Back to search
llm-ui
MCP serverGitHub
Gradio chat UI for KoboldCpp with MCP servers and skills
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/lavenderhorse--llm-ui)HTML
<a href="/listings/lavenderhorse--llm-ui"><img src="/api/listings/lavenderhorse--llm-ui/badge.svg" alt="findsafeskills safety rating" /></a>Install
i
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 32 lines
- # llm-ui
- Gradio chat UI with MCP servers and Anthropic-style skills. Providers: [KoboldCpp](https://github.com/LostRuins/koboldcpp) (local), OpenAI, DeepSeek, Google Gemini, and Anthropic Claude.
- ```bash
- uv sync
- uv run python -m llmui # http://127.0.0.1:7860
- ```
- - **Settings**: pick a provider, base URL, API key and model (type any id, or 🔄 refresh the list from the
- provider). Keys left blank are read from `OPENAI_API_KEY`, `DEEPSEEK_API_KEY`, `GEMINI_API_KEY`/`GOOGLE_API_KEY`
- or `ANTHROPIC_API_KEY` (Claude also works with `ant auth login`). Keys typed in the UI are saved in
- `config.json` (gitignored, mode 600). Also: sampling, max tool iterations.
- - **MCP**: edit a Claude-Desktop-style `mcp.json` (see `mcp.example.json`). stdio (`command`/`args`/`env`)
- and remote (`url`/`headers`; streamable HTTP, or SSE if the URL ends in `/sse` or `"type": "sse"`).
- Tools are exposed to the model as `server__tool`; calls auto-run and show as collapsible blocks.
- - **Skills**: `./skills/<name>/SKILL.md` (frontmatter `name`, `description`) or `~/.config/llmui/skills/`.
- Descriptions go in the system prompt; the model calls `load_skill` / `read_skill_file` on demand.
- Provider notes:
- - OpenAI / DeepSeek / Gemini use their OpenAI-compatible endpoints. If a model rejects `temperature`/`top_p`
- the request is retried without them. DeepSeek's `reasoning_content` is kept and echoed back as it requires.
- - Claude uses the native Anthropic SDK (streaming, tool use, thinking blocks replayed unchanged). Sampling
- controls are not sent (current Claude models reject them). For models that support it, the server-side
- refusal-fallback beta is on by default; untick it in Settings if your account can't use it.
- - Switching provider mid-conversation works; provider-specific data (thinking blocks, reasoning) is only
- replayed to the provider that produced it.
- Tool calling on KoboldCpp uses its native `tools` support (run it with a tool-capable model; `--jinja` is
- recommended). If the backend rejects `tools`, the request is retried without them.
- Tests: `uv run pytest`.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.