Detailed Analysis
A Show HN post introduces Keen Code, a solo-built, open-source coding agent written in Go that positions itself as a lean, context-efficient alternative to established tools like Claude Code and Codex. Rather than simply cloning existing agent architectures, the developer describes building it through "agentic engineering," using AI-assisted development to bootstrap the tool—including using Keen Code to develop itself. The project supports multiple model providers, MCP (Model Context Protocol) integrations, skills, subagent orchestration, and automatic context compaction, placing it squarely in the growing category of general-purpose CLI coding agents that has exploded since Claude Code's release popularized the pattern.
The most technically interesting contribution is "Turn Memory," a context-management scheme that discards tool outputs between conversational turns while preserving only the trace of which tools were called. The rationale is that most tool results—file reads, bash output, web fetches—are cheap to regenerate and rarely need to persist across turns, so retaining them wastes context window space. The developer notes that Claude Code and similar agents already tend to re-read files redundantly even when they've seen them before, suggesting that discarding stale tool output isn't a huge behavioral departure from how agents already operate, but formalizing it as a deliberate memory policy yields measurable context savings (the post claims context usage can drop from 20% to 1% at the start of a new turn). This is a pragmatic trade-off: it sacrifices some continuity in exchange for dramatically slower context saturation in long multi-turn sessions, which is one of the most persistent pain points in agentic coding workflows.
The second idea, "Skill-Driven MCP," directly echoes and extends Anthropic's own tool-search-tool concept for managing MCP server bloat. Instead of pre-loading full tool schemas for every connected MCP server—a common source of context window pressure once several servers are attached—Keen Code exposes only lightweight skill frontmatter by default. The agent progressively discovers a server's capabilities: first loading a generated skill file listing available tools, then pulling detailed schemas only for the specific tool it intends to call, with results cached locally after first discovery. This lazy-loading approach trades a small number of extra file reads for substantially reduced baseline context consumption, a sensible trade given that MCP schema bloat has become a well-known scaling problem as users attach more integrations to their agents.
Broadly, this release is a data point in the maturing ecosystem around Claude and Claude Code specifically: independent developers are now treating Anthropic's agent design patterns—tool search, skills, compaction, subagents—as a shared vocabulary to build upon, remix, and optimize rather than as proprietary features. Anthropic's decisions to open the Model Context Protocol and to publish design rationale (like tool-search-tool) appear to be seeding a broader wave of experimentation in context-window economics, which is emerging as one of the defining engineering challenges of agentic coding tools as context windows, while growing, remain a bottleneck for long-running autonomous sessions. Keen Code's explicit comparisons to Claude Code and Codex, and its adoption of Anthropic-originated context-optimization philosophy, illustrate how quickly patterns pioneered by frontier labs propagate into smaller, independently maintained tools, and how the coding-agent space is increasingly defined by competition on context efficiency and memory management rather than raw model capability alone.
Read original article →