Repository navigation
MCP Hub
decompile-release-bot edited this page Nov 21, 2026
·
1 revision
The hub layer that lets decompile fan out to multiple backend MCP
servers behind one call.
hub_connect host port? name? # register an upstream MCP
hub_disconnect name
hub_list
hub_call name tool args # one backend
hub_consensus_call tool args backends? # fan out to N, return majority / all
smart_decompile path # auto-pick + fallback chain
The composite most users should call. Behavior:
- Pick the preferred backend for the target file type
(
pick_re_backendunder the hood). - If that backend is already running (
re_backend_status), use it. - Else try to
launch_re_backendit. If launch fails, fall back to the next-best backend for the same file type. - Call the decompile action.
- Store the result under a content-hash key via the decompile cache
(
decompile-cache.ts). Subsequent calls on the same content hash return instantly.
Fallback chain (per file type):
| File type | Order |
|---|---|
| PE x86-64 | IDA → Ghidra → Binja → radare2 |
| ELF x86-64 | IDA → Ghidra → Binja → radare2 |
| Mach-O | Hopper → IDA → Ghidra → Binja → radare2 |
| DEX | JADX → Ghidra |
| WASM | Binja → radare2 |
| .NET | ilspy → dnspy → Ghidra |
hub_consensus_call tool=decompile args={"addr":"0x401000"} backends=["ghidra","ida","binja"]
Dispatches to each backend in parallel, scores the results, returns:
- The majority answer if N agree
- Or all results tagged with their backend when they diverge, for the caller to reconcile
Useful when a decompiler's output is suspect (e.g. heavy ORM / VM protection where Ghidra's model differs from IDA's).
decompile-cache.ts stores decompile outputs keyed by
sha256(binary_bytes) + tool + args_hash. Default location:
.cache/decompile/. Inspect with cache_stats / clear with
cache_clear.
Cache hits are reported in every tool response so the agent knows when a result was reused rather than recomputed.
decompile v0.0.1 · MIT · npm · install.sh · install.ps1
Getting started
Core
Advanced
Project