Start inside
your agent.
The modular evolution engine from the homepage also works in Claude Code, Codex and other coding agents. Follow three short steps to connect it and review your first improvement.
One command
to get set up.
The script installs current main through pip (not a pinned release) and connects supported agents it finds on your PATH.
curl -fsSL https://raw.githubusercontent.com/Birfy/agentdescent/main/scripts/install.sh | bash
Open your agent.
Choose the agent you use. The command below shows how to launch or verify its AgentDescent integration.
The installer writes a local plugin directory. Launch Claude Code with it.
claude --plugin-dir ~/.agentdescent/plugins/claude-code
Tell it what
to improve.
Give your agent a target and examples. It shows an estimate before the run and a diff before applying changes.
Improve the skill at ~/.claude/skills/pdf-audit against eval/cases.jsonl. Show me the plan and estimated calls before starting. Show the diff before applying.
Good to know.
What about provider keys and cost?
A model-backed run needs a configured provider key. Run agentdescent doctor to check your setup. Your agent shows the plan and estimated call count before starting. agentdescent demo runs offline without a key.
Can I explore the algorithm ports?
Research examples are source-only and are not included in the wheel. Clone the matching tag:
git clone --branch v0.7.0 https://github.com/Birfy/agentdescent cd agentdescent pip install -e ".[dev]" python -m examples.run_demo
All 23 algorithm ports print their configuration with --dry-run and make no model call. The eight benchmark-faithful ports return before loading data; the fourteen MethodPolicy ports build their policy first, so a real-benchmark port may download and cache its split. See the port catalog.
Can I install from PyPI?
Yes. pip install "agentdescent[mcp]>=0.7.0" installs this release’s CLI and MCP tools. Python 3.9 supports the core package; MCP requires Python 3.10+. The one-command and pip tabs above install current main, not a pinned release. See the plugin guide for the release workflow.
Can I work from a source checkout?
Clone main and install the MCP extra from that checkout (Python 3.10+):
git clone --branch main https://github.com/Birfy/agentdescent cd agentdescent pip install -e ".[mcp]"
This provides an editable checkout for the CLI, MCP tools and source-only examples. Use pip install -e ".[dev]" in the same checkout for development and the research examples.
How do I build these docs?
From a checkout: pip install -e ".[docs]", then mkdocs serve for a local preview or mkdocs build --strict for a site build.
One path into a bigger framework.
This agent workflow is a starting point. To evolve custom artifacts or plug in different algorithms, explore the same evolve() engine shown on the homepage.