PostPilot

About the project

A squad of AI agents for brand promotion on Telegram: find the right groups, find people who need the product, talk to them and follow up. It runs on NousResearch's open-source Hermes Agent framework, with a dashboard I built.

PostPilot is a team of AI agents that promotes a brand on Telegram. You tell it about your brand. It finds the communities where your customers are, finds the people in them who need what you sell, talks to them and follows up, while you watch everything from a dashboard.

It runs on NousResearch's open-source Hermes Agent framework. I built the agents, their tools, the Telegram layer and the dashboard on top of it. Version 0.2 works end to end; we paused it before launch.

How it works

1. 1

   Set up the brand

   Paste your website. The AI fills in your brand details, then interviews you out loud to finish the setup.

2. 2

   Plan

   Every 5 minutes the Leader agent looks at where things stand, writes a plan as JSON and carries it out.

3. 3

   Find

   The Search agent finds relevant Telegram groups. An account joins and reads the members and their bios.

4. 4

   Qualify

   The Research agent scores each person from 0 to 100. A score of 60 or more makes them a prospect.

5. 5

   Talk

   The Sales agent sends a warm message that mentions the group you share, then up to 3 follow-ups two days apart, within each account's daily limit.

Under the hood

Agents

Python on Hermes Agent: Leader, Search, Research, Talk and Sales, each with its own tools.

Telegram

Real user accounts through Pyrogram, with phone, OTP and 2FA login from the dashboard.

Dashboard

Next.js 16: accounts, communities, leads, conversations, activity, learnings and costs.

Data

PostgreSQL with Drizzle, shared with the Python agents.

What I did

• Version 0.1: three agents that research topics, write branded posts and post them to a channel
• Rebuilt it as 0.2 with a Leader plus Search, Research, Talk and Sales agents, each with its own tools
• The Leader runs every 5 minutes, writes a plan as JSON and actually carries it out, starting agents and waking early when one finishes
• Moved Telegram to real user accounts instead of bots, with phone, OTP and 2FA login from the dashboard
• The funnel: find groups, read members and bios, score them (60+ is a prospect), then a warm message and up to 3 follow-ups two days apart
• Limits per account (15 messages a day by default), monitoring per community, auto-reply per user, and replying from the dashboard
• The dashboard: accounts, communities, leads, conversations, an activity timeline and what the agents learned
• Voice onboarding: paste a URL, the AI fills in the brand, then interviews you out loud
• A cost tracker per agent and model, in rupees, and crash recovery with a redeploy button
• Moved from SQLite to Postgres, with a small layer so the Python agents work with either

What was new

• A leader agent that executes its plan instead of just logging it
• Agents that act through real accounts, which means rate limits and account health matter
• Setting up a brand by talking, not filling a form

Problems I hit, and how I fixed them

1. 01

   The Research agent crashed on every tool call, so no lead was ever scored.

   Fix The framework passes extra arguments; all 8 tool handlers now accept them.

2. 02

   The Python agents didn't load their settings, so AI calls failed.

   Fix Load the environment in all four agents.

3. 03

   Replies never reached Telegram.

   Fix Await the reply properly.

4. 04

   Expired login codes locked users out.

   Fix File-based sessions, automatic resend and a Resend button.

5. 05

   A crashing agent restarted in a tight loop and burned API credits.

   Fix At least 30 seconds between cycles.

6. 06

   Brand setup took about 15 seconds.

   Fix A faster model: about 3 seconds.

Screenshots

To be clear

Hermes Agent is NousResearch's framework, not mine. I built the agents, tools, Telegram layer and dashboard on top of it.