Every path resolves under the root you named, with no `..` escape, and `run_command` takes an argv list rather than a shell string — so the interesting decision is not which tools you get but how tightly you draw the boundary. Set `AGENT_GODMODE_ALLOWED_COMMANDS` and only those binaries can be launched; leave it unset and anything under the sandbox can run.
A workspace toolkit for building coding agents. It ships five tools — read, write, edit, list and run — every one of them scoped to a single workspace root, and exposes the same operations over MCP and in-process through `AgentWorkspace`. The model and the API key stay in your application; this package is the execution and sandboxing half.
- A UTF-8 text file read with an optional line range and byte cap — `read_file`
- Create, overwrite or append text, with parent directories made for you — `write_file`
- Search-and-replace in an existing file, where `old_string` must match exactly once unless `replace_all` is set — `edit_file`
- Directory listings with optional recursion, glob, depth cap and dotfile control — `list_files`
- A subprocess started from an `argv` list only, never a shell — `run_command`
- `OPENAI_TOOL_DEFINITIONS` and `SYSTEM_PROMPT_V1` so the same tool shapes can be handed to a chat-completions loop
`pip install agent-godmode`, then run the `agent-godmode` CLI over stdio. The workspace root is mandatory: set `AGENT_GODMODE_ROOT` or pass `--root`. Optional limits are all environment variables — `AGENT_GODMODE_MAX_READ_BYTES` (default `512000`), `AGENT_GODMODE_COMMAND_TIMEOUT` (default `120`), `AGENT_GODMODE_MAX_COMMAND_OUTPUT_BYTES` (default `256000`), `AGENT_GODMODE_LIST_MAX_ENTRIES` (default `2000`) and `AGENT_GODMODE_ALLOWED_COMMANDS`, a comma-separated list of basenames allowed as `argv[0]`. The published version is 0.1.9.
One command — pip install agent-godmode