pqc-migration-mcp

Give your AI agent the post-quantum migration facts it keeps guessing at.

Six tools over MCP: credential sizes, fragment counts, the reassembly window, the 39-family failure taxonomy, and benchmark scoring. Ask Claude "will our ML-KEM-768 handshake fit in a BLE MTU?" and it computes the answer instead of estimating one.

📖 Full documentation, tutorial and conceptual guide: https://nickharris808.github.io/pqc-toolkit/

---

Why this exists

Agents are increasingly doing PQC migration work, and they are confidently wrong about exactly the things that matter: how big a credential actually is, how many fragments it becomes, and whether a safe reassembly cap exists at your concurrency. Those are arithmetic, not judgement — so hand the agent the arithmetic.

The protocol layer here is dependency-free. MCP is JSON-RPC 2.0 over line-delimited stdio, which is small enough to implement directly and keeps the install trivial.

Install

pip install git+https://github.com/nickharris808/pqc-migration-mcp

This pulls pqc-sizes and pqc-mfb from their repositories too. Not on PyPI yet, so pip install pqc-migration-mcp does not work today.

30-second quickstart

# talk to it directly -- it is line-delimited JSON-RPC on stdio
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}' | pqc-migration-mcp

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "pqc-migration": {
      "command": "pqc-migration-mcp"
    }
  }
}

Restart Claude Desktop. The six tools appear under the connector.

Tools

Tool	Answers
credential_size	How many bytes is a KEM+signature credential, component by component?
fragments	How many fragments on this transport — and is fragmentation now mandatory?
reassembly_window	Does a safe capacity cap exist at all? If not, what concurrency would work?
list_failure_families	All 39 failure families, with case counts and published analogues
describe_family	What breaks in this family, in which designs, and what did each do?
score_submission	Score a PQC-MFB submission: coverage, regressions, zero-coverage families

Worked example — actual output

The transport is line-delimited JSON — one complete object per line. Keep the request on a single line; a request wrapped across two lines arrives as two incomplete ones and comes back as two -32700 parse errors.

$ echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"reassembly_window","arguments":{"largest_legitimate_object":12000,"memory_budget":32768,"concurrency":3}}}' | pqc-migration-mcp

The server replies with one JSON object per line. Pretty-printed, the content payload of that reply is:

{
  "budget": 32768,
  "ceiling": 10922,
  "concurrency": 3,
  "explanation": "EMPTY WINDOW: floor 12,000 B > ceiling 10,922 B (short by 1,078 B). No capacity cap is both feasible and safe. Raise the budget to at least 36,000 B, reduce concurrency to at most 2, or choose a smaller credential.",
  "floor": 12000,
  "is_empty": true,
  "max_safe_concurrency": 2,
  "recommended_cap": null
}

The agent gets a verdict and the number that would fix it, so it can propose a concrete change rather than reporting a problem.

What this server will not tell you

It exposes detection. It does not expose repairs.

An agent can learn that a design fails krack_retransmission and exactly what the unrepaired design did. It cannot obtain the mechanism that closes it. That boundary is deliberate: an MCP tool returning repairs would let any user enumerate the entire closed set in an afternoon.

There is a test that calls describe_family for all 39 families plus every other tool, concatenates the responses, and fails if repair_mechanism, repaired_detail or repaired_held appears anywhere in the output.

Error semantics

Domain errors — an unknown algorithm, an unknown family — come back as a tool result with isError: true and a message naming the valid options, so the agent can correct itself. Only protocol faults become JSON-RPC errors (-32601 unknown method/tool, -32602 bad arguments, -32700 unparseable line).

A malformed line does not kill the loop; the server replies with a parse error and keeps serving.

Tests

pip install -e ".[dev]" && pytest      # 57 passed

Tests cover the protocol, every tool, the moat boundary, and the real stdio transport driven as a subprocess — including a check that stderr stays empty, since MCP clients read stdout as protocol and stray warnings confuse them.

Scope

Arithmetic, taxonomy lookup and scoring. No cryptography, no network, no telemetry. It does not inspect your implementation. A clean answer means your configuration is sound, not that your code enforces it.

Related

pqc-sizes · pqc-mfb · pqc-guard-action · pqc-dos-embedded

Closing the 39 families is what the closed core does. Relevant subject matter is covered by a filed provisional patent application. For commercial use of the full envelope, open a GitHub Discussion or an issue on this repository.

Honest scope

What this proves. That the arithmetic and taxonomy an agent is reasoning with are correct: real credential sizes, real fragment counts, a real window verdict, and the real failure taxonomy.

What it does NOT prove.

• Not that the agent used the answer. This supplies facts; it does not supervise what is done with them.
• Not an inspection of your code. No tool here reads your implementation.
• Not a repair channel. Every tool exposes detection only. A test calls describe_family for all 39 families plus every other tool and fails if a repair field appears anywhere in the output.

Errors. Domain problems come back as tool results with isError: true and a message naming valid options, so an agent can self-correct. Only protocol faults become JSON-RPC errors.

---

The PQC migration toolkit

Eleven free tools for teams moving authenticated key exchange to post-quantum. They find and measure; they do not repair.

Tool	What it does	Where
pqc-sizes	Sizes, fragment counts, and the two-sided reassembly window	PyPI
pqc-sizes-js	The same arithmetic for Node and the browser	npm
pqc-guard-action	Fail the build when the window is empty	GitHub Action
pqc-dos-embedded	169 lines of C: the failure on a real 64 KB device	source
farkas-check	Re-verify the bound on-device, no SMT solver	source
pqc-migration-mcp ← you are here	Six MCP tools for AI agents	PyPI
pqc-mfb	322 cases · 39 failure families · scorer	PyPI
pqc-mfb (data)	The benchmark as a dataset	HF
pqc-formal-corpus	122 named formal results, 6 provers	HF
pqc-bounds-lean	The same bound in Lean 4 — 0 sorry, 0 imports	source
pqc-dos-gate-rtl	The gate in synthesizable RTL, 5 Yosys proofs	source
pqc-explorer	Try it in your browser, no install	HF Space

New here? The end-to-end tutorial walks one realistic migration through all of them in about ten minutes: sizes -> window -> CI gate -> benchmark.

In a hurry? pqc-sizes tells you in five seconds whether your credential fragments and whether a safe cap exists. pqc-explorer does the same in a browser, with no install.

The closed core

Closing the 39 failure families — downgrade binding, retransmission-safe installation, fragmentation transcripts, roaming forward secrecy, multi-link key separation, admission control, group-key binding — is a separate proprietary codebase. Relevant subject matter is covered by a filed provisional patent application.

That split is measured, not asserted: under a replicate noise control only 4 of 32 repair mechanisms are externally distinguishable, so publishing these detectors does not disclose the repairs.

For commercial licensing, open a GitHub Discussion or an issue on any of these repos.

License

Apache-2.0. See LICENSE and CONTRIBUTING.md.