Read this first
Your agent will write an ECDH handshake without blinking — and never mention that it is on a deprecation clock, that recorded traffic is already being harvested against the day it breaks, or which NIST standard replaces it. The quantakrypto MCP closes that gap: the agent can inventory the cryptography in your code, flag what a quantum computer breaks, and draft hybrid replacements, inside the editor. Below is the loop we run, and the line between what to delegate and what to decide.
There is a specific blind spot in every AI coding agent: it is fluent in cryptography and unaware of its post-quantum status. It knows the APIs cold and nothing about the calendar behind them — NIST deprecates the classical public-key algorithms after 2030 and disallows them after 2035, and harvest-now-decrypt-later means the exposure is already present tense. The quantakrypto MCP hands the agent the missing sense as tools it can call, so the model reasons from your detectors instead of guessing.
Two ways to run it — and when to use each
The transport you pick changes what the agent can see, so choose deliberately:
- Local (stdio) — `npx @quantakrypto/mcp` runs on your machine and reads your workspace. Use it for whole-repository inventory and anything that must not leave your machine.
- Hosted (remote) — the shared endpoint at mcp.quantakrypto.com is reached over an OAuth sign-in with a 30-day token. Zero install, easy to share across a team, and it only ever sees the content you pass a tool — no filesystem access.
The loop
Post-quantum migration is not a library swap; it is a sequenced programme, and the agent is good at the parts that are tedious rather than hard. The loop we run:
- Inventory first. Ask the agent to map every algorithm, key, certificate, and protocol — and what data each protects. You cannot prioritise what you cannot see; this is the cryptographic inventory, and it should live in CI, not a spreadsheet.
- Rank by exposure. Long-lived secrets migrate first, even when they are the hardest — that is the harvest-now-decrypt-later calculus. The agent can order findings by that exposure.
- Draft the hybrid. For key exchange, the move is X25519MLKEM768 — a hybrid that stays secure if either half holds and remains interoperable with classical clients. Ask the agent to propose it in place.
- Verify the change. After an edit, have the agent re-run the detectors over the snippet to confirm no classical asymmetric cryptography remains — a removed vulnerability is a claim worth checking, not assuming.
Delegate the grind; keep the decisions
The agent excels at finding every place an RSA key or an elliptic curve hides, drafting a hybrid replacement, and confirming a fix landed. The sequencing — what to break, when, with which rollback — is still an engineering judgement, and so is the risk trade-off on a system you cannot take down. Treat tool output as evidence, not a verdict, and keep a human on the load-bearing calls.
Guardrails
- Do not paste secrets. The analysis tools want the shape of the code, not your credentials — send the function, not the environment file.
- Conformance is a separate axis from migration. Moving to ML-KEM and shipping a conformant ML-KEM are different claims; passing NIST's test vectors does not prove the input-validation behaviour the standard requires. Conformance-test the implementation you ship.
- Prefer hybrids, stage the rollout, keep a fallback. Backward compatibility is what lets you migrate without an outage.
Everything the agent calls is open source and auditable — the tools in pqc-tools and the hosted gateway in mcp-gateway. Point your agent at the MCP, start with an inventory, and let the tooling you already use begin noticing what your cryptography is made of.