Mcp2cli登場、MCPより96-99%少ないトークンで全APIをCLI化 Show HN: Mcp2cli – One CLI for every API, 96-99% fewer tokens than native MCP

Mcp2cli is an open-source project that wraps arbitrary APIs as a single command-line interface designed for LLM agents, claiming a 96-99% reduction in token usage compared to native Model Context Protocol (MCP) servers. It tackles one of the more practical pain points of the current agent stack: context bloat caused by tool schemas.

MCP, introduced by Anthropic, has become a de facto standard for connecting LLMs to external tools and data sources, with adoption across Claude Desktop, Cursor, and a growing ecosystem of community servers. The protocol's design, however, front-loads every connected server's tool definitions into the model's context window. Once a user wires up several MCP servers, tens of thousands of tokens can be consumed before any real work begins, and agents must reason over a long catalog of tools they may never use.

Mcp2cli inverts that pattern. Instead of exposing each API operation as a discrete MCP tool with a full JSON schema, it presents the API as a CLI with subcommands. The agent sees only a lightweight entry point and discovers capabilities on demand by invoking --help on specific subcommands, much like a human developer exploring an unfamiliar tool. This brings classic Unix-style discoverability into the LLM reasoning loop, keeping the initial context small and deferring detail until it is actually needed.

The idea is not isolated. Cloudflare recently described a related pattern called code mode, in which tools are exposed as code that the agent imports as needed, and Anthropic itself has discussed treating tools as on-demand modules rather than always-on schemas. Together these approaches suggest the agent community is shifting away from eager schema injection toward lazy, code-like discovery, with token economics as a primary driver.

That said, the headline 96-99% figure should be read with caution. Such savings likely reflect specific benchmarks comparing a fully loaded MCP setup against a minimal CLI entry point, and real-world gains will depend on the granularity of the underlying API, how often the agent needs to consult help text, and how chatty the resulting tool calls become. There may also be tradeoffs around features that MCP clients handle natively, such as streaming progress, structured errors, sampling, or capability negotiation, which a CLI shim cannot fully replicate without extra plumbing.

For teams already feeling the cost and latency impact of large MCP toolsets, Mcp2cli is worth evaluating as a complementary or alternative integration layer, particularly for read-heavy REST APIs where help-driven exploration maps cleanly to subcommands. Whether this style supplants schema-first MCP or merges into it as a discovery optimization remains to be seen, but it reflects a healthy maturation of how the field thinks about giving agents access to the outside world.