working with Vexa, run node skills/vexa/scripts/onboard.mjs --check-webhook. If webhook_configured is false, advise adding the vexa mapping (see references/webhook-setup.md). For onboarding-only details (webhook requirements, pipeline validation, mock webhook), load references/onboarding-flow.md.

When secrets are missing: Ask them to get their key from https://vexa.ai/dashboard/api-keys and paste it in chat — you write it to skills/vexa/secrets/vexa.env. Alternatively, share the path and format so they can add it manually (better for security if they prefer not to paste the key).

Secrets location: skills/vexa/secrets/ holds env files and vexa-state.json. This dir is gitignored. When publishing the skill to ClawHub, ensure secrets/ is excluded.

Per-endpoint API keys: The CLI supports separate env files per endpoint: vexa-prod.env, vexa-local.env, etc. When switching endpoints with env:use, the matching vexa-<name>.env is loaded automatically. Falls back to vexa.env if no endpoint-specific file exists.

Non-interactive (for scripting): INLINECODE33

Quick workflows

1) User drops a meeting link → send bot

• - After successfully sending the bot, proactively run --check-webhook. If not configured, offer to set it up so finished meetings auto-trigger reports.
• Parse/normalize link (or pass explicit ID):

• - Start bot directly from URL:

2) Start bot from calendar meeting links

If a calendar tool/skill is available (for example gog):

1. 1. Fetch upcoming events.
2. Extract meeting links (Google Meet/Teams).
3. For each selected event, call bots:start --meeting_url ....
4. Optionally save event title into Vexa metadata:

3) Read transcript during meeting or after meeting

• - Poll current transcript:

• - For near real-time streaming, use Vexa WebSocket API (see references/user-api-guide-notes.md for endpoints and notes).
• After transcript is available, summarize and store key updates.

4) Stop bot

• - INLINECODE43

5) Create meeting report (after meeting finished)

After stopping the bot (or once the meeting has ended and transcript is finalized), create a basic meeting report:

• - INLINECODE44
• or INLINECODE45

Writes to memory/meetings/YYYY-MM-DD-<slug>.md with: meeting info, summary placeholders, key decisions, action items, and full transcript.

6) Get or update the Ultravox voice agent system prompt

The voice agent system prompt controls how the Vexa bot behaves in meetings (personality, language, what it does when triggered). It is stored per-user and applied when the next bot starts.

• - Get current prompt (null = using service default):

• - Set a custom prompt:

• - Reset to service default:

Note: The updated prompt takes effect on the next bot started — it does not affect bots already in a meeting.

Core commands

• - Bot status:

• - Request bot (explicit fields):

• - Update active bot language:

• - List meetings:

• - Update metadata (title/participants/languages/notes):

• - Generate share URL:

• - Set Vexa user webhook URL:

Recordings

• - List recordings:

• - Get a single recording:

• - Delete a recording (destructive):

• - Get download URL for a media file:

• - Get recording config:

• - Update recording config:

Meeting bundle (post-meeting)

Get everything about a meeting in one call — transcript, recordings, share link:

• - INLINECODE63
• INLINECODE64

Options:

• - --segments — include transcript segments (omitted by default to keep output small)
• INLINECODE66 — skip creating a share link
• INLINECODE67 — skip recordings metadata
• INLINECODE68 — resolve download URLs for each recording media file
• INLINECODE69 — share link TTL

Webhook (meeting finished → report)

When Vexa sends a "meeting finished" webhook, the transform (scripts/vexa-transform.mjs) instructs the agent to create a report. See references/webhook-setup.md for hooks mapping config. Requires hooks.transformsDir = workspace root and transform.module = skills/vexa/scripts/vexa-transform.mjs.

OpenClaw ingestion helpers

• - Create basic meeting report (meeting info, transcript, placeholders for summary/decisions/actions):

• - Audit meetings for likely test calls / cleanup candidates:

Platform rules

• - Supported: google_meet, teams, INLINECODE82
• Teams native_meeting_id must be numeric ID only.
• Teams bot join requires passcode (from ?p= in Teams URL).
• Zoom native_meeting_id is 10-11 digit numeric ID. Passcode (?pwd=) is optional.

Deletion safety (strict)

INLINECODE87 purges transcripts and anonymizes data.

Rules:

1. 1. Never call delete without explicit user request for that exact meeting.
2. Verify platform + native_meeting_id first.
3. Prefer non-destructive cleanup (meetings:update) whenever possible.
4. Require guard flag: