DEV Community: Fahad R

I asked my AI agent to research its own competitors. Here's what it found.

Fahad R — Wed, 13 May 2026 07:39:04 +0000

I built a self-hosted AI agent called Daemora.

Last week I gave it one task:

"Research your own competitors. Find their weaknesses.
Build me a go-to-market strategy."

I didn't touch it. Here's what came back.

What I asked

The full prompt was roughly:

"You are Daemora. Research the personal AI agent landscape in 2026. Find the top competitors, analyse their pain points using real user complaints from Reddit, HN, and Twitter. Identify target customers. Write outreach scripts. Build a complete GTM strategy."

No other input. No guidance. Just that.

What it did

It searched Reddit, Hacker News, Twitter, LinkedIn, Medium, and product review sites. Cross-referenced findings. Cited every source. Wrote a 3,000 word document with a full competitor analysis, 40 named target users with personalised pitches, and an outbound playbook.

The whole thing ran while I was doing something else.

What it found

Pain point 1 — Memory

34% of all AI tool complaints analyzed across 500 Reddit posts came down to one thing:

"It forgets everything every conversation."

Every major competitor — OpenClaw, Lindy, CrewAI, LangGraph — has no persistent memory architecture. Every session starts from zero.

Pain point 2 — Security

This is where it got interesting.

Daemora found CVE-2026-25253.

A confirmed prompt injection and token theft vulnerability in OpenClaw, verified by Jamf Threat Labs. Users on HN describing it as a "stack of vulnerabilities." Reddit posts titled "Do Not Use OpenClaw." One user reported burning $300 in tokens in 2 days.

The agent found this by itself. I didn't ask it to look for CVEs. It just followed the research trail.

Pain point 3 — Token costs

Managed SaaS agents are expensive and underdeliver.One Artisan AI user reported a 1% meeting booking success rate. 11x.ai was exposed on HN for claiming customers it didn't have. Lindy reviews described pricing that would "drain your wallet faster than a supermodel on a shopping spree in Paris."

Pain point 4 — Setup complexity

OpenClaw has half a million lines of code, 53 config files, and 70+ dependencies. One developer documented spending four weekends trying to get it running on a Mac Mini.

What it recommended

The agent identified 5 customer personas:

The OpenClaw refugee burned by the CVE
The indie hacker who wants ops handled while they build
The privacy-conscious dev running Ollama locally
The agency owner tired of $500/month SaaS bills
The business needing research-as-a-service

It wrote personalised DM scripts for each one.
It identified 40 specific people to reach out to
by name, with their Twitter handles and exact pitch angles.
It cited every source with working URLs.

Why this matters

This isn't a demo I staged.

I gave it a task. It used webSearch, webFetch, browserAction, writeFile, and createDocument autonomously. It decided what to search, which sources to trust, how to structure the output, and what conclusions to draw.

The CVE discovery wasn't planned. It found that by following the research trail on its own.

That's the difference between a chat tool and an agent. ChatGPT would have answered the question.
Daemora went and found the answer.

The irony

The most compelling demo of what Daemora can do turned out to be Daemora researching itself.

An AI agent that finds security vulnerabilities in its own competitors, identifies 40 target customers, and writes the outreach scripts —
while you're doing something else.

That's what we built.

Try it yourself

I gave my AI agent 3 tasks. Here's exactly what happened.

Fahad R — Sun, 03 May 2026 18:25:26 +0000

I've been building a self-hosted AI agent for 2 months.

Last week I stopped adding features and just... used it.

Here are 3 things I asked it to do, exactly what I typed,
and exactly what happened.

Task 1 — Research catnip and send me the file on Telegram

That's the whole message. Nothing else.

Here's what happened:

It searched the web across multiple sources
Cross-referenced and synthesised the findings
Wrote a structured research report
Saved it as a file on my machine
Sent the file directly to me on Telegram

I opened my phone and the file was just there. A proper
research report. Not a chat message, an actual document
I could open, save, and share.

The whole thing took about 90 seconds.

What got me wasn't the speed. It was that I didn't have
to tell it how. I just said what I wanted and it figured
out the steps.

Task 2 — Find the top 10 skipping ropes on Amazon,

analyse them, create a Google Doc and email it

This one I was skeptical about.

Four separate things chained together. Web research,
document creation, Google Docs integration, and email
delivery. I expected it to fail somewhere in the middle.

It didn't.

Searched and scraped Amazon product listings
Analysed and ranked the top 10 by reviews, price, and features
Created a formatted Google Doc with the full analysis
Emailed it to the address I gave

I walked away while it ran. Came back. The Google Doc
existed. The email had been sent.

That was the moment I went "wait, it actually did that?"

Even after 2 months of building it, it still catches
me off guard.

Task 3 — Fix the bug in this repo, open a PR,

run the tests and health check every minute

This one is for the developers.

I pointed it at a GitHub repo with a known bug. One message.

Here's what it did:

Cloned and read the codebase
Identified the bug
Wrote the fix
Ran the test suite locally, all passing
Opened a pull request on GitHub
Set up a cron job to health check the service every 60 seconds
Sent me a Telegram ping with the health status each time it ran

The PR was real. The tests passed. The health checks
kept coming.

I made coffee during this one.

What's actually happening under the hood

All three tasks ran through the same agent loop.

Task 1 used webSearch, writeFile, and replyWithFile.

Task 2 used browserAction, createDocument, useMCP
for Google, and sendEmail.

Task 3 used gitTool, executeCommand, useMCP for GitHub,
cron, and messageChannel.

The agent decides which tools to use, in what order,
and handles errors along the way. You just describe
what you want.

The part that matters

None of this runs on anyone else's server.

It runs on my machine. My API keys stay in an
AES-256-GCM encrypted vault. My files stay local.
Nothing leaves my infrastructure except the tokens
I send to the model APIs I choose.

I can use OpenAI, Anthropic, Google, Groq, DeepSeek,
Mistral, or Ollama locally with no API key needed at all.

Try it yourself

npm install -g daemora
daemora setup
daemora start

Then message your Telegram or Discord bot. Tell it
what you want. See what happens.

GitHub: https://github.com/CodeAndCanvasLabs/Daemora
Website: https://daemora.com
npm: https://npmjs.com/package/daemora

What would you ask it to do first?

I built a self-hosted AI agent that markets itself. Here's how.

Fahad R — Fri, 01 May 2026 20:39:30 +0000

Two months ago my co-founder and I got tired of paying for AI tools
that live on someone else's server, phone home constantly, and charge
per seat.

So we built our own.

Today we're open sourcing it. It's called Daemora.

What it actually does

You install it with one command:

npm install -g daemora
daemora setup
daemora start

Then you message your Telegram or Discord bot like a teammate.

This week I asked it to:

Find the top 10 skipping ropes on Amazon, analyse them, create a Google Doc and email it to someone
Research catnip, write a report, save it to my machine and send it to me on Telegram
Fix a bug in a GitHub repo, open a PR, run the tests and health check the service every minute

All three worked. While I made coffee.

The part that surprised us most

We thought the hard part would be the LLM integration.

It wasn't.

The hard part was security. When you give an AI agent access
to your actual machine, you have to think about things like:

Preventing it from dumping your environment variables
Stopping it from reading files outside its sandbox
Stripping secrets from subprocess environments so they can't be leaked via executeCommand
Blocking network egress if a response contains a known secret value

We ended up with a 14-layer security model. AES-256-GCM
encrypted vault, filesystem sandbox, subprocess isolation,
egress guard, prompt injection tagging. You can run
daemora doctor for a full scored security audit.

The thing we're most proud of

Three-layer memory.

Daemora has semantic, episodic, and procedural memory with
automatic extraction, composite-scored recall, and confidence
decay. It learns from every task without you manually saving
anything. Unified session across all channels - you can start
a task on Telegram and follow up on Discord.

It also has smart loop detection. If the agent starts burning
tokens in repetitive patterns - exact repeat, ping-pong,
semantic repeat, or polling loops - it detects and stops
automatically. This one saved us a lot of money during
development.

The meta part

This week I gave Daemora a prompt to market itself.

Set up watchers for competitor mentions. Draft Reddit posts.
Monitor GitHub for star spikes. Send me a Telegram report
every Friday with growth metrics. Generate short-form videos
via the Media Studio crew and post them automatically.

It's now running its own growth campaign. Autonomously.
On my laptop.

That felt like the moment we knew it was real.

What's inside

52 built-in tools - files, web, shell, email, scheduling, vision, media, multi-agent delegation
20 messaging channels - Telegram, Slack, Discord, WhatsApp, email, Signal, Teams and more
7 AI providers including Ollama - works fully local, no API key needed
Native MCP integration - GitHub, Notion, Linear, Postgres in one command
Media Studio crew - generate AI images, videos, and music. Edit videos with Remotion - add captions, transitions, effects programmatically
Provider failover - automatic retry with exponential backoff. If a provider goes down, switches automatically
Three-layer memory - semantic, episodic, procedural. Learns from every task, no manual saving needed
60 built-in skills - coding, research, email, GitHub, Slack, Google Workspace, macOS automation and more

The stack if you're curious

Runtime: Node.js 20+ - ES modules, no build step
AI SDK: Vercel AI SDK - model-agnostic, 25+ providers
Storage: SQLite + file-based (Markdown, JSONL)
Scheduling: croner - production-grade cron
Vault: Node.js crypto built-in - AES-256-GCM + scrypt
Voice: Twilio + OpenAI Realtime STT + ElevenLabs TTS
Video: Remotion - React-based programmatic video editing

No Docker required. No binary dependencies for the vault.
Runs as a system daemon on macOS, Linux, or Windows WSL.

Try it

npm install -g daemora

AGPL-3.0. Everything runs on your machine. Nothing leaves
your infrastructure except the tokens you send to model APIs.

⭐ GitHub: https://github.com/CodeAndCanvasLabs/Daemora

🌐 Website: https://daemora.com

📦 npm: https://npmjs.com/package/daemora

Would love to hear what you'd automate first — drop it
in the comments.