The open-source models are already at parity. The bottleneck is the orchestration tax.

Because the bottleneck for a consulting practice in 2026 is not model quality, it is orchestration. Llama 3.1 8B running locally on Ollama can summarize a Zoom transcript, draft a follow-up email in your voice, and extract action items at quality high enough for a paying client. What the open-source stack does not include is the operator that takes the summary, opens your HubSpot tab, clicks the right contact, pastes the action items, and queues a Gmail draft for your review. That is a 'computer agent' problem, and there is no first-class open-source consulting-grade computer agent today. Clone fills that gap. The honest framing is: keep the open-source half of the stack, rent the orchestration half. Principle 3 of Clone's architecture (architecture.tsx lines 56 to 58) commits to driving whatever apps you already use, including the open-source ones.

Sorted by hours-to-first-useful-ritual on a consultant's Mac: Ollama (the LLM runner), Llama 3.1 8B or Mistral 7B (the model), Whisper.cpp (transcription), Obsidian or Logseq (markdown knowledge base), Open WebUI (optional chat surface), n8n or Activepieces (workflow runner if you choose to self-host one), EspoCRM or SuiteCRM (CRM if you choose to self-host one), and Invoice Ninja or Akaunting (invoicing if you choose to self-host one). All have MIT, Apache, AGPL, or Elastic licenses. All are real and shipped. None of them are a consulting operator on their own.

An honest count for a careful consultant: 12 to 25 hours of focused work, spread across one to two weekends. Two evenings to land Ollama plus a model. Two evenings to set up Whisper.cpp watching a Zoom directory. One weekend to stand up n8n with the post-call workflow. A few hours per CRM and invoicing tool you self-host. Plus weekly maintenance for Docker updates, model upgrades, n8n breakage, and credential rotation. The orchestration tax is real, and it is not paid in dollars but in evenings.

Clone's Planner accepts an OpenAI-compatible base URL. Point it at your local Ollama at http://localhost:11434/v1 and Clone routes inference through llama3.1:8b. Whisper.cpp runs on its own and writes .vtt files to a directory. Inside the ritual file, you tell Clone to read the latest .vtt and pass it to the Planner. Clone's Computer Agent then drives your real browser tabs (HubSpot, Gmail, your invoicing tool). Memory writes to ~/.clone/memory/, which is plain markdown a consultant can drop into Obsidian or Logseq. The architecture (six layers, four principles) lives at architecture.tsx lines 5 to 65 on the shipped marketing site.

No. Clone is a proprietary product, $49 per month on Solo. The page is honest about that. The argument is not that everything in a consultant's stack must be open-source; the argument is that the model layer, the transcription layer, and the knowledge layer can comfortably be open-source today, and the orchestration layer is where commercial software is still the practical answer for a solo consultant who would rather bill than write Python. If you self-host the rest of the stack and rent only the orchestration, your subprocessor list shrinks to one line.

Principle 1 ('Runs on your machine,' lines 46 to 48) is the desktop-native commitment; it explains why Clone runs on a Mac and not in a cloud. Principle 3 ('Tool agnostic by design,' lines 56 to 58) is the open-source-pairing commitment; it explains why Clone does not care whether the LLM behind it is Anthropic, OpenAI, or your local Llama 3.1, and whether the CRM tab it drives is HubSpot or self-hosted EspoCRM. Together they are the architectural reason the open-source half of the stack stays under the consultant's control after Clone is added.

Three. First, on a Mac with less than 16 GB of RAM, the local Llama 3.1 8B will be slow; you will end up routing to a cloud model anyway. Second, an open-source self-hosted CRM (EspoCRM, SuiteCRM) is a real operations job: backups, TLS, version upgrades. If your billable rate is $300 per hour, you may find a SaaS CRM is cheaper end to end. Third, Clone is the proprietary line in the budget; if your goal is zero proprietary lines, Clone is not the answer. Browser-Use and Anthropic's Computer Use samples are the closest open-source operators, but neither is consulting-ready in April 2026.

Direct enumeration of what an honest open-source consulting stack requires on disk: an Ollama Modelfile (or at minimum a model name configuration), a docker-compose.yml for Open WebUI or n8n, a Whisper.cpp config or shell wrapper, an n8n workflow JSON for the post-call automation, a LiteLLM proxy YAML if you put a router in front of Ollama, an Obsidian or Logseq vault path, a launchd plist or cron schedule that watches the Zoom directory, and at least one OAuth credentials JSON per integrated service. Clone replaces all eight with one markdown file at ~/.clone/memory/rituals/<name>.md, containing one trigger line and one English paragraph.

Yes, and many consultants do. The page is not arguing against that path; it is arguing for clear-eyed accounting. If you enjoy the engineering, the open-source-only path is real and free of subscription fees. The trade is your evenings for $49 a month. If you would rather spend those evenings on billable work or on rest, the proprietary orchestration line is the one most consultants choose. Either way, the inference, transcription, and knowledge layers stay open-source.

There is not literally one; that is the orchestration tax. There is, however, one sentence that runs once Clone is connected to your local Ollama: 'after every Zoom call, summarize the local transcript through my Ollama, draft a 5-line follow-up in my Gmail, log action items to my HubSpot contact, and write the session to ~/.clone/memory/sessions/.' Clone saves that ritual to a markdown file. Next call you take, the ritual fires when Zoom quits.