THU CyberCraze Agent

Interactive terminal coding agent powered by the THU lab proxy (OpenAI-compatible API). It runs in your current terminal, works in your current directory, can inspect files, propose shell commands, and wait for your approval before running them.

1. Installation

1.1 Get an API key

Create a key here:

https://lab.cs.tsinghua.edu.cn/ai-platform/c/new

Base URL:

https://lab.cs.tsinghua.edu.cn/ai-platform/api/v1

Set environment variables:

export THU_LAB_PROXY_API_KEY='your_proxy_key_here'
export THU_LAB_PROXY_BASE_URL='https://lab.cs.tsinghua.edu.cn/ai-platform/api/v1'

Windows PowerShell:

$env:THU_LAB_PROXY_API_KEY='your_proxy_key_here'
$env:THU_LAB_PROXY_BASE_URL='https://lab.cs.tsinghua.edu.cn/ai-platform/api/v1'

You can also launch the agent and paste the key when prompted. The agent saves it into a per-user global config file:

• Linux and macOS: ~/.thu-cybercraze-agent/.env
• Windows: %USERPROFILE%\.thu-cybercraze-agent\.env

1.2 Run the agent

Linux (built binary):

./dist/thu-agent

Windows (built on Windows):

.\dist\thu-agent.exe

macOS (run Python directly):

python3 agent.py

1.3 Build the binaries (if needed)

Linux build:

bash build_agent.sh

Result:

dist/thu-agent

Windows build (run on Windows, not inside WSL):

py -3 -m pip install pyinstaller
powershell -ExecutionPolicy Bypass -File .\build_agent_windows.ps1

Result:

dist\thu-agent.exe

1.4 Optional: Run globally

Linux:

sudo install -m 755 dist/thu-agent /usr/local/bin/thu-agent

Windows: add the repo dist directory to PATH, or copy the .exe into a directory already on PATH.

Example (PowerShell):

[Environment]::SetEnvironmentVariable(
  "Path",
  $env:Path + ";C:\Users\USER\Downloads\THU-deepseek-glm-api-mcp-server\dist",
  "User"
)

Open a new terminal and run:

thu-agent.exe

2. Usage

Start the agent:

./dist/thu-agent

Or run with Python:

python3 agent.py

Pass model and key directly if you want:

python3 agent.py --model deepseek-v3.2 --api-key "$THU_LAB_PROXY_API_KEY"

Default model:

deepseek-v3.2

Current models:

• qwen3-max-thinking
• qwen3-max
• glm-5
• glm-5-thinking
• glm-4.7-thinking
• kimi-k2.5
• kimi-k2.5-thinking
• minimax-m2.5
• minimax-m2.5-thinking
• qwen3.5-plus
• qwen3.5-plus-thinking
• qwen3.5-mini
• deepseek-v3.2-thinking
• deepseek-v3.2

While the agent is thinking or running a command, press Ctrl+C to interrupt. It will ask for a follow-up instruction. Type /stop there to discard the interrupted turn, or type a new instruction to continue.

3. Function List

Slash commands available:

• /help
• /save [name]
• /autosave
• /context
• /compact [keep]
• /clear
• /status
• /attach <path> [instruction]
• /stop
• /sessions
• /load <id|name>
• /fork <id|name> [new-name]
• /new [name]
• /delete <id|name>
• /update
• /model
• /key
• /pwd
• /alwaysRun
• /exit

4. Function Explanation

Session and memory:

• /save [name] saves the current session to disk. Sessions are manual-save by default.
• /autosave toggles automatic saving for this session.
• /sessions lists saved sessions with an ID, summary, and last-used time.
• /load <id|name> loads a saved session.
• /fork <id|name> [new-name] creates a new session from a saved one.
• /new [name] starts a new session with a fresh context.
• /delete <id|name> deletes a saved session.

Context management:

• /context refreshes and displays the startup context snapshot (date, git status, nearby memory files like AGENTS.md or CLAUDE.md).
• /compact [keep] summarizes older messages and keeps recent turns to reduce context size.
• /clear clears in-memory conversation while preserving current project context.
• /status shows version, model, session name, autosave state, message count, and context size.

Commands and execution:

• /alwaysRun toggles auto-approval for shell commands.
• /stop is used only after an interrupt prompt to discard the interrupted turn.

Attachments:

• /attach path/to/file.txt explain this file inlines small text/code files into the next model turn.
• Non-text files are passed as file references for the agent to inspect with commands.
• Image files can be sent as multimodal content only when the selected model/proxy supports it and THU_AGENT_MULTIMODAL=1 is set. Otherwise they are treated as file references.

Models and keys:

• /model reselects the model (this resets the conversation context).
• /key updates the API key and saves it to the global .env.

Updates:

• At startup, the agent compares its embedded version with the GitHub VERSION file and reminds you if it is behind.
• /update clones the GitHub repo to a temp directory, rebuilds the binary, installs it to the current executable path (or /usr/local/bin/thu-agent on Linux), then removes the temp clone. On Windows it stages a post-exit replacement of the running .exe.

Other:

• /pwd prints the current working directory.
• /help shows the command list.
• /exit quits the agent.

Notes:

• The MCP server in server.py is separate from the interactive agent in agent.py.