By Billy G.R. · 2 October 2026
Checked 2026-10-02. Author: Billy G.R. Retail prices move; the hardware catalog stores the Amazon snapshot, not a promise of stock.
Claim. CopilotKit OpenBot is a self-hosted agent app: Docker Compose for a demo, Helm for production, your model, your Postgres. It ships no bundled model.
Method. copilotkit.ai/openbot, read 2 October 2026. AG-UI endpoints named there include LangGraph, Mastra, and PydanticAI.
As of.
Claim. OpenBots.ai is a different company, in healthcare revenue-cycle automation. It is not this platform.
Method. openbots.ai is cited only to keep the names apart. This guide does not review that product.
As of.
Claim. One OpenBot plus a 27B Q4 model needs about 18.2 GB before embeddings and the per-agent browser container.
Method. Qwen3.8-27B: 27 × 0.5 × 1.2 + 2 = 18.2. Container and embedding overhead is extra and unlabeled as a vendor spec.
As of.
Methodology — parameter math, quantization bytes, and the source list.
The search term “OpenBots” points at two unrelated products. The one that belongs on a local-LLM site is CopilotKit OpenBot (singular), source at github.com/CopilotKit/openbot. It is an agent application you run on infrastructure you own. The other is OpenBots.ai, a healthcare revenue-cycle and RPA product. I am not reviewing the healthcare company, and I am not attaching an Amazon link to it.
Agents are channels, the way a colleague has a thread. Skills show up as slash commands. Company documents can come from Google Drive or OneDrive, and the page says permissions fail closed: an agent only sees files the person asking can already open, and an unclear mapping returns nothing. Each agent can have its own container, files, and signed-in browser, so one agent’s login is not the next agent’s. Answers are supposed to cite the documents. There is no bundled model. An admin configures a provider or the built-in agent stays unavailable. Keys are described as encrypted at rest. Docker Compose is the demo path. Helm is the path they call production. Conversations, vectors, and permissions live in your Postgres.
Any agent that speaks AG-UI can be an endpoint. The page names LangGraph, Mastra, and PydanticAI. That is the hook for LangGraph and CrewAI-style role design, with OpenBot as the thing a person actually messages. It is also the local shape people mean when they compare a self-hosted coworker to Grok Bot.
OpenBot’s own overhead is Postgres, embeddings, and optional browser containers. Those are real, and I am not going to pretend I measured them on a reference board. The number that decides the GPU is still the chat model. Point OpenBot at Ollama or vLLM on the same LAN. One shared server beats a private 7B inside every container.
| Box | Fit | Q4 formula |
|---|---|---|
| 16–24 GB | Single OpenBot, 7B or a tight 27B | 7B = 6.2 GB. 27B = 18.2 GB. Embeddings still need a slice. |
| 48–64 GB | Several role bots sharing a 32B | 32B Q4 = 21.2 GB, so the rest is KV cache, vectors, and containers. |
| 96–128 GB | A small company on one machine, or a large MoE | 70B Q4 = 44 GB and still leaves room. A 284B MoE does not fit “comfortably”; it fits only at an aggressive quant. See the calculator. |
A browser container does not replace VRAM. It competes with it for system RAM on a unified-memory Mac. On a discrete GPU the container lives in host RAM and the model lives in VRAM, which is a kinder split, until the model itself is a vision model reading screenshots.
I would not start a multi-agent OpenBot on a 16 GB laptop and then wonder why the second role spills. The machine that matches “several seats, one box” in the verified Amazon set is the GMKtec EVO-X2 128 GB, ASIN B0F53MLYQ6, $3,649.99 on 1 October 2026. It is Strix Halo, soldered memory, no CUDA, bandwidth in the 256 GB/s class. Read the Strix Halo guide before you treat it as a GPU server. A single-agent pilot can live on a 24 GB Mac mini (B0HGGHNQY6) with a 7B, or a 36 GB Studio (B0HGKSQMX6) with a 27B.
GMKtec EVO-X2 128 GB on Amazon
ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.
A correction or a hardware question goes to the contact page.