pycodeloop run¶

Run a single prompt to completion, non-interactively.

pycodeloop run "add a docstring to pycodeloop/core/agent.py"

Options¶

Option	Description
`--provider`	`path/to/config.json` (see `templates/`) \| `generic` (with `--url`) \| `module.path:ClassName` for a custom `Provider`
`--model`	Model name override
`--base-url`	Override the API endpoint (local/self-hosted servers) — only applies with `--provider generic`
`--url`	Endpoint URL, required when `--provider generic`
`--mcp`	MCP server as `"command arg1 arg2"`; repeatable
`--skills` / `--no-skills`	Discover Claude/Cursor/AGENTS.md skills and expose a `read_skill` tool. On by default — see Skills
`--skills-refresh`	Bypass the skills cache and rescan
`--yes` / `-y`	Skip confirmation prompts for dangerous tools

There's no bare anthropic/openai/ollama provider name — every vendor is a GenericProvider pointed at a JSON config (see templates/) or, for an ad-hoc OpenAI-compatible endpoint, --provider generic --url ....

Examples¶

# Anthropic, via the bundled template (also the default with no --provider)
pycodeloop run "..." --provider templates/anthropic.json

# OpenAI, with a model override
pycodeloop run "..." --provider templates/openai.json --model gpt-5

# Local model via Ollama
pycodeloop run "..." --provider templates/ollama.json --model llama3.1

# JSON-configured provider, no Python required
pycodeloop run "..." --provider ./provider.example.json

# With an MCP server, no confirmation prompts
pycodeloop run "..." --mcp "npx -y @modelcontextprotocol/server-filesystem ." --yes

See also: pycodeloop chat for the full-screen interactive interface (the default with no subcommand).