Objdiff CLI MCP Server

[HaydnTrigg]

MCP server for AI-driven decomp matching.

Important

• Benchmarking of this MCP server has been limited. Testing so far has used Anthropic models exclusively due to other AI models being incapable of producing usable output on tested projects.
• A future experiment (not part of this PR scope) is switching to a TOML/Markdown output format, which reportedly could cut token usage further.

Usage

Usage: objdiff-cli mcp [--transport <transport>] [--bind <bind>] [--project <project>]

Starts an MCP server exposing objdiff's diffing for decompilation matching.

Options:
      --transport <transport>  Transport: "stdio" or "http". (Default: stdio)
      --bind <bind>            Bind address for the http transport. (Default: 127.0.0.1:3001)
      --project <project>      Optionally remember a project directory on startup.

Runs persistently over stdio (--transport stdio) or a shared HTTP instance (--transport http --bind ...). State (loaded project + diff config) is shared across sessions for the life of the process.

Tools:

• open_project / list_units - load objdiff.json, refer to units by name
• list_units(filter?) - units + resolved target/base paths
• diff_function(symbol, unit | target+base, config?) - side-by-side per-instruction diff + match% for one function
• diff_overview(unit | target+base, only_mismatches?, limit?, config?) - functions ranked worst-match-first
• build(unit, target?) - run the project's build command for a unit (via objdiff-core run_make)
• set_config(key, value) - persistent diff/disassembly config overrides
• list_symbol_mappings(unit) - manual target=base mappings
• set_symbol_mapping(unit, target_symbol?, base_symbol?)
• version()

diff_function / diff_overview accept either a project unit or explicit target/base object paths, plus per-call config overrides.

Testing

This code has been tested for over a month across two extremely large decompilation projects. Using three Claude Max 20x subscriptions using models Opus 4.8, Opus 5, Fable 5 in Ultracode mode. Rough estimates put token usage 30–50% lower than the command-line and logging workflow depending on context.

This seems to be caused by the models having short-circuits to use MCP servers compared to thinking about command line usage (we've switched over to a few other MCP servers before this and saw similar token usage reductions). Due to Anthropic mostly being a black box one can only speculate.