| description | Install AgentDescent in one command or with pip, connect Claude Code, Codex, OpenCode or DeepSeek Harness, and ask your agent to improve a skill. |
|---|
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.
<section class="qs3-step" id="connect"><div class="qs3-step-head"><span>01 / INSTALL</span><h2>One command<br>to get set up.</h2><p>The script installs current <code>main</code> through pip (not a pinned release) and connects supported agents it finds on your PATH.</p></div><div class="qs3-install-card"><div class="qs3-tabbar" role="tablist" aria-label="Installation method"><button role="tab" aria-selected="true" class="active" data-install-mode="script">One-command setup</button><button role="tab" aria-selected="false" data-install-mode="pip">Install with pip</button></div><div class="qs3-terminal"><div class="qs3-terminal-head"><span>TERMINAL</span><button id="qs3-copy-install" class="exp-copy">Copy</button></div><pre id="qs3-install-command">curl -fsSL https://raw.githubusercontent.com/ThinkFlowLab/agentdescent/main/scripts/install.sh | bash</pre></div><div class="qs3-install-foot"><span id="qs3-install-note">Installs AgentDescent and connects the agent CLIs already on your PATH.</span><a href="https://github.com/ThinkFlowLab/agentdescent/blob/main/scripts/install.sh">Read the script ↗</a></div></div></section>
<section class="qs3-step qs3-connect" id="agent"><div class="qs3-step-head"><span>02 / CONNECT</span><h2>Open your agent.</h2><p>Choose the agent you use. The command below shows how to launch or verify its AgentDescent integration.</p></div><div class="qs3-host-card"><div class="qs3-host-tabs" role="tablist" aria-label="Choose your agent"><button role="tab" aria-selected="true" class="active" data-host="claude">Claude Code</button><button role="tab" aria-selected="false" data-host="codex">Codex</button><button role="tab" aria-selected="false" data-host="opencode">OpenCode</button><button role="tab" aria-selected="false" data-host="dsh">DeepSeek Harness</button></div><div class="qs3-host-output"><span id="qs3-host-label">LAUNCH CLAUDE CODE</span><p id="qs3-host-help">The installer writes a local plugin directory. Launch Claude Code with it.</p><div class="qs3-terminal compact"><div class="qs3-terminal-head"><span>TERMINAL</span><button id="qs3-copy-host" class="exp-copy">Copy</button></div><pre id="qs3-host-command">claude --plugin-dir ~/.agentdescent/plugins/claude-code</pre></div></div></div></section>
<section class="qs3-step qs3-ask" id="ask"><div class="qs3-step-head"><span>03 / ASK</span><h2>Tell it what<br>to improve.</h2><p>Give your agent a target and examples. It shows an estimate before the run and a diff before applying changes.</p></div><div class="qs3-prompt"><div class="qs3-prompt-top"><span>PASTE INTO YOUR AGENT</span><button id="qs3-copy-prompt" class="exp-copy">Copy prompt</button></div><p id="qs3-prompt-text">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.</p><div class="qs3-prompt-foot"><span>Plan first</span><span>Review the diff</span><span>Apply with approval</span></div></div></section>
<section class="qs3-more"><h2>Good to know.</h2><details><summary>What about provider keys and cost?</summary><p>A model-backed run needs a configured provider key. Run <code>agentdescent doctor</code> to check your setup. Your agent shows the plan and estimated call count before starting. <code>agentdescent demo</code> runs offline without a key.</p></details><details id="source-examples"><summary>Can I explore the algorithm ports?</summary><p>Research examples are source-only and are not included in the wheel. Clone the matching tag:</p><pre>git clone --branch v0.9.0 https://github.com/ThinkFlowLab/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]>=0.9.0" 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/ThinkFlowLab/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.
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.