llm-cli

owui — CLI agent: Open WebUI + Open Terminal via MCP

A terminal-native agent that gets its brain from Open WebUI (LLM) and its hands from Open Terminal (sandboxed shell), connected via the Model Context Protocol.

┌─────────┐     tools (from MCP)       ┌───────────────┐
│         │ ──── chat/completions ────→ │  Open WebUI    │
│  owui   │ ←─── stream + tool_calls ── │  (LLM)         │
│  (CLI)  │                             └───────────────┘
│         │     MCP (streamable-http)   ┌───────────────┐
│         │ ──── CallTool ────────────→ │ Open Terminal  │
│         │ ←─── results ────────────── │ (MCP server)   │
└─────────┘                             └───────────────┘

Tools are discovered dynamically. At startup, owui connects to Open Terminal's MCP server and calls ListTools — whatever endpoints your instance exposes (execute, files, search, terminals, notebooks…) become available to the model automatically. No hardcoded tool definitions, no version drift.

Built with Kukicha and its standard library:

• stdlib/llm — OpenAI-compatible chat completions (streaming, tool calls)
• stdlib/mcp — MCP client (connect, list tools, call tools, bearer auth)
• stdlib/cli — fatal error handling
• stdlib/fetch — HTTP requests
• stdlib/input — readline prompts
• stdlib/json — JSON marshalling

Install

kukicha build examples/llm-cli/
# binary at ./owui

Configure

# Environment
export OWUI_WEBUI_URL="http://localhost:3000"
export OWUI_WEBUI_API_KEY="sk-..."
export OWUI_MODEL="llama3.1"
export OWUI_TERMINAL_MCP_URL="http://127.0.0.1:9000/mcp"
export OWUI_TERMINAL_MCP_API_KEY="..."  # optional, for authenticated MCP servers

# Or interactive
owui configure

# Verify
owui health
owui tools    # see what the MCP server exposes
owui models   # see what LLMs are available

Usage

Agent mode (LLM + MCP tools)

# One-shot
owui "set up a python project with fastapi and write a hello world"

# Pipe context
cat error.log | owui "diagnose and fix this"

# Raw output for piping
owui "list installed python packages" --raw | grep torch

# Interactive chat
owui -c

# Seed the system prompt
owui -c -S "you are a senior Go developer. be terse."

# Pipe context then chat
cat main.go | owui -c "review this code"

Tool calls show on stderr:

MCP connected: 12 tools from http://127.0.0.1:9000/mcp
-> execute_command {"command":"ls -la /workspace"}
<- execute_command total 24 drwxr-xr-x 3 user user 4096...
-> write_file {"path":"/workspace/main.py","content":"from fa...
<- write_file Wrote 342 bytes to /workspace/main.py
The FastAPI project is set up...
(3 tool rounds)

Inspect

owui tools     # list all tools discovered from MCP
owui models    # list available LLMs
owui health    # check connectivity to both services

Flags

Architecture

main.kuki        CLI entrypoint, flag parsing, subcommands, interactive chat
config.kuki      Config loading (env > file > defaults)
agent.kuki       Agent loop: LLM → tool calls → MCP dispatch → repeat
bridge.kuki      MCP client → connects (with optional bearer auth),
                 discovers tools, converts to stdlib/llm format,
                 dispatches CallTool via MCP protocol

The key file is bridge.kuki:

1. Connect — mcp.Connect (or mcp.BearerConnect for authenticated servers) to Open Terminal via streamable HTTP
2. Discover — mcp.ListTools gets all available tools with schemas
3. Convert — MCP tool schemas → llm.Tool format for the LLM
4. Dispatch — mcp.CallTool executes tools and returns results

No hardcoded tool definitions anywhere. If Open Terminal adds a new endpoint, your CLI picks it up on the next run.

Prerequisites

• Open WebUI running (for the LLM)
• Open Terminal running with MCP enabled:

  pip install open-terminal[mcp]
  open-terminal mcp --transport streamable-http
  

  Or the standard HTTP server works too — FastMCP.from_fastapi exposes /mcp alongside the REST API.

License

MIT