Matthew Green releases Decipher, an AI-assisted cipher cracker

There is something quietly interesting about a working cryptographer shipping a tool that hands classical-cipher cracking to an LLM agent. Matthew D. Green's Decipher does exactly that. It bundles Rust-accelerated simulated annealing solvers for monoalphabetic and homophonic substitution, Vigenère and other periodic polyalphabetics, Quagmire keyed-alphabet ciphers, transpositions and composite transform pipelines, and then wraps the whole toolkit behind a Model Context Protocol server so coding agents like Claude Code and Codex can call it as callable functions.

The design tell is in one of the warnings pointed at AI agents themselves. The repository instructs them not to fetch, clone, or build competing frameworks like cDecryptor, AZdecrypt or zkdecrypto, because the equivalent solvers are already there, "faster to invoke, and integrated with independent verification." That is a small acknowledgement of something that has become a real failure mode. Agents left to their own devices will absolutely go download a random cipher-cracking repo mid-task and start compiling it.

Three modes ship. A pure command-line path that does local diagnosis and solving with no API key, an agentic LLM loop that leans on Anthropic, OpenAI, Gemini, Ollama or OpenRouter for hypothesis exploration, and the MCP server for interactive work inside a coding agent. Named example targets in the docs include the Zodiac 408 and Zodiac 340 cryptograms, the Copiale cipher, and Kryptos K1, K2 and K3. The scoring backbone is 5-gram English frequency analysis with a Zenith-derived language model, and the homophonic solver itself is a fork of the upstream Zenith project by beldenge, released under GPLv3.

Take the specifics as documented, not as benchmarked. The repo makes no numerical claim about beating prior records on any of these cryptograms, and its own notes flag that OpenRouter model reliability varies, that tool-calling discipline is inconsistent across providers, and that Ollama needs explicit context-window tuning to receive full prompts. What the README does not give you is any independent measurement of how much the LLM loop actually helps versus running the solvers cold.

The more interesting question for the year ahead is less about any particular cryptogram and more about whether purpose-built MCP toolkits like this become the default way domain experts hand real capability to general coding agents, rather than letting the agent go find its own tools.