MCP SERVER VENDOR QUESTIONNAIRE Version 1.0, 18 November 2026. https://isokron.ai/mws/vendor-questionnaire The questions come from the MCP Write-Safety Requirements (MWS), an open spec, draft 0.9, in public comment. Each one names the requirement IDs it rests on, so a vendor knows exactly what is being asked and you can read the requirement and its test method yourself. Requirements: https://isokron.ai/mws HOW TO ANSWER Answer each question in one of two ways: 1. A dated test record. The mcp-writecheck record for each requirement ID shown, for each write tool, with its verdict, the date of the run, the server version and the backend it ran against. 2. Yes or no, with evidence. A tool schema, a captured tool result, a test or the line of code that shows it. A yes without evidence is a claim. Reading a record: - held means the condition was forced and the server did what the requirement says in every trial. violated means the forbidden behavior was reproduced. - platform means the backend behaves the same way when called directly; the server passes it on, and the question becomes what the vendor does about it. - not observed, not tested and not applicable are not a yes. Each record of that kind states its reason, so read the reason. - A record is about one requirement and one write tool. MWS defines no total and no result for a server as a whole, so ask for the records, not a summary. The free, open-source tester mcp-writecheck can produce these records against the vendor’s own test instance with no account; as a preview it tests some requirements only in part and reports the rest as not tested. https://github.com/isokron-ai/mcp-write-safety/tree/main/packages/writecheck QUESTIONS 1. Do your write tools leave the backend’s own concurrency controls, such as version checks and locks, switched on unless the caller explicitly overrides them? Why it matters: A tool that forces an overwrite removes the one check that would have stopped an agent from replacing someone else’s edit: in our audit of the Grafana MCP server, dashboard patches skipped the version check that refused a stale save in 10 of 10 trials through Grafana’s own API. (https://isokron.ai/research/grafana-mcp-server-audit#findings-register) Requirements: MWS-CONC-1 (MUST), https://isokron.ai/mws#mws-conc-1 A good answer: A record with the verdict held for MWS-CONC-1 on every write tool whose backend operation has a control, with the control arm shown. Or yes, with each override argument and its default, which is off. Vendor answer: Evidence or record, with its date: 2. When an agent writes from data it read earlier and the object has changed since, do the tools that replace or delete an object, or that take a version, refuse the write with a conflict? Why it matters: An agent acts on a read that may be many steps old: in our audit of the GitHub MCP server, the file-delete tool took no version and deleted content written after the caller’s read. (https://isokron.ai/research/github-mcp-server-audit#findings-register) Requirements: MWS-CONC-2 (MUST), https://isokron.ai/mws#mws-conc-2 A good answer: A held record for MWS-CONC-2 on each replace, delete and versioned tool. Where the backend cannot detect the change, a platform record and a tool description that says stale writes are not detected. Or yes, with the name of the version argument on each tool and whether it is required. Vendor answer: Evidence or record, with its date: 3. If two callers change different fields of one object at the same moment, is every call that reports success actually kept? Why it matters: A lost update that reports success leaves nothing for anyone to notice: in our audit of the GitHub MCP server, 19 of 30 simultaneous pairs lost a change while both calls returned success. (https://isokron.ai/research/github-mcp-server-audit#findings-register) Requirements: MWS-CONC-3 (MUST), https://isokron.ai/mws#mws-conc-3 A good answer: A held record for MWS-CONC-3 from a forced schedule, or a not observed record from simultaneous pairs that states how many pairs overlapped (at least 30). A platform record shows both counts, through the server and straight to the backend. Or yes, with the script that sent two simultaneous calls on different fields many times and how many changes survived. Vendor answer: Evidence or record, with its date: 4. When a write is refused because of a conflict, is nothing applied, and is the refusal returned as a tool error that says so? Why it matters: An agent can recover from a conflict only if it can tell one happened: in our audit of the Payload MCP plugin, a refused write came back as plain text without the error flag in 3 of 3 trials. (https://isokron.ai/research/payload-mcp-server-audit#payload-findings) Requirements: MWS-CONC-4 (MUST), https://isokron.ai/mws#mws-conc-4 A good answer: A held record for MWS-CONC-4. Or yes, with a captured conflict result that has isError: true and says the write was not applied. Vendor answer: Evidence or record, with its date: 5. Does a write tool report success only when the state the caller asked for is in effect afterwards? Why it matters: An agent acts on what the result says, not on what the backend holds: in our audit of the GitHub MCP server, an update that skipped an empty title or body returned success with nothing changed. (https://isokron.ai/research/github-mcp-server-audit#findings-register) Requirements: MWS-RES-1 (MUST), https://isokron.ai/mws#mws-res-1 A good answer: A held record for MWS-RES-1 on each update and delete tool, with the settle interval the run used. Or yes, with what each tool returns when it changes nothing. Vendor answer: Evidence or record, with its date: 6. When a write fails, lands only in part or times out, does the result say what remains applied, or that the outcome is unknown, as a tool error? Why it matters: A failure reported as success is never retried, and an unknown outcome reported as a failure is retried into a duplicate: in our audit of the Payload MCP plugin, bulk updates and deletes reported success in 5 of 5 trials when some documents had failed. (https://isokron.ai/research/payload-mcp-server-audit#payload-findings) Requirements: MWS-RES-2 (MUST), https://isokron.ai/mws#mws-res-2 A good answer: A held record for MWS-RES-2 that includes the trial in which the backend applies the write and the answer is lost. Or the text of the error an agent sees after a backend timeout, which says the outcome is unknown and names the read that settles it. Vendor answer: Evidence or record, with its date: 7. Does every create tool return the identifier of what it created, including when a later step of the same call fails? Why it matters: Without the identifier an agent cannot find, fix or remove what it just made: in our audit of the GitHub MCP server, a pull request was created, a later step failed, and the error dropped the new pull request’s number and URL. (https://isokron.ai/research/github-mcp-server-audit#findings-register) Requirements: MWS-RES-3 (MUST), https://isokron.ai/mws#mws-res-3 A good answer: A held record for MWS-RES-3 on each create tool. Or yes, with the error result of a create whose later step failed, showing the identifier. Vendor answer: Evidence or record, with its date: 8. Can a caller repeat a create after a lost answer without making a second object, and is a write repeated with the same idempotency key applied at most once? Why it matters: A caller that gets no answer sends the write again: our audit of the GitHub MCP server notes two create tools that send no idempotency key, so a retry after a lost response makes a duplicate, which the audit records as behavior of the platform’s API that the server does not guard against. (https://isokron.ai/research/github-mcp-server-audit#other-observations) Requirements: MWS-RETRY-2 (MUST), https://isokron.ai/mws#mws-retry-2; MWS-RETRY-1 (SHOULD), https://isokron.ai/mws#mws-retry-1 A good answer: A held record for MWS-RETRY-2 on each tool that takes a key: the same key and arguments are applied once and answered with the first outcome, and the same key with different arguments is refused. Or yes, with the name of the key argument on each create tool and how long a key is remembered. Vendor answer: Evidence or record, with its date: 9. When a tool stops to ask the caller for more input, has it written nothing yet? Why it matters: In MCP revision 2026-07-28 the client answers and then sends the original request again, so a tool that wrote before asking can write twice; no published audit shows this yet, and the requirement rests on the protocol text. Requirements: MWS-RETRY-3 (MUST), https://isokron.ai/mws#mws-retry-3 A good answer: A not applicable record for MWS-RETRY-3 where no tool returns an input_required result, or a held or not observed record for each tool that does. Or yes, with the list of tools that ask for input. Vendor answer: Evidence or record, with its date: 10. When an agent sends an argument the tool does not define, or a value of the wrong type, does the server say so instead of dropping the argument or changing the value? Why it matters: A misspelled argument that is dropped becomes a success that did not do what was asked: in our audit of the PostHog MCP server, generated schemas read the string "false" as true and null as 0. (https://isokron.ai/research/posthog-mcp-server-audit#findings-register) Requirements: MWS-IN-1 (MUST), https://isokron.ai/mws#mws-in-1; MWS-IN-2 (MUST), https://isokron.ai/mws#mws-in-2 A good answer: held records for MWS-IN-1 and MWS-IN-2 on each tool tested, with the tools not exercised listed as not tested. Or yes, with the result of a call that carries one misspelled argument. Vendor answer: Evidence or record, with its date: 11. Does the server enforce the limits its tool schemas declare, and return a validation failure as a tool error that names the argument and the reason? Why it matters: A schema is a promise to the caller, and an error an agent cannot read is one it cannot fix: in our audit of the GitHub MCP server, page bounds advertised on four tools were not enforced, and one out-of-range value came back as a raw decode error. (https://isokron.ai/research/github-mcp-server-audit#findings-register) Requirements: MWS-IN-3 (MUST), https://isokron.ai/mws#mws-in-3; MWS-IN-4 (MUST), https://isokron.ai/mws#mws-in-4 A good answer: held records for MWS-IN-3 and MWS-IN-4. Or yes, with one captured validation error that has isError: true, the argument’s name and the reason. Vendor answer: Evidence or record, with its date: 12. If we turn on read-only mode or a tool deny-list, does every write tool refuse, and does a mistyped value or tool name stop the server or switch writes off instead of leaving them on? Why it matters: A safety setting that fails open protects nothing while looking as if it does: in our audit of the GitHub MCP server, a read-only environment variable set to yes, on or a typo left the write tools enabled. (https://isokron.ai/research/github-mcp-server-audit#findings-register) Requirements: MWS-CFG-1 (MUST), https://isokron.ai/mws#mws-cfg-1; MWS-CFG-2 (MUST), https://isokron.ai/mws#mws-cfg-2; MWS-CFG-3 (MUST), https://isokron.ai/mws#mws-cfg-3 A good answer: held records for MWS-CFG-1, MWS-CFG-2 and MWS-CFG-3, which are tests that start the server with each setting. Or yes, with what the server does on an unrecognized value, for example that it refuses to start. Vendor answer: Evidence or record, with its date: WHAT THESE QUESTIONS LEAVE OUT The questions leave out four areas of MWS that matter for some servers: updates that reset fields the caller did not name, writes that land outside the scope or on more objects than the caller named, multi-item writes that are applied in part, and annotations that do not match behavior. They also leave out sign-in, permissions and everything else a security review covers. For the full list, ask for a conformance statement, which the spec defines as every requirement with its verdicts for each write tool tested. PROCUREMENT CLAUSE Before the Customer enables agent access to the Supplier’s MCP server, and within 30 days of each release that adds or changes a write tool, the Supplier will give the Customer dated test records for each write tool against the requirements of the MCP Write-Safety Requirements (MWS), draft 0.9, listed in Schedule [X], produced with mcp-writecheck or another implementation of the test methods the specification publishes, and stating the server version, the backend, the declaration used and what was not tested. The Supplier will notify the Customer within [10] business days of any record with the verdict “violated” for a requirement at MUST level, with its plan and date for a fix, and the Customer may suspend agent write access for the affected tool until a later record for that requirement and tool no longer shows that verdict. - Schedule [X] is the list of requirement IDs you care about; the 17 IDs cited above are a reasonable start. - The clause pins draft 0.9 on purpose. Requirement IDs are stable and are never renumbered, but wording and levels can change before 1.0. - It asks for records and notice, not for a result for the server as a whole, because the spec defines none. - This is a starting point for your own counsel to adapt, not legal advice. MCP Write-Safety Requirements (MWS) is an open spec, draft 0.9, in public comment. IsoKron is its steward and maintains the tester.