Apideck CLI — Agent Interface That Uses 32x Fewer Tokens Than MCP
Apideck has released the Apideck CLI, a static Go binary that serves as an AI-agent interface to the Apideck Unified API while consuming dramatically fewer tokens than MCP servers. The blog post, trending on Hacker News with 116 points, presents striking benchmarks: MCP tool definitions consume 55,000+ tokens before an agent processes a single message, while the Apideck CLI replaces this with an approximately 80-token system prompt.
The key insight is progressive disclosure through --help commands instead of loading the entire API schema upfront. In testing, MCP used 4-32x more tokens for identical operations and showed a 28% failure rate, while the CLI runs locally without remote dependency issues.
Safety is structurally enforced in the binary rather than via prompt-based guardrails: GET requests are auto-approved, POST/PUT/PATCH require confirmation, and DELETE is blocked by default. The CLI works with any agent framework that can execute shell commands — no MCP protocol support required.
This challenges a core assumption in the MCP ecosystem: that agents need rich tool schemas loaded into context. Apideck argues that agents can discover capabilities progressively, just like human developers use --help.
GitHub: https://github.com/apideck-libraries/cli
Blog: https://www.apideck.com/blog/mcp-server-eating-context-window-cli-alternative
← Back to all articles
The key insight is progressive disclosure through --help commands instead of loading the entire API schema upfront. In testing, MCP used 4-32x more tokens for identical operations and showed a 28% failure rate, while the CLI runs locally without remote dependency issues.
Safety is structurally enforced in the binary rather than via prompt-based guardrails: GET requests are auto-approved, POST/PUT/PATCH require confirmation, and DELETE is blocked by default. The CLI works with any agent framework that can execute shell commands — no MCP protocol support required.
This challenges a core assumption in the MCP ecosystem: that agents need rich tool schemas loaded into context. Apideck argues that agents can discover capabilities progressively, just like human developers use --help.
GitHub: https://github.com/apideck-libraries/cli
Blog: https://www.apideck.com/blog/mcp-server-eating-context-window-cli-alternative