Most coding agents burn through tokens like there's no tomorrow. Maki, a new Rust-based terminal coding agent from developer tontinton, takes a different approach. It parses code into compact skeletons using tree-sitter, showing imports, type definitions, and function signatures with line ranges instead of full file contents. The result: it adds roughly 59 tokens per turn but saves 224 on reads. Since reads make up about 65% of token usage, this adds up fast. Benchmarks show a 40% cost reduction and 2x speed improvement. Maki also runs a sandboxed Python interpreter. Models can execute code, filter results, and process data without dumping intermediate output into your context window. A script that might consume 18,000 tokens gets compressed to around 40 tokens of actual output. The agent picks different model tiers for different subtasks too. Claude Haiku handles grep-heavy research. Claude Opus handles architecture decisions. Each subagent gets its own chat window, so you can see what's happening and intervene if needed. The interface runs at 60 FPS as a native binary with no JavaScript runtime. You get syntax highlighting and fuzzy search out of the box. Memory persists across sessions, and MCP support plus a headless mode that plays nice with Claude Code output. The permission system parses bash commands with tree-sitter, so it catches dangerous side effects like 'git diff && rm -rf /' instead of blindly approving it. Over 90% of Maki's own codebase was written by the agent itself, which is either a strong endorsement or clever marketing, depending on your perspective.