Production AI agents on Hermes & OpenClaw.
Self-hosted. Model-agnostic. Yours.

Who it's for

Built for people who want to own their AI.

• Founders & indie hackers

  You want a personal AI that lives in your Telegram, knows your projects, follows up on your tasks, runs nightly reports while you sleep — and isn't tied to one vendor's whims.

• Agencies & operators

  You want agents that actually do the work: triage inbox, draft client updates, run research loops, post to channels, escalate exceptions. Hooked into your tools, not floating in a third-party UI.

• Businesses building intelligence layers

  You want a domain-specific brain — a sales copilot, a finance reviewer, a support tier-zero agent — that learns from your data, lives behind your firewall, and gets sharper every week.

What I build

Three things, done end-to-end.

• Personal AI assistant

  A senior chief-of-staff that lives in Telegram or Slack. Reads your inbox, manages your calendar, runs scheduled briefings, remembers everyone you've talked to, and gets better the more you use it.

• Business intelligence agent

  An agent wired into your data — Postgres, Notion, Stripe, GA4, your CRM — that answers questions in plain English, writes weekly reports, flags anomalies, and acts on them when you say go.

• Ops & workflow agent

  The unsexy, high-leverage one. Lead triage. Customer support tier-zero. Internal RAG. Scheduled reports. Alert routing. Code review companions. Anything repetitive you do today.

Custom builds, advisory, or build + run as a managed retainer. We figure out the right shape on the call.

The stack — Hermes & OpenClaw

The two open-source frameworks I build on, and why it matters.

Proof

I build and run my own 14-agent stack on Hermes.

A working autonomous agent operating system on a private VPS — continuous uptime, real tasks shipped end-to-end, audit trail for every decision. The same architecture I bring to client builds.

• 14 specialist agents in production

  Commander, Concierge, Research, Dreamer, Coder, QA, OSINT, Content-Studio, SEO-Recon, and others. Each with a sharp identity and a single domain. No overlapping ownership.

• Identity-first design

  Every agent loads a SOUL.md (who they are, hard limits, refusal triggers) plus an AGENTS.md (mission, data paths, decision process, cross-agent matrix). Vague agents do vague work.

• Coordination via SQLite kanban

  Single source of truth for every task — atomic claims, dependencies, retries, full audit trail. No message queues, no event buses, no service discovery to deploy.

• Reliability as architecture, not a wishlist

  Credential pool auto-rotates exhausted API keys; a fallback chain swaps providers when one's down. The agent never sees the failure. Free-tier rate limits stop being a problem.

• Discord-first ops with cron

  I talk to one agent (the Commander) in a private channel. It delegates via kanban to the right specialist. Cron auto-fires recurring work — daily briefs, weekly retention sweeps. Output lands in a structured vault.

• Private-by-default networking

  Tailscale mesh only. No public ports. SSH and dashboards reach the VPS from my laptop and phone — and nowhere else.

• Two-layer memory

  Automatic recent-context (holographic, runtime-native) plus deliberate cross-session knowledge (vector store, semantic recall). Different time horizons, different memory shapes.

How we work

How a project runs.

1. 01

   Discovery call (20 min, free)

   You describe the problem. I tell you whether agents are the right answer, and what the smallest version looks like.

2. 02

   Scoping doc (within 48 hours)

   A one-pager: the agent's job, where it lives, what it touches, success metrics, and a fixed-scope plan. No hourly games.

3. 03

   Build sprint

   I design the agent, wire the tools, write the skills, set up the gateway and infra on your stack. You see progress in your Telegram or Slack within the first week.

4. 04

   Production hardening

   Approval flows, command allowlists, container isolation, logging, and a runbook. We don't ship toys.

5. 05

   Run, learn, expand

   Optional managed retainer where I keep the agent improving, add skills as new edges show up, and review what it's learning each month.

Why me

Start a project

Tell me what you want to build.

The more concrete, the better. One business day reply.

Things people ask

Things people ask before we start.

Will my data leave my infrastructure?

Only if you decide it should. The default deployment is self-hosted on your servers — a $5 VPS works, a GPU cluster works, your existing AWS / Hetzner / Render works. The agent only talks to the model provider you choose. If you want fully local — Llama, Mistral, your own endpoint — that's supported on day one.

Which model should I use?

Whatever fits your task and budget. Hermes is model-agnostic — you can switch between OpenAI, Anthropic, OpenRouter, Nous Portal, or self-hosted weights with one config change. We pick on the discovery call.

Where does the agent live?

Telegram, Slack, Discord, WhatsApp, Signal, a CLI, a web dashboard, an internal tool — or several of those at once. Hermes was designed for messaging-first; OpenClaw is too. Email, voice memos, and scheduled tasks all work out of the box.

How is this different from buying ChatGPT Enterprise / Claude Teams / a Zapier agent?

Those are great until you want memory that persists, tools that aren't on the vendor's allowlist, models other than the one they sell you, or your data not crossing their wire. The moment any of that shows up, you need an agent you actually own. That's what I build.

Do you do build-only, or build + run?

Both. Default is a fixed-scope build with a runbook so your team can operate it. If you'd rather I keep improving it monthly — adding skills, watching telemetry, hardening edges — there's a managed retainer.

What about security and approvals?

Every command the agent runs can be allowlisted, sandboxed, or require human approval. Hermes ships with command approval, DM pairing, and container isolation; I extend that with whatever your security review needs. Nothing ships without an audit trail.

Timeline?

Most first agents reach a usable v1 in 1–3 weeks. Production hardening adds 1–2 weeks. Anything bigger we phase.

Pricing?

Quoted per project after the discovery call. Fixed scope, fixed price, no hourly games. Retainers are monthly with a clear deliverable list.