BendClaw

A self-evolving agent engine — fully observable, token-efficient, built for long-running complex work.

Everything Claude Code / Codex can do — half the tokens, fully observable, in 20k+ lines of Rust.

The engine behind evot.ai

News · Why · Install · Quickstart · Dev · Community

📢 News

• 2026-04-11 🧠 [Memory] Auto-load Claude Code memories for the current project.
• 2026-04-10 🎯 [Plan Mode] Add ask_user tool for interactive option selection.
• 2026-04-10 🌐 [Web Fetch] SPA support via headless Chrome fallback.

---

⚡ Why BendClaw

Claude Code and Codex dump everything into context — bloated outputs, stale history, noise. Tokens wasted. Quality drops. No visibility into why.

BendClaw doesn't waste a single token — and proves it.

• Clean context, always. Every prompt to the LLM is minimal, high-signal, zero-waste.
• Blazing fast. Fewer wasted tokens → fewer turns → complex tasks done in half the time.
• Fully observable. Every LLM call, tool execution, and compaction tracked end-to-end. This data feeds back into the engine — BendClaw evolves its strategy so the next prompt is always leaner than the last.

Not a CLI wrapper. The agent engine you build on — ships with REPL, CLI, and server.

Installation

One-liner (recommended)

curl -fsSL https://raw.githubusercontent.com/evotai/bendclaw/main/install.sh | bash

Build from source

cargo install --git https://github.com/evotai/bendclaw.git

Quickstart

Create ~/.evotai/bendclaw.env:

# Provider: "anthropic" or "openai"
BENDCLAW_LLM_PROVIDER=anthropic

# Anthropic
BENDCLAW_ANTHROPIC_API_KEY=sk-ant-...
BENDCLAW_ANTHROPIC_MODEL=claude-sonnet-4-20250514

# OpenAI
BENDCLAW_OPENAI_API_KEY=sk-...
BENDCLAW_OPENAI_MODEL=gpt-4o

Only the active provider's keys are required. Set BENDCLAW_LLM_PROVIDER to switch.

bendclaw                              # interactive REPL
bendclaw -p "summarize today's PRs"   # one-shot task

CLI flags & options

Flag	Description
-p, --prompt	Run a single prompt and exit
--resume <id>	Resume an existing session
--model <model>	Override the configured model
--output-format text|stream-json	Output format (default: text)
--max-turns <n>	Limit agent turns (default: 512)
--max-tokens <n>	Limit total tokens
--max-duration <secs>	Session timeout in seconds (default: 3600)
--append-system-prompt "..."	Inject extra system instructions
--verbose	Enable info-level logging

Development

make setup      # install Rust toolchain, git hooks
make check      # fmt + clippy
make test       # unit + integration tests

Community

GitHub Issues Bug reports / Feature

Twitter @Evot_AI Updates & announcements

team@evot.ai Reach the team directly

License

Apache-2.0

---

Built with 🦀 by Evot AI