OnPrem.LLM AgentExecutor: Launch Sandboxed AI Agents with Built-in Tools

The AgentExecutor from OnPrem.LLM enables autonomous AI agents to execute complex tasks using both cloud and local models. The pipeline works with any LiteLLM-supported model that supports tool-calling, including cloud models like OpenAI's GPT-5.2-Codex, Anthropic's Claude Sonnet 4.5, and Google's Gemini 1.5 Pro, as well as local models through Ollama, vLLM, or llama.cpp.
Built-in Tools
By default, AgentExecutor provides access to nine built-in tools:
read_file- Read complete file contentsread_lines- Read specific line ranges from filesedit_file- Edit files via find/replacewrite_file- Write complete file contentsgrep- Search for patterns in filesfind- Find files by glob patternrun_shell- Execute shell commandsweb_search- Search the web for informationweb_fetch- Fetch and read content from URLs
Configuration Examples
You can customize tool access based on your security requirements:
# Use defaults (all tools including shell):
executor = AgentExecutor(model='anthropic/claude-sonnet-4-5')
Defaults but no shell access (safer):
executor = AgentExecutor(
model='openai/gpt-5-mini',
disable_shell=True
)
Minimal tools:
executor = AgentExecutor(
model='openai/gpt-5-mini',
enabled_tools=['read_file', 'write_file']
)
Web research only:
executor = AgentExecutor(
model='openai/gpt-5-mini',
enabled_tools=['web_search', 'web_fetch']
)
Sandboxed Execution
For security, you can run agents in ephemeral containers using sandbox=True. This is important because agents with shell access can potentially read or modify files outside the working directory. The agent operates within the specified working directory and cannot read or write outside it unless given shell access.
Basic example with sandboxing:
executor = AgentExecutor(
model='anthropic/claude-sonnet-4-5',
sandbox=True,
)
result = executor.run(
task="""
Create a simple Python calculator module with the following:
- calculator.py with add, subtract, multiply, divide functions
- test_calculator.py with pytest tests
- All tests must pass
""",
working_dir='./calculator_project'
)
This approach is useful for developers who need to automate coding tasks while maintaining security boundaries. The tool requires installing PatchPal with pip install patchpal.
📖 Read the full source: HN AI Agents
👀 See Also

FixAI Dev: A Consumer Rights Game Using Claude Haiku with Strict JSON Contracts
A developer built a browser game where Claude Haiku acts as a corporate AI denying consumer requests; players argue using real consumer protection laws across 37 cases in EU, US, UK, and Australia. The architecture uses Haiku for language only, with server-side game logic and strict JSON contracts between components.

Benchmark Results: 331 GGUF Models Tested on Mac Mini M4 16GB
A benchmark of 331 GGUF models on a Mac Mini M4 with 16GB RAM reveals only 11 Pareto-optimal models, all Mixture-of-Experts architectures. Mixture-of-Experts models dominate performance with median 20.0 tokens/second versus 4.4 for dense models.

Custom Status Line for Claude Code Shows Context Usage, Cost, and Git Branch
A Reddit user created a bash script that leverages Claude Code's statusLine setting to display real-time information including context window usage, session cost, active model, and current git branch. The script requires jq and is available on GitHub.

Offline Voice-to-Text Tool for macOS Using Local Whisper via MLX
A developer has open-sourced whisper-dictate, a macOS tool that provides fully offline voice-to-text transcription with real-time translation capabilities using Whisper running locally through MLX on Apple Silicon. Transcription takes about 500ms after speaking stops.