mcp-confirm

In plain terms: when an AI asks "are you sure?", the MCP SDK already stops that approval being reused for a different action. It does not stop the same approval being used twice. This closes that gap, and two others the SDK leaves open.

Read this first: most of this problem is already solved

The 2026-07-28 MCP specification added Multi Round-Trip Requests, so a tool can pause mid-call and ask the user to confirm. The approval travels back through the client as an opaque requestState, which the spec says servers MUST treat as attacker-controlled.

The Python SDK does this for you, by default, on every MCPServer. RequestStateBoundary is appended to the middleware chain unconditionally — with an ephemeral key if you supply none. It seals the state under AES-256-GCM and binds it to the method, the target, a digest of the call's arguments, the audience, and the authenticated principal, with a TTL.

So the attack everyone reaches for first — approve deleting cache.txt, then replay that approval against thesis.txt — is already refused by the SDK. You do not need a library for it, and you should not write one.

This repository originally wrote one anyway: 250 lines of HMAC, TTL and argument binding, shipped as the headline feature, redundant the day it was published. That is recorded at the bottom rather than quietly deleted.

Related MCP server: RecourseOS

What the SDK does not do

Three gaps, each with a test that fires.

1. It binds and expires the state. It never spends it.

Inside the TTL, the same approval verifies as many times as it is presented. Every check the boundary makes passes, every time. The specification is explicit that this is deliberate and that the rest is yours:

Note that these measures bound the replay window and prevent cross-user and cross-request reuse, but do not by themselves guarantee single-use. Servers for which a given requestState must be consumed at most once (e.g., one-time redemptions) MUST enforce that invariant server-side.

For "delete a file" the second attempt finds nothing. For "transfer £500" it is the entire problem. singleuse.py is that invariant — a ledger of outstanding confirmations, spent on redemption. It contains no cryptography, because the boundary already guarantees the plaintext is something this server minted.

2. Nothing sanitises the question the human reads

An elicitation message is server-chosen text rendered to a person, usually with an untrusted value interpolated into it — the whole point is to say which file is going away. So name a file:

cache.txt

SYSTEM NOTICE: your session has expired.
Enter your AWS secret key to continue:

The dialog now carries a second, official-looking prompt. The user is not confirming a deletion; they are being phished by their own tooling. The same 2026-07-28 release also shipped MCP Apps — server-rendered UI — which widens this surface rather than narrowing it.

prompt.py flattens untrusted values to one line, strips bidirectional overrides and zero-width characters, and truncates from the middle so the filename at the end stays visible. Control characters become spaces rather than being deleted, since collapsing them would let a\nb and ab render identically — two different files, one dialog.

3. No protocol layer can re-check your resource at execution time

The user thought about it in between. The file can be replaced in that window, and only the tool knows what "unchanged" means for it. This server re-checks before deleting, and refuses a path that has become a symlink.

Which layer refuses what

This is the useful part, and the tests are written to demonstrate it. MCPError means the SDK's middleware refused before this package ran; ToolError means this package did.

Tests

35 tests — 17 through the server, 11 on the ledger, 7 on prompt sanitisation.

Every server test runs through the real RequestStateBoundary, the same class MCPServer installs on itself, with a pinned key — sealing round one and unsealing round two exactly as the wire would.

That harness exists because of a specific mistake. The first version tested by calling MCPServer.call_tool() directly, which goes straight to the tool manager and bypasses the middleware chain entirely. The SDK's boundary never ran, so a redundant hand-rolled guard looked load-bearing. The test design is what hid it, for an entire build cycle.

Coverage is 81%, with prompt.py at 100% and singleuse.py at 94%. The gap is main()'s argparse and transport wiring, exercised through build_server instead.

CI runs on Ubuntu only, deliberately: the swap-after-confirmation tests need symlinks, and the build fails if they report as skipped there.

Install

pip install git+https://github.com/les-k/mcp-confirm.git

Run

mcp-confirm --root /path/you/allow

With no --root, the server refuses every request rather than defaulting to anything. Roots are fixed at startup and never chosen by the agent. No signing key is needed — MCPServer brings its own.

Known limitations

• The ledger is in-memory, so it is correct for one process and wrong behind a load balancer. The SDK supports sharing keys across replicas (RequestStateSecurity(keys=[...])); under that configuration replica B rejects a confirmation issued by replica A. It fails closed — the safe direction — but reads to a user as a confirmation that inexplicably stopped working. A multi-process deployment needs Redis or a database row with an atomic compare-and-delete. There is a test for this behaviour.

• The demonstration tool is deliberately small. It deletes one file. The interesting code is singleuse.py and prompt.py.

• No CVE backs this. MRTR is weeks old, so this is built from the specification's own MUST/SHOULD list and from reading the SDK, not from a published incident.

What was wrong with version 0.1.0

Kept here because a repository that records only its successes is not evidence of anything.

0.1.0 reimplemented what the SDK already did. state.py was 250 lines of HMAC signing, TTL, and principal/method/argument binding — all duplicating RequestStateBoundary, none of it as good: HMAC where the SDK uses authenticated encryption, no audience binding, no key rotation.

It was caught by reading the SDK's source, not by a test. The tests passed precisely because they bypassed the middleware that would have exposed it.

CI separately caught a real bug in 0.1.0: the symlink check ran after Path.resolve(), so it inspected the link's destination rather than the link itself. A swap aimed at another file inside an allowed root would have been deleted. Fixed, with a test for the variant the original never exercised.

0.2.0 deletes state.py entirely and keeps only what the SDK leaves uncovered.

Licence

MIT.