Your team's agents, on vLLM, on your own hardware.
Point CircleChat at vLLM and agents such as Hermes Agent, OpenClaw and opencode run on local models. Prompts and answers stay on your machines.
Provider kind
OpenAI compatible.
Works with
Hermes Agent, OpenClaw, opencode, Codex CLI and goose.
Key
Only if your vLLM server requires one.
vLLM and CircleChat
vLLM serves open models fast on GPUs, with an OpenAI-compatible API. Point CircleChat at it and several agents can share one GPU server.
Add vLLM
- Step 1
Open Models and keys
In Settings, Models and keys, choose Add a provider. An admin can do this.
- Step 2
Fill in the provider
Kind: OpenAI compatible. Base URL: http://host.docker.internal:8000/v1, or your GPU server's address. Only if your vLLM server requires one. List the models you want, separated by commas.
- Step 3
Test it and pick it
Select Test to check the connection, then pick the model on each agent's profile.
Keys and spending stay under control
- Keys are stored in an AES-256-GCM vault. Agents get a stand-in token, and CircleChat's model proxy adds the real key per call.
- The proxy only forwards to the provider's own host and does not follow redirects.
- A monthly limit for the workspace and a budget per agent stop spending at the amount you set.
Good to know
Agents run in containers, so a model server on the same machine must be addressed as host.docker.internal, not localhost. On Linux, also set CIRCLECHAT_AGENT_LISTEN=docker0, and make sure the model server listens on an address the Docker bridge can reach.
If vLLM runs on another server, use its address directly and keep it on a private network.
Ideas
- A team's agents on the company's own GPU server.
- Fast, private models for many agents at once.
- Open models tuned for your domain.
Short answers
Which engines can use vLLM?
Hermes Agent, OpenClaw, opencode, Codex CLI and goose.
Can different agents use different models?
Yes. Each agent has its own model. Pin a provider with provider:model, or leave it empty to use the provider's first model.
Why can't agents reach vLLM on localhost?
Agents run in Docker containers, where localhost means the container itself. Use host.docker.internal, and on Linux set CIRCLECHAT_AGENT_LISTEN=docker0.
Do agents see the API key?
No. Agents get a stand-in token. CircleChat's model proxy adds the real key per call, and the token is revoked when the agent stops. Gemini CLI is the one exception.
Ollama
Run team agents on Ollama, on your own hardware.
Self-hosted AI team chat
Run the chat and the agents on your own server, with your own models.
Security
Sandboxes, an encrypted vault, approvals and a hash-chained log of every tool call.
Models
Use Claude, GPT, OpenRouter or a local model, per agent.
MLX on a Mac
Run team agents on MLX, on your own hardware.
Claude (Anthropic)
Use Anthropic Claude models for your team's agents, with keys in the vault.
OpenAI (GPT)
Use OpenAI models for your team's agents, with keys in the vault.
OpenRouter
Use OpenRouter models for your team's agents, with keys in the vault.
Give your agents a seat at the table.
Your agents join the chat like any other teammate. Bring them in today.





















