English · Español

Yapr

Talk to a Telegram bot, get a tidy .md back.

    

Yapr turns your voice notes into structured Markdown documents. Walk out of a meeting, a site visit, or just have an idea rattling around: record a voice note, ramble as much as you want, and the bot sends back a .md file with everything organized into clear sections.

It's 100% self-hosted, BYOK (bring your own API key) and open source.

---

How it works

Voice note (Telegram)
  → download the audio
  → transcription (Whisper, via Groq)
  → organized by an LLM according to the selected mode (gpt-oss-120b, via Groq)
  → .md file sent back on Telegram

Everything runs as a single n8n workflow on your own instance.

Modes

Each mode is a different system prompt that defines how the note gets organized. Every chat picks its own with /modo.

Mode	What it's for
general (default)	Takes notes or carries out requests, depending on what you say: lists, reminders, summaries, questions. It tells apart an order ("summarize X") from something you just want written down.
negocios	Notes from business visits and sales meetings: how the business operates, pain points, opportunities, objections and next steps. Keeps what the client said, what you observed and what you think apart.
ideas	Organizes an idea, plan or project: what's already defined, what's missing, contradictions and questions to decide.

All modes share the same fidelity rules: they don't invent data, they flag likely transcription errors with (¿?), and they turn relative dates ("on Thursday") into concrete ones.

Commands

Command	What it does
(voice note)	Processes it with the active mode and sends back the .md
/modos	Lists the available modes
/modo <name>	Switches this chat's active mode (for example, /modo negocios)

---

Requirements

• Self-hosted n8n. Tested on n8n 2.29. Needs the Data tables feature.
• A public HTTPS URL for your n8n instance. Telegram notifies n8n of every message through a webhook, so localhost won't work. If you run n8n at home, a tunnel (Cloudflare Tunnel, ngrok, etc.) solves this.
• A Telegram bot (free, created with @BotFather).
• A Groq API key. The free tier is enough for personal use.

Setup

1. Create the Telegram bot

Message @BotFather, send /newbot and follow the steps. It gives you a token: keep it.

2. Get your Groq API key

Create one at console.groq.com. It starts with gsk_.

3. Create the data tables

In n8n, go to Data tables and create these two tables with these exact names and columns (all of type String):

modes: name, description, prompt

chat_settings: chat_id, mode

Then import modes.csv into the modes table. It contains the three base modes. chat_settings stays empty and fills itself in whenever someone uses /modo (you can also create it by importing chat_settings.csv).

4. Import the workflow

In n8n: Create workflow → ⋯ → Import from file and pick workflow.json.

5. Set up the credentials

Create these two credentials and assign them to the nodes:

Credential	Type	Values	Nodes
Telegram	Telegram API	the BotFather token	all Telegram nodes
groq api	Header Auth	Name: Authorization · Value: Bearer gsk_your_key	STT Request and HTTP Request

In the Groq credential, Authorization and the word Bearer (followed by a space) go in exactly like that.

6. Re-select the data tables in the nodes

When you import a workflow, n8n loses the reference to the tables. Open each of these nodes and pick the right table from the dropdown:

Node	Table
Get all modes	modes
Get mode	modes
Get mode prompt	modes
Get chat mode	chat_settings
Upsert row(s)	chat_settings

7. Fill in the configuration

Open the Config node:

Field	What to put
allowed_ids	Your numeric Telegram ID. Several IDs go separated by commas.
default_mode	The mode used when a chat hasn't picked one. general by default.
language	Language code for the audio and the date (es, en, ...).

If allowed_ids is left empty, the bot answers no one. That's on purpose: anyone who finds your bot could burn through your API key. If you don't know your ID, send the bot a message and it will reply with your ID so you can copy it.

8. Activate the workflow

Activate it and send a voice note to your bot.

---

Language

Yapr was built in Spanish: the three base modes and the bot's messages (confirmations, errors, the mode list) are written in Spanish. The language field in the Config node only changes the language Whisper transcribes in and the format of the date passed to the model.

To use it in another language, translate the prompts in the modes table and the texts of the Telegram nodes (Mode changed, No mode, No mode fallback, Unauthorized, Build modes list, Send .md file). Each prompt tells the model to write in the language of the audio, so the rest should adapt on its own.

Create your own modes

A mode is a row in the modes table:

• name: what the user types after /modo (lowercase, no spaces).
• description: a one-line summary shown in /modos.
• prompt: the full system prompt.

To write a good prompt, look at the three base modes. What works best is telling the model who is speaking, in what situation, and what the note will be used for, not only which sections to include.

Built a mode that could help others? Open a pull request adding it to modes.csv.

Privacy

Audio and transcriptions are sent to Groq for processing. Yapr doesn't store the audio or the notes: it only stores which mode each chat picked. If that doesn't work for you, you can change base_url and model in the STT and Text IA nodes to any OpenAI-compatible provider (OpenAI, OpenRouter, local Ollama, etc.). Only Groq has been tested.

Known limitations

• Telegram lets bots download files up to 20 MB and Groq accepts up to 25 MB. That's plenty for normal voice notes, but very long recordings may fail.
• There's no error handling yet: if Groq fails or you hit the free tier limit, the bot stays silent.
• /start and unknown commands get the "that mode doesn't exist" reply.
• The bot's messages and the base prompts are in Spanish (see Language).

License

MIT. n8n is distributed under its own Sustainable Use License.