Most coding assistants phone home to the cloud. AgentTool does the opposite — it runs an agentic coding loop entirely on models served from your own machine, keeping your code local and your context private.
What it does
A terminal-based agentic coding assistant powered by locally hosted large language models:
- OpenAI-compatible tool-calling loop that auto-discovers Ollama and LM Studio model servers.
- Tools for web search, file operations, and shell command execution.
- Robust fallback JSON parsing for non-conformant model outputs, plus graceful interrupt handling.
- Comprehensive pytest suite with CI integration for reliable, test-backed behavior.
Figure 1 — The agent observes, picks a tool, acts (search/file/shell), and iterates toward a goal.
Architecture
The assistant is a loop: observe the current state, let the model choose a tool call, execute it, feed the result back, and repeat until the task is done. The tool-calling protocol is the OpenAI-compatible schema, so any server exposing that interface — Ollama, LM Studio — works without code changes.
Figure 2 — The bundled toolset spans web search, file manipulation, and shell execution.
Key challenges
- Model non-conformance — local models often emit malformed JSON; the fallback parser recovers usable calls.
- Interruption safety — Ctrl-C must unwind cleanly without corrupting state.
- Local-first reliability — tests guard the tool loop so regressions surface before they reach your terminal.
Repository
Further reading
- The changing work of software engineers — the broader shift toward agentic coding workflows.