Ertas + OpenClaw for AI Agencies

AI agencies deploying OpenClaw for clients face the same margin problem as every API-dependent tool: per-token costs that grow with each client. With Ertas, agencies train a model for each client, export it, and run it on their own hardware behind OpenClaw, so per-token API bills give way to a predictable hosting cost and each agent knows its client's domain.

The Challenge

OpenClaw is the most capable AI agent framework available: autonomous task execution through messaging platforms that clients already use, browser automation, file management, email triage, and cron-based monitoring. For AI agencies, it is a natural evolution from simple chatbot deployments to full-spectrum AI assistants.

But the economics do not work at scale with cloud APIs. Every client's OpenClaw instance routes every interaction through GPT-4o or Claude, generating per-token charges on every email triaged, every report generated, every message sent. A single active client can cost AU$150-300/month in API pass-through. Across 10-20 clients, the agency is looking at AU$2,000-4,000/month in variable costs that eat directly into fixed-retainer margins.

Worse, the costs are unpredictable. A client's marketing campaign goes viral and their support volume triples? The API bill triples too, but the retainer stays the same. An OpenClaw cron job scans an inbox every 30 minutes? That is steady token throughput 24/7, burning through credits even when nothing actionable happens.

The differentiation problem is equally pressing. If every agency deploys OpenClaw with the same GPT-4o backend, clients are effectively getting the same AI, just wrapped in different branding. There is no moat. A technically inclined client can replicate the setup themselves.

The Solution

Ertas moves agency OpenClaw deployments from API pass-through to models the agency owns. The pattern: train a model on each client's domain data in Ertas Studio, export it, run it locally with Ollama or another OpenAI-compatible server, and point that client's OpenClaw instance at the local endpoint.

The economics change once the hardware is in place. A request to a local model costs electricity and hardware the agency already owns, so a client whose volume triples no longer triples the bill. Training happens in Studio on Ertas's cloud GPUs through a visual canvas, so the agency needs no training infrastructure of its own, and several candidate runs can train in parallel and be compared before one goes live. Training data is uploaded to Ertas for this step, so strip personal details from the logs first.

The quality gain is the real selling point. A model trained for a real estate client has learned that client's listings, pricing terminology and buyer communication style. A model trained for a dental practice knows the practice's appointment types, insurance panels and patient tone. A general model has to approximate all of that from a system prompt on every request. Agencies can measure the difference on each client's own held-out examples, which gives them evidence to show the client and a basis for premium pricing.

Key Features

Studio

Per-Client Fine-Tuning

Upload a client's conversation logs, product catalogue or domain data as JSONL, or build the dataset in Data Craft, then configure a LoRA training run on Studio's visual canvas. Training runs on Ertas's cloud GPUs and produces a model that captures that client's needs, with no GPU infrastructure to manage.

Hub

Export and Run Locally

Hub lists every model the agency has trained. Download each as GGUF for Ollama, llama.cpp or LM Studio, or as safetensors for vLLM, and serve it on the agency's own hardware. Point each client's OpenClaw instance at its model, and the agent runs with no Ertas dependency and no per-token bill.

Hub

A Separate Model Per Client

Each client's agent runs on its own trained model, so one client's data never shapes another client's replies. The agency decides where each model runs, and every agent's requests and outputs stay on that hardware. Training datasets can be deleted from Studio when an engagement ends.

Studio

Base Models for Agent Work

Choose from Studio's catalogue of open models, including Qwen, Llama, Gemma, Mistral and Phi, at sizes from a few hundred million to 8 billion parameters. Train two or three candidates in parallel on the same client data and compare them on email triage, scheduling or reporting tasks before picking the one to deploy.

Example Workflow

A Sydney-based AI automation agency manages OpenClaw deployments for 12 small business clients across real estate, hospitality and professional services. Each client has an OpenClaw agent handling email triage, appointment scheduling and customer inquiries through WhatsApp and email. On cloud APIs the agency spends AU$2,800 a month, and three high-volume clients (a real estate agency, a boutique hotel and an accounting firm) account for AU$1,500 of it. The agency moves to Ertas one client at a time. For each client the team exports three to six months of conversation history, removes personal details, formats it as JSONL and trains a model in Ertas Studio, running two base model sizes in parallel and keeping the one that does best on a held-out set of that client's real inquiries. Each chosen model is exported from Hub as GGUF and served with Ollama on a Mac Studio in the agency's office, and each client's OpenClaw instance points at its own model. Before switching a client over, the agency compares the new model's inquiry classification and replies with what GPT-4o produced for the same messages, and has the client review a sample of replies for brand voice. Once moved, those clients cost the agency an Ertas subscription plus electricity and hardware it owns, and the agency can sell a model trained on each client's own data as a premium tier.

Compliance & Security

When each client's model runs on the agency's hardware, the agent's requests and outputs stay there and are never sent to a third-party AI provider, which gives agencies a clear answer in RFP responses and enterprise procurement questionnaires. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas: remove or mask personal details in conversation logs before upload, and delete datasets from Studio when an engagement ends. Training a separate model for each client keeps one client's data out of another client's model. Securing the agency's own servers, and meeting any contractual or regulatory obligations to its clients, remain the agency's responsibility.