Compare harnesses
Same catalog as the rest of this section. Tool-calling and long-context cells are paraphrases of the cited docs — “not documented” means we did not find a first-party claim. As of 2026-10-12.
Pick a harnessConnect snippets
| Harness | Category | Surfaces | Local /v1 | Hybrid | Tools (cited) | On our rig |
|---|---|---|---|---|---|---|
| Hermes Agent Nous Research | Agent harnesses | CLI, Desktop | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Native agent tools. Docs require at least 64K context or Hermes rejects the model at startup. | Used on our rig |
| OpenHands All Hands | Agent harnesses | Web, CLI | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Agent tools via the OpenHands stack. Local models must be strong at tool-following — not documented as guaranteed on every GGUF. | — |
| Aider Aider-AI | Agent harnesses | CLI | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Repo edits and git. Unknown models warn until you add .aider.model.settings.yml metadata. | — |
| Cline Cline | Agent harnesses | IDE, CLI | Yes http://127.0.0.1:8080/v1 | Local + hybrid | File and terminal tools. Quality depends on the model’s tool-calling, not on Cline inventing a protocol. | — |
| Roo Code Roo Code | Agent harnesses | IDE | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Native OpenAI tool calling is required — Roo’s docs say a model without it cannot be used. | — |
| Continue Continue | Agent harnesses | IDE | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Set capabilities: [tool_use] on the openai provider when the server actually implements tools. Otherwise not documented as working. | — |
| OpenCode OpenCode / SST | Agent harnesses | CLI | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Agent tools in the product. Use @ai-sdk/openai-compatible for /v1/chat/completions; @ai-sdk/openai if the model uses /v1/responses. | — |
| Goose Block | Agent harnesses | CLI, Desktop | Yes http://127.0.0.1:8080 | Local + hybrid | Tool calling depends on the model. Ollama path documents a toolshim; OpenAI-compatible is native function-calling if the server implements it. | — |
| Claude Code Anthropic | Agent harnesses | CLI | Yes http://127.0.0.1:8080 | Local + hybrid | Official tool use is Anthropic-shaped. Whether llama.cpp’s Messages endpoint maps every Claude Code tool is not documented here. | — |
| Codex CLI OpenAI | Agent harnesses | CLI | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Codex agent tools. Completeness against llama.cpp’s Responses translator is not documented here. | — |
| Cursor Cursor | Agent harnesses | IDE | Not first-class https://your-reachable-host/v1 | Hybrid-shaped | Agent/Ask may send Responses-shaped payloads. Tab/autocomplete stay on Cursor’s own models per staff. | — |
| OpenClaw OpenClaw Foundation | Agent harnesses | CLI, Web | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Channel tools + whatever the model supports. Follow OpenClaw allowlists — do not bind to the internet for a first test. | — |
| Open WebUI Open WebUI | Chat UIs | Web | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Optional tools / pipelines. Default is chat. | — |
| LM Studio LM Studio | Chat UIs | Desktop | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Documented OpenAI endpoints include chat completions and Responses. Tool-calling completeness on Radeon GGUF is not documented. | — |
| Jan Menlo / Jan | Chat UIs | Desktop | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Chat completions documented. Images/audio endpoints listed as not yet supported on the API-server page we opened. | — |
| SillyTavern SillyTavern | Chat UIs | Web | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Optional function calling in Chat Completion sources — not documented as our daily-driver path. | — |
| LibreChat LibreChat | Chat UIs | Web | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Product has tools/agents; whether they fire against llama-server tool-calling is not documented here. | — |
| AnythingLLM Mintplex Labs | Chat UIs | Desktop, Web | Yes http://127.0.0.1:8080/v1 | Local + hybrid | Workspace features; agent completeness against llama-server is not documented. | — |
| llama.cpp built-in web UI ggml-org | Chat UIs | Web | Yes http://127.0.0.1:8080 | Local only | Server README documents function calling / --tools. Completeness on Qwen3.6 is not measured here. | — |
| ComfyUI on ROCm comfyanonymous / AMD docs | Image / video | Web | Yes n/a | Local only | Node graph, not LLM tools. | — |
Local vs hybrid is the same question asfully localversushybridon the calculator. Cursor’s override is hybrid-shaped even when the weights are local — forum staff say requests still go through Cursor’s backend.