Repository navigation
Releases
ReverseAI 0.9.0
Version 0.9.0: resumable, model-independent hive runs. Explicit local fallback after a provider refusal (--fallback-*, scoped to a run, roles or named bees) and recovery with --resume-refused; distinct bee outcomes, a saved frozen plan, a physical CLI-start limit and forecast, and bounded group planning (phase B). Configured self-hosted servers are checked before planning through the actual CLI: a real ReverseAI MCP call on a surveyed function, structured output and active context metadata from Ollama or vLLM (--required-context). Consecutive MCP argument-schema rejections stop the agent (--max-tool-rejections); they are not provider refusals. Provenance distinguishes server_kind cloud, local and remote. The user chooses the model; passing the startup check is no guarantee of completion. A development pilot using the earlier startup check stopped in the first planner with 143/143 observed export_function calls rejected for argument schema; no local reconstruction was accepted.
ReverseAI 0.8.0: reconstruction exports and the AI analyst team
ReverseAI 0.8.0 adds reconstruction exports and an optional AI analyst team, the hive, that rebuilds native programs as C or C++ from ReverseAI's evidence.
Reconstruction exports
reverseai prepare(MCPprepare_file) runs Ghidra's analysis once per file and keeps the project in a private, bounded cache.- Whole programs and single functions are exported in pages: functions, calls, references, strings, complete pseudocode and assembly, over the command line and MCP.
- MCP answers are compact by default; their text escapes every non-ASCII character, so invisible or bidi characters from a file never reach a model unseen.
The hive (experimental, python3 -m hive run from a checkout)
- Queen in code: plans with one planner agent, freezes a shared contract header, schedules the bees, compiles, links and computes the verdict. She never writes code, votes or overrules a check.
- Builder bees in parallel: one module each, each with its own ReverseAI MCP server and a folder it alone can write; Claude bees have no shell and no web.
- Independent council: three non-authors per module try to refute it against the assembly; every seat must answer, and code changed after a council read goes back to the council.
- Deterministic final word: an evidence checker re-runs every cited question; the compiler runs in a closed file system with inline assembly disabled; the original runs on conformance cases only with
--run-original. - A pre-flight stops packed programs before any agent is spent; C or C++ is chosen from the program; optional Codex bees are refused while their sandbox can reach the network.
Measured
| Program | Functions rebuilt | Council | Evidence checker | Behaviour |
|---|---|---|---|---|
| wordtally (stripped, never published) | 9 in 2 modules | 2/2 accepted | 16/16 claims | 43/43 and 92/92 cases identical |
| lz4 1.10.0 (254 KB) | 282 in 30 modules | 30/30 accepted | 448/448 claims | 68/68 and 127/127 independent cases identical |
The lz4 run used the queen as it was before the review of 2026-11-01, which let 13 modules change after their council read; see the note in the README.
Proofs
- Whole suite, 2,200 tests, and the same wheel on macOS ARM64 (Python 3.15) and in a Linux ARM64 container (Python 3.12).
- R1 native verifier 67/67 and Ghidra containment 32/32 on the new pinned Ghidra r7 image.
Limits
Contained tools run on ARM64 (Apple silicon, Linux ARM64) today; x86-64 images are next. The hive needs Claude Code logged in with your own subscription (no API key) and a Clang compiler it can confine. Use it only on programs you are entitled to analyse and rebuild, and never run malware with --run-original. Report security problems through the private vulnerability form (see SECURITY.md).
ReverseAI 0.7.0: MCP server for AI agents
MCP server for AI agents. reverseai mcp gives Claude Code, Codex, Cursor and any other Model Context Protocol client the same verified questions as the command line, over standard input and output, confined to a folder of files you approved (protocol revision 2026-07-28, older clients supported). Contained tools run only when you enable them, and text quoted from the files is marked as untrusted.
Easier to start. Use cases with runnable examples, an infrastructure diagram, START.md, a setup prompt for coding agents and verified install paths (pip, uv, wheel, editable).
Tested and proven. Tested on real local files (Mach-O, PE, APK, npm tarball, Electron archive); the whole suite and the installed package are proven on macOS ARM64 (Python 3.15) and in a Linux ARM64 container (Python 3.12).
Install from a clone, offline and without dependencies: see the quick start. Earlier versions are listed in the changelog.