MCP server scan ยท 9 October 2026
0 of 266

tools that may change something, across 24 popular MCP servers, take an idempotency key. When a call times out, an agent can't retry any of them without risking doing it twice.

24servers scanned
489tools listed
266not marked read-only
0take an idempotency key
48named like reads, not marked read-only

Why the key matters

A tool call can time out, or the agent's process can die mid-call. Then nobody knows whether the issue was filed, the refund issued, the message sent. Retrying risks doing it twice; giving up risks never doing it.

Payment APIs settled this years ago with an idempotency key: the caller sends a key with the request, and the server answers a repeat with the first result instead of acting again. MCP's tools/call has no key field (spec issue #3394), so a server can only offer one as a tool argument. We checked how many do.

What a retry does, on real GitHub

The same create-issue call, sent twice to GitHub's official MCP server, as a client does when the first answer is lost. Then the same through agentsafe-mcp, a proxy that logs each call before forwarding it. Private scratch repository, github-mcp-server v1.14.0.

Straight to the serverduplicate
call 1: issues/6
call 2: issues/7
2issues created
Through agentsafe-mcpreplayed
call 1: issues/8
call 2: issues/8  replayed from the log
1issue created

Each of GitHub's write tools we called for real (create issue, comment, create branch, commit a file) made exactly one effect per call: the server doesn't duplicate anything itself. The duplicate comes from the retry, and GitHub's API has no key that could stop it.

Every server

Tools each server lists, and how many of them aren't marked read-only. The key column is the one that matters.

all toolsnot marked read-only
ServerVersionToolsToolsNot read-onlyWith a keyAnnotated

Three things the scan turned up

No keys, anywhere

Names were matched against idempotency_key, request_id, client_token and variants, then every schema was searched by hand for idempot, dedup, nonce, reference and receipt. The only hits were Razorpay's merchant references, which aren't documented as idempotency keys. Often the API underneath has none either (GitHub, Slack, Jira, Notion).

Hints that don't describe the tools

Razorpay marks all 41 of its tools destructive and not read-only, including 25 fetch_* reads. Nine servers send no annotations at all, so MCP's defaults call every tool a destructive write. A client that asks before destructive calls asks about everything.

One tool for the whole API

Square's server exposes everything through make_api_request(service, method, request). Listing customers and charging a card are the same tool, so a rule written per tool (an approval, a rate limit, a key) covers both or neither.

What this does and doesn't show

Run it yourself

Needs Go, Node and uv. run.sh downloads and runs each server's code on your machine and calls none of its tools.

git clone https://github.com/Ashutosh2308Bhardwaj/agentsafe
cd agentsafe/mcp/scan
./run.sh                 # inspects every server in servers.tsv, writes RESULTS.md
SANDBOX=you/scratch-repo AGENTSAFE_SANDBOX_GITHUB_TOKEN=... ./retry-github.sh

The server list, every result and the method are in mcp/scan. A server missing from the list is a pull request away.

If you run agents against these servers

Something between the agent and the server has to remember what was already done. agentsafe-mcp is one way: it logs each call before forwarding it, answers an exact repeat from that log, and when a call's outcome is lost to a crash or timeout it records it as unknown and never retries it blindly. It starts from a policy that agentsafe-mcp inspect writes for any server above.