Tale supports audio transcription two ways, and this tutorial walks through the fully local option. If you only need to summarise ad-hoc recordings, dropping an audio or video file into chat is the simplest path — the platform’s transcription pipeline handles it server-side (see Chat attachments). But for a full meeting-capture workflow where the raw audio never leaves the presenter’s laptop, you pair Tale with a local tool that handles the audio path. Meetily is an MIT-licensed, 100% local meeting recorder that transcribes with Whisper.cpp on-device and sends only the transcript to an LLM for summarisation.
That split matters: raw audio never leaves the endpoint, Whisper runs on the presenter’s laptop, and Tale only ever sees text. Your existing row-level security, audit logging, and governance policies cover the full sensitive path because everything that reaches Tale is a real agent conversation thread.
Pick the server-side flow when convenience matters and you trust your configured transcription provider with the audio; pick this Meetily flow when compliance or network policy requires that audio bytes never cross the endpoint boundary.
What you will build
A meeting flow where Meetily captures and transcribes audio locally, then hands the transcript to a Tale agent that produces a structured summary — participants, decisions, action items — stored as a normal conversation thread under that agent.
Prerequisites
- A Tale instance reachable on HTTPS, with Admin access and at least one agent suitable for summarisation. The system prompt in Step 1 below is a good starting point.
- Meetily installed on the workstation that will record the meeting — see the project’s latest release. macOS and Windows are supported.
Create a dedicated agent so summaries use the model, tone, and format you want. A starting system prompt:
You are the meeting-summary agent.
Input: a raw transcript of a meeting, possibly with imperfect speaker labels.
Output, in Markdown:
1. A one-paragraph summary.
2. Decisions — bullet list, each with the person responsible.
3. Action items — bullet list in the format "Owner — task — due date (if mentioned)".
4. Open questions — bullet list of things that were raised but not resolved.
Rules:
- Do not invent content. If something is unclear, say so.
- Preserve the language of the transcript.
- Never include raw quotes longer than one sentence.
Step 2 — Create a webhook for the agent
Open the agent’s Webhook tab and click Create. Tale generates a URL of the form https://<your-tale-instance>/api/agents/wh/<TOKEN>. Copy it — you’ll paste it into Meetily in Step 4.
Treat the webhook URL like an API key: anyone holding it can invoke this specific agent. Disable or delete the webhook to revoke access.
Step 3 — Install Meetily
Download and install Meetily from the project’s releases page; the project docs at meetily.ai and the GitHub README cover per-OS install steps, including the first-run permission grants for system audio. Verify you can record a short clip and see a transcript appear before moving on — that confirms Whisper is working locally.
Step 4 — Point Meetily at the Tale webhook
In Meetily’s Settings > LLM provider (label varies by version), choose Custom OpenAI-compatible and set:
| Field | Value |
|---|
| Base URL | The webhook URL from Step 2 (e.g. https://<tale>/api/agents/wh/<TOKEN>) |
| API key | Any non-empty value — the URL token is the credential |
| Model | A model ID from the agent’s supportedModels (e.g. openai/gpt-4o). Unrecognised values silently fall back to the default. |
/chat/completions to the Base URL automatically; every summary it generates now flows through the configured Tale agent.
Step 5 — Record and summarise a meeting
Click Start recording at the top of the next meeting. Meetily shows live transcription in a side panel — audio is transcribed on the CPU/GPU of the laptop, nothing is uploaded. When the meeting ends, click Stop and then Generate summary. The transcript is posted to Tale, the agent runs, and the summary appears alongside the transcript in Meetily.
In Tale, the request becomes a real conversation thread under the summarisation agent — visible in the agent’s conversation history, counted against the org’s usage ledger, tagged in the audit log, and governed by the agent’s team and knowledge rules.
Privacy notes
- Audio never crosses the network. Whisper.cpp runs locally.
- The transcript does cross the network — to your Tale instance, over HTTPS, under your reverse proxy and auth. It does not go to any third party.
- Retention follows Tale’s standard rules. If your org retention policy is seven days, the summary thread expires on that schedule; see Governance.
- Access to the summary is scoped by the agent’s Knowledge tab and team rules — standard agent RLS.
- Client system prompt. Any
system message Meetily sends is concatenated after the agent’s own system prompt — agent rules frame identity and output format; Meetily’s prompt adds use-case detail.
Troubleshooting
- Context-window errors on long meetings — the transcript exceeds the model’s input limit. Switch the agent to a model with a larger context, or pre-chunk the transcript in Meetily’s summary settings. See Agent concepts — Model.
- Meetily timed out — Meetily’s client-side timeout is 300 seconds. On slow providers a long transcript can exceed it. Options: switch to a faster model, shorten the transcript, or retry after generation completes (the thread in Tale will still hold the full summary).
- Summaries in the wrong language — either the transcript was mixed-language, or the system prompt did not pin the output language. Tighten the prompt’s “Rules” section.
- Empty or refused summary — inspect the agent’s conversation thread in Tale; the full model response (including any refusal or governance message) appears there.
- 401 Unauthorized — the webhook URL is invalid or the webhook is disabled. Check Settings > agent > Webhook tab; toggle active, or delete and regenerate.
Next
Last modified on April 21, 2026
