Starting point
Agent harnessesUsed on our rigLocal + hybridCLIDesktop
CLI agent on a local /v1. Hermes is used on our rig (wtpai1), pointed at Strata :8083 and llama-server. Aider and OpenCode are the other well-documented terminal attaches.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open Hermes AgentConnect to :8080 / :8083Permalink
Also consider
- Aider — CLI coding in the repo you already have. Prefix the model with openai/.
- OpenCode — CLI agent that you point at llama-server via opencode.json.
- Goose — Desktop/CLI agent with recipes. Local path is OpenAI-compatible or Ollama.
Starting point
Agent harnessesUsed on our rigLocal + hybridCLIDesktop
Hermes Agent is William’s daily setup on wtpai1 — desktop/CLI on top of Strata :8083 and llama-server. Goose is the other local-first desktop/CLI agent with an OpenAI host override.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open Hermes AgentConnect to :8080 / :8083Permalink
Also consider
- Goose — Desktop/CLI agent with recipes. Local path is OpenAI-compatible or Ollama.
Starting point
Agent harnessesLocal + hybridIDE
In-editor chat/edit/apply. Continue documents both an openai apiBase (…/v1) and a native llama.cpp provider. Cline and Roo are the other OpenAI-compatible IDE agents. Cursor’s Override OpenAI Base URL is BYOK and not a first-class localhost attach.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open ContinueConnect to :8080 / :8083Permalink
Also consider
- Cline — In-editor agent that asks before it runs commands.
- Roo Code — In-editor agent with filesystem and terminal tools.
Starting point
Agent harnessesLocal + hybridIDE
In-editor chat/edit/apply. Continue documents both an openai apiBase (…/v1) and a native llama.cpp provider. Cline and Roo are the other OpenAI-compatible IDE agents. Cursor’s Override OpenAI Base URL is BYOK and not a first-class localhost attach.
Hybrid (the same toggle as /calculator?usage=hybrid): keep private or always-on turns on the Radeon; send token-heavy reasoning to a cloud model the harness already supports. This page does not quote cloud prices.
Open ContinueConnect to :8080 / :8083Permalink
Also consider
- Cline — In-editor agent that asks before it runs commands.
- Roo Code — In-editor agent with filesystem and terminal tools.
- Cursor — Chat/Agent against a custom OpenAI-compatible URL where the override is supported.
Starting point
Agent harnessesLocal + hybridWebCLI
A browser software agent (OpenHands) or a multi-channel gateway (OpenClaw). Both attach to LLM_BASE_URL / a /v1 you already curled.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open OpenHandsConnect to :8080 / :8083Permalink
Also consider
- OpenClaw — Message the home agent from apps you already use. Already documented on this site.
Starting point
Chat UIsLocal + hybridDesktop
Desktop chat apps that also expose a local /v1. LM Studio’s documented server port is 1234; Jan’s is 1337. Pointing either *at* our llama-server is a UI check — Jan’s API-server page does not document that client path.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open LM StudioConnect to :8080 / :8083Permalink
Also consider
- Jan — Local chat and a loopback /v1 for other tools.
- AnythingLLM — Docs + chat with a local /v1. RAG-shaped, not a coding agent.
Starting point
Chat UIsLocal + hybridWeb
Browser chat against any OpenAI-compatible server. Open WebUI is the usual self-hosted pick. The llama.cpp built-in web UI is the zero-extra-process option on :8080.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open Open WebUIConnect to :8080 / :8083Permalink
Starting point
Image / videoLocal onlyWeb
Image and video graphs on ROCm PyTorch. ComfyUI does not speak llama-server /v1 — it is a different stack from the daily-driver GGUF.
Fully local (the same toggle as /calculator?usage=local): the box on the desk does the whole job. Prefer a harness that talks to 127.0.0.1 without a vendor proxy.
Open ComfyUI on ROCmConnect to :8080 / :8083Permalink