Skip to main content
The bash_tool primitive runs arbitrary shell commands inside the chat’s sandbox. Combined with the pre-installed stack, the model can do anything a developer can do.

​
Languages

RuntimeVersion
Python3.12.3 (107 packages, incl. pandas/numpy/opencv/playwright)
Node.js22 (21 packages, incl. TypeScript/pdf-lib/mermaid-cli)
Bunlatest
JavaOpenJDK 21

​
Behavior

  • Streaming output. The tool streams stdout/stderr back to the model with 15-second heartbeats — no silent hangs.
  • Truncation-aware. Output caps at 30K chars (first 15K + last 15K), so a runaway grep doesn’t blow the context window.
  • Semantic exit codes. grep returning 1 is treated as “no match,” not an error.
  • Timeout. Commands die after COMMAND_TIMEOUT seconds (default 120). Override per call.

​
Typical patterns

Data wrangling 
python3 -c "
import pandas as pd
df = pd.read_csv('/mnt/user-data/uploads/sales.csv')
print(df.groupby('region')['revenue'].sum())
"
Media 
ffmpeg -i /mnt/user-data/uploads/clip.mov -vf scale=720:-1 /mnt/user-data/outputs/clip.mp4
Git 
git clone https://github.com/org/repo /tmp/repo && cd /tmp/repo && npm test
Server 
cd /home/assistant && python3 -m http.server 3000
The sandbox has its own network namespace — long-running servers stay up until the container is GC’d.

​
When to prefer a skill

If a task fits a skill (docx, xlsx, pptx, pdf, playwright-cli, …), the skill has curated scripts that are faster, safer, and produce more polished output than hand-rolled bash. The model picks automatically based on the <available_skills> block in its system prompt.