How does ghidra-mcp interact with ghidra-server?

[oliv3r]

I've been running ghidra server for a while, which is more of a collaboration/storage solution if you will. Since ghidra-mcp is also headless, I wonder what/how that relation (could) looks like?

[bethington]

Sorry for the very late reply on this — the answer has gotten more complete since you asked.

ghidra-mcp composes well with a Ghidra Server. Concretely:

• /server/ endpoints* — auth (/server/connect), repo and folder listing, checkout, checkin, launch_codebrowser, status checks, etc. So you can do the full collaboration loop (lock → edit → commit) from MCP.
• GUI plugin (FrontEnd mode) — runs in the Project Manager window and uses a FrontEndProgramProvider that opens server-managed programs on demand from DomainFile. That means an AI agent can read/write annotations on shared-server files the same way a human collaborator would, and save_program + checkin persists the edits back to the server for the rest of the team.
• Multi-version / multi-binary — every program-scoped tool accepts an optional program query parameter, so you can target a specific server-managed program by path (?program=/Vanilla/1.00/D2Common.dll) without juggling switch_program.
• Headless container — the /server/* listing/checkout side works, but headless does not yet open server-checked-out programs end-to-end. That gap is tracked in Can't open projects from Shared Ghidra Server via Headless API #119.

If you're running both ghidra-mcp and ghidra-server, the typical setup is: server stays the source of truth for the team, ghidra-mcp's GUI plugin runs locally in your FrontEnd, and AI tooling reads/writes against the shared project just like another teammate would.