qrp-mcp

A local cryptographic inventory for developers and AI agents that carries verifiable coverage and its own limits inside the CBOM.

Every signature in your wallet, contract and validator rests on elliptic-curve cryptography, and a large quantum computer breaks it. Plenty of tools will tell you what they found. This one also tells you what it read, what it could not read, and which question it is not answering — in the document itself, where an auditor can check it rather than take your word.

An MCP server that scans a local directory for cryptography that Shor's algorithm defeats — secp256k1, Ed25519, BLS, Schnorr, RSA — plus weak primitives and CI signing commands, and classifies each one: broken by a quantum computer, post-quantum, or neither.

Everything runs on your machine. No network calls, no account, no API key, nothing uploaded. A tool that reads your keys' surroundings has no business phoning home, so this one makes zero outbound connections — enforced by a test, not promised in a paragraph. The only process it starts is a local git, to pin what it read, with the scanned repository's own hooks and filters disarmed.

Why this matters for chains and wallets

Bitcoin and Ethereum authenticate with ECDSA over secp256k1. Solana, Cardano and Polkadot use Ed25519. Ethereum's consensus layer aggregates with BLS12-381. Taproot adds Schnorr.

All four are public-key schemes whose security rests on discrete-log hardness — and all four fall to the same quantum algorithm. The practical consequence is specific: once a public key is exposed, the private key becomes derivable. Reused addresses, on-chain public keys, and long-lived validator keys are where that exposure already exists today.

None of this is a prediction about dates. It is an inventory question: which of my code paths sign with what? That question has an answer right now, and this tool gives it.

Quick start

Add it to your MCP client — no installation step, uvx fetches and runs it:

{
  "mcpServers": {
    "qrp": {
      "command": "uvx",
      "args": ["qrp-mcp"]
    }
  }
}

Then ask your agent:

Scan ~/code/my-protocol for quantum-vulnerable cryptography.

As a Claude Code plugin

The same server, packaged with a skill, so there is no config file to edit:

/plugin marketplace add StanimirTenev/qrp-mcp
/plugin install qrp@quantumreadiness

Then /qrp:pqc-scan in any project. Both routes need uv on your PATH, since uvx is what fetches and runs the server.

Without an agent: write the result to a file

uvx qrp-mcp scan ~/code/my-protocol --out result.json

This produces the same result as the scan_repo tool, written to a file you can read before it goes anywhere. Nothing is sent. --level masked stars out everything in the quoted line but the algorithm name, and --level trimmed removes the line altogether; both keep each file and line number. The SHA-256 of the written bytes is printed, so anyone you send the file to can quote back exactly what they received.

uvx qrp-mcp cbom ~/code/my-protocol --out cbom.json
uvx qrp-mcp closure before.json after.json --out closure.json

cbom writes the same CycloneDX document as the export_cbom tool, masked by default because it is the document built to be sent. closure writes the prove_closure verdict for two saved scans, and names both inputs by their SHA-256, so the three files can be tied together by whoever receives them. Both print the SHA-256 of what they wrote.

Tools

What it looks at

Chain and wallet code — secp256k1, ecrecover, ethers, web3, bitcoinjs, ECPair, btcec, tweetnacl, @solana/web3.js, solana_program, bls12-381, blst, @chainsafe/bls, BIP340/Taproot Schnorr. Solidity (.sol), Rust (.rs), Move and Cairo are scanned alongside Python, Go, Java, JS/TS, Ruby, PHP, C/C++/C# — headers included — PowerShell, Perl and shell.

Classical crypto anywhere else — RSA, DSA, DH, ECDSA and elliptic-curve usage, plus MD5, SHA-1, RC4 and DES/3DES. Not only through library calls: the names the protocols themselves use (ssh-rsa, rsa-sha2-512, ssh-dss, key types such as rsa-2048 and RSA_4096), the modern OpenSSL 3 form where the algorithm is a string argument (EVP_PKEY_Q_keygen(libctx, propq, "RSA", bits), the EVP_*_fetch calls), and the hash idioms people actually write (hashes.MD5(), hashlib.new('md5'), MD5Init, <sha1.h>, Go sha1.Sum).

Cipher suites, decomposed — ECDHE-RSA-AES128-GCM-SHA256 is ECDH and RSA, and DHE-DSS-… and DES-CBC3-SHA name families that reading the suite as one word never sees. IANA TLS_* names are read the same way. A banned component is not a use: !MD5 and !3DES in a cipher list are exclusions, and they are treated as such.

Protocols and dependencies, counted without inventing an algorithm — a pinned TLS version (MinVersion: tls.VersionTLS12, ssl_protocols, SslProtocols.Tls12, SSL3_VERSION), an SSH transport line, and a cryptographic library declared in package.json, go.mod, requirements.txt, Cargo.toml, pom.xml or a Gemfile. These are real facts about a repository and they are reported — in their own buckets, never in detected_algorithms. Each carries a basis: configured_protocol or declared_dependency, never observed_call, because an installed library is not a line of code that calls it. A version is not a verdict either: TLSv1.0 is marked deprecated, and quantum vulnerability is not claimed from a version number, since TLS 1.3 is vulnerable over X25519 and is not over X25519MLKEM768. -SSLv3 in an SSLProtocol line is a ban, and is recorded as one. Manifests are parsed structurally, so a library named in a comment is not a dependency.

Key sizes — a size named on the line (key_size=1024, rsa:1024, genrsa 1024, GenerateKey(..., 1024)) travels with the family, so a weak RSA key is reported as weak rather than as one more RSA. algorithm_key_sizes holds the smallest size seen per family — the figure that answers is anything weak here — and algorithm_key_sizes_observed holds every size seen, which is a different question and became the operational one when RSA-896 was factored on a data-centre fleet: where is RSA-1024 still in use?

From 0.17.0 the size reaches the CBOM component. Where one size was observed it is stated as cryptoProperties.algorithmProperties.parameterSetIdentifier — the field whose own example in the schema is a key length. Where several were, that field is left out and all of them are named in qrp:observedKeySizes instead: choosing one would put a number in the document that looks measured and is not, and a reader could not tell the difference.

Hybrids and composites — the RFC 10024 TLS groups X25519MLKEM768, SecP256r1MLKEM768 and SecP384r1MLKEM1024, OpenSSH 10's default mlkem768x25519-sha256, and the composite certificate algorithms of draft-ietf-lamps-pq-composite-sigs such as id-MLDSA44-RSA2048-PSS-SHA256. A hybrid holds if either half holds, so the post-quantum scheme leads — and the classical half is carried in also_present rather than dropped, since it is the component Shor breaks.

Post-quantum schemes, by family — ML-KEM, ML-DSA, SLH-DSA, Falcon (FN-DSA), NTRU, Classic McEliece, BIKE, HQC, FrodoKEM, XMSS, and the stateful LMS/HSS of SP 800-208 that CNSA 2.0 requires for firmware signing. The nine schemes NIST advanced to its third additional-signatures round in May 2026 — FAEST, HAWK, MAYO, MQOM, QR-UOV, SDitH, SNOVA, SQIsign, UOV — are recognised as candidates, and CROSS as dropped from that process.

Each carries the mathematical family it rests on (structured or unstructured lattice, code-based, hash-based, isogeny-based, multivariate, symmetric-based) and where it stands: standardised, selected, candidate, withdrawn, eliminated or broken. SIKE is reported as broken and HAWK as withdrawn rather than counted as quantum-resistant — "post-quantum" is a category, not an assessment.

Certificates and keys — .pem, .der, .crt, .cer, .cert, .csr, .key, .pub, .p12, .pfx. Algorithms are resolved from the object identifiers inside the DER and from PEM labels and OpenSSH key types, and private key material is reported separately. This does not parse X.509: it matches the object identifiers already in the classifier against the bytes. A file that is not a certificate but contains an identifier's bytes is reported, and the finding says so — "observed in the file; the file was not decoded as a certificate". An earlier version of this paragraph promised that a malformed certificate yields nothing; an independent analysis put a valid RSA identifier into arbitrary bytes and got a finding, which is what the code measures. For a certificate register — issuer, validity, the device — this tool is the wrong instrument and says so: that data comes from a structural parser or an external inventory.

Configuration — nginx.conf, sshd_config, openssl.cnf, swanctl.conf, .ini, .toml, .properties, .hcl, .json, and any YAML that is not a manifest. This is where a TLS or SSH hybrid group is chosen: X25519MLKEM768 and mlkem768x25519-sha256 are almost never strings in code. IKE proposal syntax is read here too — ecp384 is NIST P-384, modp2048 is group 14.

Quantum-resistant mechanisms, not only algorithms — RFC 8784 mixes a postquantum preshared key into IKEv2 key derivation, so a tunnel resists a quantum adversary with no post-quantum algorithm present. A scanner matching algorithm names cannot see that by construction, and would report a protected deployment as classical_only. PPK is matched by the directives that switch it on and classified as quantum_resistant_mechanism — deliberately not pqc_ready, because a preshared key is not ML-KEM. Whether it holds depends on the entropy of the key and on out-of-band distribution, neither of which is visible in a file, and the finding says so.

CI/CD pipelines — signing commands such as gpg --sign, cosign sign, signtool, jarsigner, codesign.

Infrastructure as code — Terraform and Kubernetes key algorithms, and private key material committed by mistake.

Build files, includes, test recipes and key inventories — Makefile, CMakeLists.txt, .in, .cmake (a build file enumerates which algorithms a tree implements at all), .inc (the assembly and C fragments of implementations — OpenSSL keeps its ML-DSA there), .t (Perl test recipes: tests are code, and skipping them quietly is exactly what this tool argues against), and known_hosts / authorized_keys, where the key type is named on every line.

Not read, on purpose — documentation: .txt, .md, .pod, .rst. Measured across five real repositories, reading them would have added more than 11,000 "findings" from help texts and changelogs. An algorithm mentioned in prose is not a deployment. The boundary is pinned by a test, so it is a declared scope rather than a silent skip — the same standard this tool asks of others. Binaries and images are not read either.

Accidents of the alphabet are not findings — a match inside a long hexadecimal or base64 run does not count. A NIST test vector in OpenSSH spells ed448 inside its message bytes.

Real run against OpenZeppelin's contracts (711 files, about five seconds):

{
  "detected_algorithms": ["ECDSA", "RSA"],
  "summary": {
    "quantum_vulnerable_count": 2,
    "pqc_ready_count": 0,
    "highest_severity": "high",
    "pqc_readiness": "classical_only"
  }
}

What the coverage block says

Every scan carries one, because a coverage figure without its conditions is not comparable to another coverage figure. Four things, each answering something the percentage cannot:

• instrument — the version and the emitter's own commit. Two runs of this package once reported the same version from code that differed by a commit, so the version alone does not identify what did the reading. Where the tool runs from an installed wheel there is no commit, and the field says which absence rather than going quiet.
• corpus — the commit that was read, whether the tree was dirty, and whether the clone was shallow. A commit identifies a tracked tree; a scanner walks a filesystem, and the two are not the same thing.
• window — when it was read.
• scope — the denominator, the numerator, and every file that was in the first and not the second, with a reason. The reasons are a closed set: type_not_claimed is a boundary this tool declares, unreadable is a failure it hit, and they are never collapsed. A reason with no instances is reported at zero rather than omitted.

Measured across five real repositories (certbot, OpenSSH, Vault, Bitcoin, OpenSSL) at this release, the scan reads 67% of the files present — 74% of OpenSSL, 85% of OpenSSH, 66% of certbot, 77% of Bitcoin, 58% of Vault. The rest is counted and named with a reason. A directory that cannot be entered or listed is reported in unreadable_directories; its files cannot be counted, so the scan then says it cannot account for every file instead of claiming it read them all.

It also states which kind of claim the numbers are. Coverage is a claim about reading, not about finding: a file can be opened, counted, and still be one this tool was blind in. Reaching a file is something a scanner can measure about itself; whether it found what was there is not, because a silent rule and an absent algorithm produce the same output. So the block reports claims.axis: reached, and declares that it holds no control — the corpus with independently established contents that would license the second claim — naming the absence rather than implying the stronger reading.

What the number is, and what it is not

A coverage figure here is the overlap between the file types this tool claims and what the corpus is made of. It is not a discovery rate. A scanner claiming 32 extensions cannot reach 100 % against a tree holding 79 kinds, so a lower number means more file types left unclaimed rather than more cryptography left unfound — and the two read identically unless the page says which it is.

The denominator stays conservative anyway: you cannot know a .txt holds no PEM block without opening it, and key files often carry no extension at all.

Concentration, and the partition it is measured over

The block reports where the mass sits, not only how large a total is — because an aggregate over a lopsided population describes its largest members and reads as describing all of them. Measured across five open-source repositories: three file kinds account for 52 % to 89 % of everything not read, and the two largest kinds present are 43 % to 55 % of everything counted.

Three values travel with every share, and the third is the one most tools omit:

• share — how much the largest members are.
• cardinality — how many kinds that share is out of. Three of 24 kinds at 69 % is five and a half times a flat split; three of 63 at 68 % is fourteen times one. Without it the same share means different things and cannot be read as high or low at all.
• partition — what a kind is. A concentration is not a property of an aggregate but of the aggregate crossed with the partition it was measured over: the same tree split by extension, by directory, or by language gives different shares with nothing changing on disk. This block partitions by file extension and says so beside every figure.

The signer's sentence carries the same three inside the bracket holding the count, because a figure lifted out of a document without them is the failure the block exists to prevent:

This scan read 822 of 1250 files (65.76%); the remaining 428 (3 of 24, by extension —
files with no extension, .rst and .txt — being 68.93% of them) are listed with a reason each.

The published measurement, with the raw artefacts: quantumreadiness.eu/evidence/scan-coverage

The CBOM it emits

export_cbom produces a CycloneDX 1.6 document, validated against the published schema; from 0.20.0 qrp-mcp cbom PATH --out FILE writes the same document to a file. Three things travel in it that a component list alone cannot say:

• compositions.aggregate — incomplete where files were not examined, and unknown where the tool cannot account for its own reading. Reading every file is not enough for complete: that word claims every asset present was found, which is the second axis, and only a held control licenses it. The document used to say complete on the strength of the denominator alone; comparing against other scanners showed findings missed inside files that had been read.
• properties — the whole coverage block. It travels there because the root object is additionalProperties: false and the format has no field for it; the awkwardness is the point rather than something to hide.
• evidence.occurrences — file, line and matched text for every asset.

Output is deterministic where it matters. The serial number is derived from the target, the two pins and a digest of what was found, so the same code over the same corpus that finds the same things gets the same serial, and a different result gets a different one. The timestamp and the coverage window record when the run happened, so those fields differ between runs.

What kind of place a finding sits in

Two things about a match are not the match itself, and both now travel with it — in the scan result and in the exported CBOM.

Evidence kind. scan_repo grades every match: call, declaration, import, reference, ban, comment. Comment evidence is kept out of the inventory, and the reason is measured — the nearest rival strips comments before matching and scored 0.542 precision on an independent corpus against this scanner's 0.93. The exported document used to drop that grade, so a sentence about certificates reached an auditor looking exactly like a signature: 78 of 465 occurrences on certbot (17%) and 1174 of 7142 on OpenSSH (16%), including 493 ML-DSA and 429 ML-KEM mentions in OpenSSH, which is a lot of prose to present as post-quantum adoption. An occurrence whose evidence is a comment now says [comment].

⚠️ A banned algorithm is not marked, it is absent: !RC4 in a cipher list never becomes a component. What the line forbids is not inventory; what it enables is.

Test code. A finding says whether it sits in test code, and the scan declares the rule it used:

"test_code": {
  "rule": "a path component named test/tests/testing/spec/specs/fixtures/testdata, or a file named test_*, conftest, *_test, *.test.*, *Test, *Tests, *_spec, *.spec",
  "meaning": "whether the finding sits in test code; named, not excluded",
  "findings_in_test_code": 144,
  "findings_in_other_code": 301
}

Components whose every sighting is test code also carry the standard field — component.scope: "excluded", which the schema has defined since 1.6 as documenting "component usage for test and other non-runtime purposes". ⚠️ v0.18.0 shipped the per-occurrence marker and no scope at all, which was a private spelling of a field the standard already had.

⚠️ And the honest measurement of that field: on certbot, not one of the 31 components is confined to test code — every family that appears in a fixture also runs in production. At component granularity excluded almost never fires on a real repository, which is exactly why the per-occurrence marker is not redundant: scope is a property of the component, [test] is a property of one sighting, and 144 of 445 findings need the second.

Those are certbot's real numbers: almost a third of its findings are in fixtures. Nothing is dropped — a fixture's RSA key is real RSA. Whether it belongs in a particular migration plan is the reader's call, and they can only make it if the document says which is which. Occurrences in test code carry [test], and a comment inside a fixture carries both.

⚠️ There is no external convention for this, which is why it is declared rather than decided. The two public corpora that could settle it — qscan's recall benchmark and the cryben corpus of Näther & Hirsch — are 100% synthetic fixtures with no test/production distinction. The rule here is a path convention, not a fact, so it is printed with the numbers it produced; a reader who disagrees with the rule can see exactly what it caught.

Whether a fix closed it

A finding missing from the second scan has not necessarily been fixed. It may have moved down the file, moved to another file, landed in a directory the second run could not open, or stopped matching because the scanner changed. Each of those looks exactly like a fix when the only question is "is it still in the list".

prove_closure(before, after) takes two saved results (qrp-mcp scan PATH --out FILE) and answers in two steps:

1. Can these two runs be compared for closure? Same instrument commit, version, rules, file types and exclusions; both trees pinned to a commit and clean; both results quoting lines at the same level. Run from an installed release (uvx qrp-mcp, pip install) there is no commit to name; two runs of the same installed release are accepted on its version, and the result says so in basis — weaker than a commit, since a locally modified install would carry the same version string. The trees may differ — that is what is measured. If the runs cannot be compared, nothing is reported as closed, the reasons are listed, and the repair is named: scan the first commit again with the instrument that did the second.
2. Only then, occurrence by occurrence, without line numbers: closed, still open (and how many only moved lines), relocated to another path (a rename is not a fix), moved into or out of test code (neither is a fix), new, unverifiable — in a file the second run did not read — and removed: the whole file is gone from the second tree. A deleted file does remove its code, but it is also the cheapest way to make a finding disappear, so it is named apart and not counted as closed. A file the second run could not open is never counted as fixed.

From 0.20.0 every scan lists files_read by name, beside the counts. That is what tells a fixed line from a deleted file; it stays out of the CBOM. Two results without it (0.19.0) are reported as unestablished, not compared on a weaker rule.

Checked on a copy of certbot (521 occurrences): removing one RSA call closed exactly that one; inserting blank lines closed none and reported two moved; renaming a file closed none and reported two relocated; a masked scan against a full one was refused. Two scans of OpenSSH with the same instrument: 7,143 still open, 0 closed. A second run with an emitter that had uncommitted changes was refused as instrument_dirty. With 0.20.0 installed in a clean environment, deleting one RSA line in client.py and the whole of challenges.py gave 1 closed, 1 removed, each on the right file and line.

⚠️ What is not claimed: that the fix is correct, only that the occurrence is gone from every file the second run read. Other scanners close findings between scans too — Semgrep, SonarQube, GitHub code scanning, and Keyfactor AgileSec ("Deleted findings are automatically resolved", 3.4 release notes). Semgrep marks a finding "Removed" rather than "Fixed" when its rule changed or its path went away, and GitHub records the tool version to track changes it caused; both do this as a status after the fact. What was not found in the sources read is a refusal, before calling anything fixed, when the reader changed between the two scans. SonarQube documents that a file dropped from scope is counted as fixed. They keep a history across many scans; this compares two documents. A statement about the sources read on 2026-09-27, not a survey of every tool.

What to replace it with

Each finding that is quantum-vulnerable or deprecated carries a replacement: what to use, for which role, and the standard that says so, with its address. It is a suggestion; nothing is applied, nothing leaves the machine, no model is asked.

Transition dates come from NIST IR 8547, which is still an initial public draft (12 Nov 2024): quantum-vulnerable signatures and key establishment deprecated after 2030 at 112-bit strength, disallowed after 2035. Hybrids are named as NIST names them — accommodated, and temporary. The CBOM is not changed: CycloneDX has no standard field for a recommendation, and this project does not add a private one.

The role of an RSA or EC line

From 0.21.0 every RSA and EC occurrence carries a role, read from its own line: signature (RS256, rsa-sha2-256, SignPSS, sha256WithRSAEncryption, ECDHE-RSA-… in a cipher list), key_establishment (OAEP, EncryptPKCS1v15, TLS_RSA_WITH_…, a kex table) or undetermined. A key generated or declared on a line is used somewhere else, and this reads one line, so such a line stays undetermined; so does a line with signals for both, and a cipher list longer than the 200-character excerpt. The finding's replacement counts the roles of its lines of code (comments carry a role but are not counted) in roles_seen, and drops a path only when no line showed that role and no line was undetermined. Every path says how many lines it is for.

⚠️ Against this: on every corpus measured, both paths stayed. Undetermined lines of code: certbot 208 of 229, OpenSSH 172 of 456, Vault 361 of 656, OpenSSL 3,354 of 3,893. The count is the gain, not a shorter list.

Role-assignment precision on a sample (not detection; judged from the source ±5 lines, not from the excerpt the rule saw; rules frozen at 3287c1c before these corpora were opened; seed 20260927):

The nine wrong key_establishment lines have two causes, both in TLS names: TLS_ECDHE_RSA_WITH_… written with underscores is read as ECDH, and in ECDH_RSA suites RSA only signs the certificate of a fixed ECDH key. Both are known and not fixed in 0.21.0; the fix will be measured on a corpus these rules were not written against.

Measured against the other scanners

In September 2026 three free tools that do the same job — CryptoScan, CBOMkit-hyperion (sonar-cryptography) and CBOMkit-theia — were run over the same repositories and the findings compared line by line. What they found and this tool did not became the 0.8.0, 0.8.1 and 0.9.0 releases. Of what this tool does that they did not, one claim needed narrowing when a wider survey was done, and it is corrected here:

• Coverage inside the document. Several tools do report what they skipped: QuantaKrypto's qscan counts scanned and unread files, IBM Quantum Safe Explorer logs each excluded file with a reason, SandboxAQ shows missed locations. What we have not found elsewhere is the coverage travelling inside the CycloneDX CBOM — a reason per group of unread files, the paths that could not be read, the directories that could not be entered, and accounts_for_every_file. qscan computes the counts and drops them on export, so it is one flag away from the same thing.
• This scanner separates reading from finding in the document itself, and refuses to say complete without a control that licenses the stronger claim. Reading is self-measurable; finding is not.
• sntrup761, the default hybrid in OpenSSH from 9.0 to 9.9 — OpenSSH 10 defaults to mlkem768x25519-sha256 — has no rule in CryptoScan; this tool reports 145 lines of it in the OpenSSH tree.
• Private keys are recognised by content, not by file name: the 45 key files CBOMkit-theia found in OpenSSH and this tool did not are read from 0.9.0, and a PEM header with the body elided — documentation — is not one.
• An excluded cipher (!MD5) is counted as a use by CryptoScan; here it is an exclusion.
• On certbot, this scanner finds algorithms in 26 files against hyperion's 7, and hyperion's one extra finding is wrong (RSA-96 where certbot defaults to 2048).

Measured on a corpus this project did not write

The nearest tool of the same kind is QuantaKrypto's qscan — lexical like this one by its own changelog, and the only other tool in this class that publishes a detection figure. In September 2026 both were run against Cryben (Näther & Hirsch, arXiv 2608.04857): an independent corpus with its own reference CBOM and its own scorer, written by neither of us. The scripts and the raw output are reproducible; the method matters more than the number.

⚠️ Every figure in that table was measured on 0.11.0 and has not been re-measured since. 0.12.0 changes what counts as a use: an external retest found that a denial word anywhere on a line -- including inside a comment, including the word weak in a variable name -- turned a real key into a ban and dropped it from the inventory. Fixing that necessarily moves both recall and precision, in directions this table cannot state until the corpora are run again. The numbers below are the previous release's, labelled as such rather than quietly carried forward.

Three denominators appear in that table and they are not the same question. 37 is the number of Cryben cases inside this tool's declared scope; 176 is the label count in qscan's recall corpus; 197 is Cryben's full finding set. A recall figure over one of them cannot be read against an F1 over another, and earlier versions of this section invited exactly that. Each row belongs to one task and one denominator; none of them combine.

200 clean is not a precision figure. Its denominator is negative files, not emitted findings. It says the scanner stayed silent on 200 files that contain nothing; it says nothing about how many of the findings it does emit are right. Precision is the row two lines above it, and it is the row this tool loses.

The independent control is not published. The Cryben corpus, its scorer and the raw runs behind 36/37, 0.667 and 0.346 are not in this repository. Until they are, those three are this author's claims rather than something a reader can check, and an external comparison said so in those words. Publishing the artefacts is the repair; restating the numbers is not.

Four things those numbers do not mean, said here rather than left to be assumed:

• Cryben is 37 cases in this tool's scope. A figure from 37 cases has a wide interval. It is a floor worth publishing, not a precision claim.
• The scorer is theirs, and it credits a finding that names nothing. qscan's metric matches per file, not per line, and 10 of its hits name no algorithm at all. A tool that reports "something cryptographic is here" scores the same on those as one that says which family it is, so detection and family classification have to be published apart. They are not, here, yet. An earlier version of this section claimed the figure was unchanged with that credit withdrawn. It was wrong: the switch that withdrew it dropped only one bucket, while elliptic-curve findings kept the credit unconditionally. Withdrawn properly, and applied to both tools, the numbers are 0.818 for this tool and 0.790 for qscan. The conclusion survived the correction; the sentence did not.
• A per-scan control is still not held. claims.control.held stays false in your scan unless you run one, and it should: a figure measured here says nothing about your repository.
• Recall is not the axis this tool is weakest on — precision is. On Cryben's full 197 findings, scored by its authors' own tool, this scanner's precision is 0.667 against qscan's 0.93. It was 0.542 one release ago. Reaching a rival's recall while reporting more that the reference does not is not the same as being the better inventory, and it would be dishonest to publish the first number without the second.

Three levels of evidence, because a quoted line of code is not always safe to send. --level full carries the line itself. --level masked keeps its shape with everything but the algorithm name starred out — *** = rsa.********_*******_***(***_****=****) — so a reader sees a call rather than a string without seeing the contents. --level trimmed carries no line at all. The file and the line number stay in all three.

The masking rule is inverted from the obvious one. A rival tool masks by position, keeping the first characters and starring the rest, which keeps whatever the line happens to begin with — a token, a key, a password. Here only the characters that spell the algorithm survive; every other letter and digit becomes a star, and punctuation stays because the shape of a call is not a secret. An independent analysis of this scanner found a synthetic token sitting on the same line as a finding, which is what prompted the level.

From 0.16.0 the tools mask by default. The level existed on qrp-mcp scan --out -- the path that writes a file for you to read before you send it -- and nowhere else. The scan_repo and export_cbom tools returned every matched line as written, and those are the path that runs on every agent call and hands its result to a model. export_cbom says in its own description that it is for "when the result has to leave the machine", and it carried each line verbatim into evidence.occurrences. The control had been built for the path we thought left the machine rather than the one that leaves on every call. Both tools now take the same level and default to masked; level="full" returns the lines. Reading is unchanged -- masking decides how a finding is quoted, never whether it is found, and no detection figure in this file moves because of it.

qscan is faster: 7.9 s against 219 s on Vault, a factor of 27.7, and a factor of 10.6 on the median of five warm runs over a smaller corpus in an independent comparison. Measured here after 0.13.0: 26.1 s for 823 files of certbot, 31.7 ms per file, of which the bulk is 47 algorithm patterns run over every line -- about 4.5 million pattern searches on that tree. A prefilter would cut it and would also change what is detected, so it is a task of its own rather than a line in this one. The figure was written here as "about 24" and was arithmetic nobody had done: 219 / 7.9 is 27.7. It is not deeper: it is lexical, by its own changelog, as this section already says. Measured per language on its own corpus it leads in none by more than three labels and trails this tool by five in Go; its real edge is in TLS, SSH and dependency lines rather than in any language. This tool reads more kinds of file, says what it did not read, and does not invent a family for an asset that names none.

The corpus that measures the other direction

A corpus of labelled findings can only ever say what a tool misses. It says nothing about the twenty lines in the same file that are not cryptography. So this release is measured against one written for the opposite question:

• 200 files with no quantum-vulnerable cryptography at all — ordinary code in fifteen languages, configuration, CI, infrastructure, manifests, lockfiles, data and documentation. Any finding on one of them is a false positive.
• 100 near misses — cryptography named, banned, discussed, tested against or imitated, but not used: an SSLProtocol line removing SSLv3, a policy listing forbidden algorithms, a lint rule quoting the pattern it forbids, a test asserting an algorithm is rejected, a docstring explaining why RSA was dropped, base64 that is a JWT payload rather than a key.

Every file was written and labelled before the scanner was run over it, and the corpus was written by someone who had not read the scanner's rules. A corpus assembled by looking at what a tool reported measures the tool against itself, which is the circularity this document criticises elsewhere.

Result: 200 of the 200 negative files are clean — no finding of any kind. On the 100 near misses, this release reports 83 findings the labels say should not read as a use, down from 136. What remains is one shape: a file whose whole purpose is to forbid. A lint rule quoting hashlib.md5(...) as the thing it bans is, line by line, indistinguishable from code calling it; telling them apart needs file-level context this release does not attempt.

The corpus is a floor, not a measurement. It was written by the same hands that wrote the tool, so a construct nobody here thought of is in neither. Its honest use is as a difference between two releases.

What is still missing here

Stated rather than hidden, and each of these is a known gap rather than a suspicion:

• A finding carries a kind, not a confidence level. Since 0.11.0 every finding says whether it is a call, an import, a comment, a declaration, a reference or a ban, and 0.12.0 decides that by the position of the match rather than by the line it sits on. What is still missing is a graded confidence. Two contexts that were read as code and are not -- a Python docstring, and a string constant holding a PEM header -- were reported by an independent comparison and are fixed in 0.13.0, with the controls that keep the fix narrow: a key pasted into a triple-quoted value is still key material, and a cipher suite named in a string is still configuration.
• SSH and TLS assets are read at 8/14 and 7/11 on qscan's corpus; X448 at 4/8.
• Speed. Reading key material by content costs about 26% over 0.8.1 on a large tree.
• A symlink is never read. Where a tree reaches content only through a link whose target sits in a directory this tool excludes, that content is not scanned. It is named as a link rather than silently skipped, but it is not read.
• Symmetric cryptography, hashes for integrity, KDFs and random number generation are out of scope by design. This tool reports what a cryptographically relevant quantum computer would break. On Cryben's full 197 findings — most of which are AES, SHA-256 and KDF — it scores 0.13, and that is the scope working, not failing.

What an outside review found

0.9.0 was reviewed by someone who did not write it, from the published ZIP, and reported nine defects. All nine reproduced here before anything was changed, and all nine are fixed in 0.10.0. Two of them were about the claims this tool makes for itself, which is the worst place to be wrong:

• A symlink carried the scan outside the directory it was given. A link inside the tree pointing at a file beside it was read, and reported under the link's name. This tool is pointed at code its user did not write, and the excerpt travels into an agent's context and into any exported CBOM, so that was the declared boundary failing. Links are no longer followed. Each is named in symlinks_not_followed, with a relative path when the target is inside the root and the words "outside the scanned directory" otherwise -- printing an outside path would leak what reading it did. A linked directory sets accounts_for_every_file to false. On certbot, which uses 48 of them, coverage falls from 68.9% to 65.8%: the links are now counted as present and not read. No algorithm is lost, because each target is still read at its own path.
• compare_coverage keyed on a commit, and a commit is not what was read. A dirty emitter still compared; an unverified corpus compared because None read as False; two different subdirectories of one commit compared at 100% and 0% coverage; a git-ignored file that the scan reads changed the corpus while both pins stayed clean. Comparability now keys on corpus.content_digest -- a hash over the files read and the text read from them, the files skipped with their sizes, and every entry not read with its reason. The git pin stays as provenance a reader can follow; it no longer carries the conclusion. dirty is three answers, and None means nobody checked.

The rest: a malformed coverage block returned an exception instead of unestablished; certificate evidence came off a set, so its order moved between interpreters; --out ~/file passed its check and failed its write; and three tests read the tool's own git pins from the environment, so they passed from a clone and failed from the published tarball -- which anyone who downloaded 0.9.0 and ran its tests saw, and we had not, because we always ran in the checkout. Extracting the release and running its tests in a fresh environment is now part of shipping one.

Why deterministic

There is no LLM inside this tool. The same input always produces the same output, and every finding points at a file and a line you can open yourself.

That is the point of handing it to an agent: the agent brings the language, the tool brings the truth. An agent guessing about your signing code is worse than nothing; an agent reading a deterministic inventory can actually reason about it.

What it is not

It reads source, configuration, CI pipelines, infrastructure-as-code, Kubernetes manifests, build files, certificates and key inventories. It does not read documentation, binaries or images.

Every file under the path is accounted for in one of three ways: scanned, unreadable, or skipped because the tool does not claim that type — the last counted by extension, so the coverage figure has a base. files_scanned + unreadable_files + files_skipped_by_type always equals files_present. A directory the scan cannot enter or list is named in unreadable_directories; its files cannot be counted, so the scan then says it cannot account for every file instead of claiming it read them all. A scan that read seven files out of nine is a different report from one that read seven out of four hundred, and only one of them is worth trusting.

A free inventory tool, not a readiness assessment. It deliberately does not do:

• risk scoring or prioritisation,
• migration planning — a replacement names what a finding becomes, not when or in what order,
• network or host scanning, or reading a system certificate store,
• tracking change over time — prove_closure compares two scans you give it; nothing keeps a history.

Those live in the Quantum Readiness Platform, the product this tool is extracted from. Nothing here is crippled to push you there — what it does, it does completely.

It also does not tell you that you are about to be hacked. It tells you what you are using.

License

Apache-2.0.