An AI CLI That Writes Its Own Tools on Demand

What it does Claude Engineer is an interactive Python CLI that wraps the Anthropic API and gives Claude the ability to generate, save, and call its own Python tools mid-conversation. It ships with basic file-system and shell primitives, but the central mechanism is a built-in create_tool function that lets the model write new code, persist it to a tools directory, and immediately import it into the running session. The longer the conversation, the more specialized gear the agent accumulates.

The interesting bit Most agent frameworks arrive with a fixed multitool; this one gives the AI a workbench and lets it build new instruments as the conversation demands. The self-improvement loop is explicit: the agent can author its own extensions, manage their dependencies, and refine them based on usage patterns, all without restarting the REPL.

Key highlights

• Autonomous tool creation: the agent writes, saves, and executes new Python tools during a conversation
• Dynamic module importing system loads freshly created tools without restarting the session
• Token usage tracking via the Anthropic API
• Built-in primitives for file I/O, shell commands (with a 60-second timeout), and web search
• Configurable model choice via the CE_MODEL environment variable, not locked to Opus

Caveats

• The README contains conflicting OS support claims: Requirements list only Windows or Mac OS, while installation instructions also include Linux
• The repository description references “Claude Opus(3-4.7),” which is unclear, though the code itself supports any Claude model
• The project structure shows a tests/ directory, but the README does not describe test coverage or CI status

Verdict Worth a look if you want an agent framework that can adapt its own capabilities without human intervention between turns. Skip it if you need a battle-tested, static toolchain with predictable, fixed behavior.