every-app-is-open-source

Every App Is Open Source

Give your AI agent a compiled program, and it works out how it's built. Ghidra does the decompiling. This is the agent's side of it: a small set of MCP tools, short answers by default, and evidence you can trace back to the binary.

The name is a provocation, not a license. Seeing how an app works doesn't give you the right to copy it. Read the disclaimer.

What it looks like

An agent connected only to EAOS was given a small compiled program and asked what quote(6) returns. These are its actual tool calls:

Tool call	What came back
search {"kind":"functions","pattern":"quote"}	_quote at 0x100000460
inspect {"addresses":["100000460"],"mode":"all"}	Assembly and decompiled code, saved as evidence fe8589bb…
read_evidence {"id":"fe8589bb…"}	The saved evidence, hash-checked

Its answer: 45. Below 4 the function returns 19, otherwise 7*n + 3, citing the address and the evidence ID.

This is one recorded session on version 0.1.1 with an already-analyzed cache, not a benchmark. The full transcript has every call and its limits.

How it compares

I've compared it with REA and GhidraMCP on three modified programs, with the same model and questions for each tool. The results depend on the target, and EAOS doesn't win every time:

Target	EAOS version	vs GhidraMCP	vs REA
Modified Doom	0.1.0	16% cheaper, 17% faster; all 18/18 correct	36% cheaper, 27% faster; all 18/18 correct
Modified Ollama	0.1.4	28% more expensive, 23% slower; GhidraMCP missed 3 of 195 cases	14% cheaper, 17% faster; both 195/195
Modified Ollama, round 2	0.1.5	6% cheaper, similar time	25% cheaper, similar time

Costs are API-equivalent model costs; timings exclude initial analysis. Each row is three runs per tool, so these are descriptive results, not a general ranking. In round 2 the other tools' missed cases were mostly header order and incomplete error text, not wrong behavior. The modified programs and answer keys are private so the tests can't be spoiled. That means you can read every run summary, scoring review and protocol, but you can't replay these exact comparisons yourself. Every number, loss and limitation is in BENCHMARKS.md.

Get your first result

These instructions target 0.2.0. Before publication, use the reviewed local wheel path in place of the package requirement.

You need an Apple Silicon Mac, uv and Apple's Command Line Tools (xcode-select --install). Windows, Linux and Intel Macs aren't supported yet.

1. Install. This downloads checksum-pinned Java and Ghidra (about 770 MB) and builds Ghidra's Apple Silicon decompiler, which the official download doesn't include:

uvx --python 3.13 --from every-app-is-open-source==0.2.0 eaos setup

Everything goes into ~/Library/Application Support/Every App Is Open Source/. It doesn't touch your shell profile, your system Java or your agent's settings. Versions, sizes, troubleshooting and uninstall are in SETUP.md.

2. Build a tiny test program and analyze it:

mkdir -p /tmp/eaos-example
cat > /tmp/eaos-example/example.c <<'C'
#include <stdio.h>
__attribute__((noinline)) int quote(int n) { return n < 4 ? 19 : n * 7 + 3; }
int main(int argc, char **argv) { puts("eaos fixture anchor"); return quote(argc); }
C
clang -O0 /tmp/eaos-example/example.c -o /tmp/eaos-example/example
uvx --python 3.13 --from every-app-is-open-source==0.2.0 eaos call /tmp/eaos-example/example open_program

You should see classification: computed, an ARM64 language, a SHA-256 and an evidence ID.

3. Connect your agent. For Claude Code:

claude mcp add --transport stdio --scope local eaos -- uvx --python 3.13 --from every-app-is-open-source==0.2.0 eaos mcp /tmp/eaos-example/example

Codex and other MCP clients are in CLIENTS.md.

4. Ask:

Find the quote function. What does quote(6) return? Explain both branches and cite the function address and saved evidence.

The correct answer is 45. This checks that your setup works; it doesn't prove the agent will understand any app. Addresses depend on the compiler, so yours will differ from the example above.

For a standalone program, use the same eaos mcp /absolute/path/to/program command. For an application directory, select its root with an explicit external workspace:

uvx --python 3.13 --from every-app-is-open-source==0.2.0 eaos mcp /Applications/Example.app --workspace /absolute/path/outside-bundles/eaos-workspace

Ask the agent to inventory the root, select one native component by its returned ID, then inspect that component. Full native analysis remains the default; quick mode is opt-in. Individual executable paths inside bundles are still refused; copy the executable out for standalone analysis. For big standalone programs, terminal pre-analysis can finish before you connect your agent.

The tools

I started with seven tools based on the calls agents made across 24 earlier test runs. Native-file sessions now expose eight, including bounded static call traversal. Application-directory sessions expose eleven, adding inventory, explicit component selection and bounded JavaScript inspection.

Tool	What it does
open_program	What the program is: architecture, memory layout, analysis status
search	Find text inside the program and where it's used, or find functions
inspect	Decompiled code, raw instructions, callers and callees
read_memory	Read static memory or Go string tables, with provenance
calculate	Safe arithmetic and byte decoding, so the agent doesn't do math in its head
decode_swift_string	Recover text hidden inside Swift code on Apple platforms
read_evidence	Pull up the full saved evidence behind any reply, hash-checked
trace_calls	Follow bounded static caller/callee relationships, retaining indirect edges and cutoffs as unresolved

Application-directory sessions also expose:

Tool	What it does
inventory_application	Inventory a selected root and supported ASAR members, with explicit gaps and budgets
select_component	Capture and select one inventoried native component outside the original bundle
inspect_javascript	Inspect captured JavaScript source; development profiles add focused preload/module inspection and bounded evidence relationships

0.2.0 extends the initial JavaScript/ASAR inspection from 0.1.9, which was limited to 1 MiB of source.

The 0.2.0 source profiles accept coherent JavaScript captures up to 8 MiB, with explicit bounded syntax-index continuation and fixed result pages. Focus one observed preload API, inspect one selected module or its surrounding loader skeleton, then compose selected saved evidence into a compact source relationship pack. Every candidate retains its component, source, table and evidence identity; unknown bindings remain explicit. Python parser dependencies come with the package; no JavaScript or application is executed.

On official Electron Fiddle 0.39.0, the recorded development check inspected its large bundles and produced 60 selected source nodes, six conditional module-lookup candidates and three preserved narrow preload relationships. It established no complete main-process IPC edge: dynamic factory replacement, loader/interop identity, inherited dispatch and process context remain unresolved. These are local static checks, not a comparative benchmark or a test of reliable agent answers. A separate bounded Ghostty trace still did not establish the complete Swift/C/Zig reload chain. Full Electron support, reconstructed applications and REA parity are not established.

Native traversal also retains available names for unprepared quick-mode candidates. That does not prepare their instruction bodies or establish missing relationships. Full native analysis remains the default.

The exact inputs, outputs and limits of each one are in the tool contract.

Every tool result says what it is

Each tool result is labelled:

• fact: read straight from the captured binary or source
• computed: worked out from it, like decompiled code or a decoded string
• guess: an inference that hasn't been verified

Decompiled code is computed, not the original source. Successful analysis results include saved evidence you can pull up with read_evidence. Status responses and errors don't include evidence IDs; a tool failure makes no analysis claim.

These labels describe the tool results, not your agent's final answer. An agent can still misread correctly labelled evidence. The labels and evidence IDs let you check a conclusion; they don't certify it.

Analyze once, reuse it

Ghidra's analysis is slow, so each standalone program gets its own eaos-analysis/<sha256>/ folder right next to it, with the analyzed project and evidence. Reopening a compatible current-version profile reuses it; different engine, architecture or analysis profiles can require fresh analysis. Application-directory captures and native caches go into the explicit external workspace. Projects made before 0.1.7 are kept but not reused: 0.1.7 analyzes the program again in its new cache format.

Each session works on a read-only copy of the exact bytes it hashed, so changing the original file doesn't affect a session that's already open. If a cached project doesn't match the program's hash, it stops instead of guessing, and it never deletes old evidence.

One Ghidra quirk: it won't create projects inside a folder whose name starts with a dot, like .cache. Keep the program you're analyzing out of those.

Analysis options

Full analysis is the default. The CLI and MCP server accept --analysis-mode quick for partial metadata and requested-function preparation, and --architecture arm64 for an explicit Mach-O slice. Quick function search offers explicit bounded index continuation, separately from view-bound result pagination. Quick mode is limited to 64-bit ARM/x86 Mach-O; it does not recover every Swift name or prove missing references are absent. No new speed or accuracy comparison has been run.

On Apple Silicon, compatible universal inputs select ARM64 automatically. Each captured whole-file hash retains separate architecture/profile/engine caches under profiles/, with shared immutable evidence. Legacy projects are preserved; the new cache format does not reuse them. Status remains available during native operations, with configurable deadlines and cancellation. See setup and the tool contract before using these options.

How the benchmarks are run

The summary is above; the details live in BENCHMARKS.md, one row per target, each linked to its run summaries and scoring.

The rules stay the same every time:

• Same model, fresh sessions, three runs per tool.
• Questions and answer key frozen before the first run. Where a model might know the answers by heart, I secretly change them and rebuild the program first.
• Same head start. Every tool begins with the program already analyzed, outside the timer.
• REA runs the way people actually install it, with its bundled skill and its full tool catalog.
• Everything is kept, including rounds I lost.

The private targets and answer keys mean you can't replay my exact comparisons. You can run the same method on your own target with the benchmark guide.

What's next

Right now it analyzes native programs on Apple Silicon Macs, with selected-root inventory and initial partial JavaScript/ASAR inspection. I'd like it to inspect more kinds of software, but there's no fixed order or schedule: new support follows a demonstrated need and reproducible tests, and each one gets its own benchmark before I claim anything.

Security

The tools expose the selected program, or the explicitly selected application root and component, plus their generated evidence. There's no general file-reading or shell tool. The Python and Java processes aren't sandboxed by the OS, though: they run with your user's permissions, and text recovered from a binary is untrusted. Read the security model before you analyze anything you don't trust.

Found a vulnerability? Report it privately, not in a public issue.

Development

uv build
PYTHONPATH=src uv run --no-project --python 3.13 python -m unittest discover -s tests -v
python3 scripts/check_distribution.py dist

What gets tested where:

Check	Where it runs
Unit, failure-case and packaging tests	CI on Ubuntu, every push
Installed-wheel CLI smoke test	CI on Ubuntu, every push
Native checks with real Ghidra on Apple Silicon (12 tests)	Locally on a configured Mac before each release; skipped in CI

The green CI badge covers the first two rows only. To run the native checks after eaos setup or ./install.sh:

EAOS_NATIVE_TESTS=1 uv run python -m unittest discover -s tests -v

To work from a checkout instead of PyPI, ./install.sh sets up everything locally, then ./eaos mcp /absolute/path/to/program starts the server.

Contributing

This is my project, and I build it myself. Bug reports and ideas are very welcome through issues. I don't accept outside code. If you want to take it somewhere else, fork it; that's what the license is for. More in CONTRIBUTING.md.

License and credits

Apache-2.0. If you build on it, keep the NOTICE file. What's changed in each release is in the changelog.

Ghidra and PyGhidra do the heavy lifting; they're the work of their own contributors. The comparison baseline is LaurieWired's GhidraMCP. The programs I benchmark on, and their licenses, are credited in BENCHMARKS.md.

Made by thomasisalwayslate. I make apps, games and small tools, usually something I want to use myself. More on my site, including the write-ups behind the benchmarks.

If this saved you some digging, a ⭐ helps other people find it.