Skip to content
IsoKron

MWS 0.9 · Draft for public comment

MCP Write-Safety Requirements (MWS)

Two agents edit one record. Both are told the write succeeded, and one change is gone. MWS is 37 testable requirements for MCP servers whose tools create, change or delete data: a write must do what the caller asked, to the object the caller named, and be reported truthfully.

The test tool is a preview: it runs the tests of 14 requirements in full and 23 in part, and reports the rest as not tested, with the reason. Coverage by requirement · What is MCP write safety?

Shell
npx -y -p @mcp-write-safety/writecheck@0.1.0 -p @mcp-write-safety/tiny-tracker@0.1.0 mcp-writecheck demo --defect MWS-CONC-2

Node 22 or later · no account · runs on 127.0.0.1, in memory; nothing leaves the machine.

Output
MWS-UPD-1    MUST   held, not applicableMWS-UPD-2    MUST   heldMWS-CONC-1   MUST   violated in ticket:delete_ticketMWS-CONC-2   MUST   violated in ticket:delete_ticketMWS-CONC-3   MUST   not observed, heldMWS-CONC-4   MUST   held, self-attested, not testedMWS-CONC-5   SHOULD self-attested, not applicable, not testedMWS-CONC-6   SHOULD held…MWS-SEC-1    MUST   not observed 285 records in 22.7 s. One line per requirement; the report has one row per write path. There is no total.mcp-writecheck is a preview: of the 37 requirements of MWS 0.9 it executes the test of 14 in full and of 23 in part; what is not built is reported "not tested" with the reason (README: Coverage in this preview).
A real mcp-writecheck demo run on tiny-tracker, the sample server, with one defect on: a delete that ignores the version read. Each line lists the verdicts across that requirement’s write paths; the report keeps one record for each.

● Draft 0.9, first public reviewComments until 16 Dec 2026Stewarded by IsoKron (our interest)Path to 1.0Decision logCiteDiscuss on GitHub

A lost update, in one picture

Panel 1 · Without a version check

  1. A reads v3 · B reads v3
  2. B adds a label → v4
  3. A saves the whole record as it read it, with a new title → v5
  4. A: “success” · B: “success”
  5. B’s label: lost

Panel 2 · With a version check (MWS-CONC-2)

  1. A reads v3 · B reads v3
  2. B adds a label → v4
  3. A’s write carries v3 → refused, nothing written
  4. A reads v4 and saves its title on top → v5
  5. B’s label: kept
Figure 1: Schematic. A version check refuses a stale version, not stale content: resending what A read at v3 under v4 would pass, so A rebuilds its write from v4. In our audit of the GitHub MCP server v2.1.0, concurrent issue_write calls on different fields of one issue lost a change in 19 of 30 pairs, both reporting success; GitHub’s per-field update lost 0 of 28.

Quickstart

Try the sample, test your own server, add the check to CI, fix what it finds.
Caution:

run and the Action’s full mode make real writes: point them only at a test environment you may change. plan writes nothing.

CLI

Shell
# 1. Try it: the sample server, every deliberate defect on (127.0.0.1, in memory)
npx -y -p @mcp-write-safety/writecheck@0.1.0 -p @mcp-write-safety/tiny-tracker@0.1.0 mcp-writecheck demo --defect all

# 2. Your server, in a test environment
npm install --save-dev @mcp-write-safety/writecheck@0.1.0
npx mcp-writecheck init --out writecheck.yaml --env BACKEND_URL -- node ./my-server.js   # drafts from tools/list; every guess is marked
npx mcp-writecheck plan -m writecheck.yaml                                                # every call a run would make, and the plan hash
npx mcp-writecheck run  -m writecheck.yaml                                                # asks you to type the target and the plan hash
Recorded run on the sample server: plan, run
$ mcp-writecheck plan -m writecheck.yamlPlan b78ef2a3aba90e28 for target tiny-tracker (environment: test)Hosts reached or given to the server: 127.0.0.1:4100At most 1601 writes and 282 created objects (write budget 20000, 1000 writes per second).No tool call was made to build this plan.…plan hash: b78ef2a3aba90e28 $ mcp-writecheck run -m writecheck.yaml --confirm-target tiny-tracker --plan-hash b78ef2a3aba90e28…ticket: stale-write…ticket: barrier-pairs…ticket: held-request……no requirement was violated in this run (see report.md for what was not tested) results: mwsc-results/2a334a745d/report.md (also results.json, junit.xml, SHA256SUMS)

run asks you to type the target and the plan hash; CI passes them as flags, and there is no --yes. Objects a run creates carry the prefix mwsc-<run>- and are deleted at the end; cleanup handles an interrupted run. A write budget and rate limit apply (default 500 writes, 5 per second), and a run stops at the first HTTP 429.

CLI README and coverage

GitHub Action

plan needs no test environment; full needs one.

YAML
# .github/workflows/mcp-writecheck.yml (excerpt of examples/mcp-writecheck.yml)
on:
  pull_request:
  push:
    branches: [main]
permissions:
  contents: read
jobs:
  plan:
    if: github.event_name == 'pull_request'
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7.1.0
        with:
          node-version: 22
      - run: npm ci && npm run build
      - uses: isokron-ai/mcp-writecheck-action@d4cfcad2506e118d49a021cb47a28b081ca7bf9c
        with:
          manifest: writecheck.yaml
          mode: plan            # no tool call, no write, no secret
  full:
    if: github.event_name == 'push'
    runs-on: ubuntu-latest
    concurrency: { group: mcp-writecheck-test-environment, cancel-in-progress: false }
    env:
      API_TOKEN: ${{ secrets.TEST_BACKEND_TOKEN }}   # valid only in the test environment
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
      - uses: actions/setup-node@949feb2413d6458794dcd2491c4babbbce0c15c1 # v7.1.0
        with:
          node-version: 22
      - run: npm ci && npm run build
      - uses: isokron-ai/mcp-writecheck-action@d4cfcad2506e118d49a021cb47a28b081ca7bf9c
        with:
          manifest: writecheck.yaml
          mode: full            # real writes to the test environment
          fail-on: must
          confirm-target: staging-tracker.internal   # the manifest's target.confirm
          plan-hash: 0123456789abcdef                # from the plan, after a person has read it

The plan hash covers the manifest, the tool list, the suite version, the limits and the hosts in the server’s environment. If any of them changes, full refuses to run until someone reads the new plan and updates the hash. The hash records which plan was accepted; it does not itself require an approval.

Action README

Kits and skills

Shell
npm install @mcp-write-safety/kit        # TypeScript, Node 20+, no runtime dependency
pip install mcp-write-safety-kit         # Python 3.10+, MCP Python SDK 1.x and 2.x (import mws_kit)
Shell
# Agent skill for coding agents that load SKILL.md folders
SKILLS_DIR=./skills
mkdir -p "$SKILLS_DIR/mcp-write-safe-builder"
curl -fsSL -o "$SKILLS_DIR/mcp-write-safe-builder/SKILL.md" \
  https://raw.githubusercontent.com/isokron-ai/mcp-writecheck-action/d4cfcad2506e118d49a021cb47a28b081ca7bf9c/skills/mcp-write-safe-builder/SKILL.md

The kits add version checks, retry keys, strict inputs and honest results to an existing server. Their version check holds within one server process, so also pass the version to the backend’s own check. The skills give a coding agent the requirements as rules. Only the tests say whether a server meets MWS.

What MWS covers

Ten areas, 37 requirements (26 MUST, 11 SHOULD, in the BCP 14 sense), each with a stable ID, a condition, a test method and the evidence behind it.

MWS tests whether writes are correct; it is not a security test. Authentication, authorization, transport security, prompt injection and agent behavior are out of scope; one narrow requirement covers credentials in tool results.

  • UPD

    Update semantics

    An update changes only the fields it names.

    2 MUST

  • CONC

    Concurrency and preconditions

    Refuse a stale write wherever the backend can detect it.

    4 MUST · 2 SHOULD

  • TGT

    Targeting

    A write stays in its named scope; an empty selector never means everything.

    3 MUST

  • RES

    Truthful results

    Success only when it is true; a failure says what remains applied, or, only when it cannot know, that the outcome is unknown.

    3 MUST · 1 SHOULD

  • ATOM

    Multi-step and multi-item writes

    A partial write is not a success and names what landed.

    1 MUST · 1 SHOULD

  • RETRY

    Retries and replays

    A write repeated with the same idempotency key is applied at most once.

    2 MUST · 1 SHOULD

  • IN

    Inputs

    No meaning-changing coercion, no silently dropped arguments.

    4 MUST · 3 SHOULD

  • CFG

    Safety settings

    Read-only mode refuses every write; a misread setting never leaves writes on.

    3 MUST · 1 SHOULD

  • ANN

    Annotations

    An explicit hint such as readOnlyHint: true must be true.

    3 MUST · 2 SHOULD

  • SEC

    Credentials in results

    No token, secret or signed URL in a tool result.

    1 MUST

Short forms only; the requirement is the sentence in section 9 of the specification.

UPD Update semantics⁠

IDLevelShort formSpecification
MWS-UPD-1⁠⁠MUSTA field patch changes only the named fields
Requirement sentence

An update tool in field-patch mode MUST change only the fields its arguments name, so that every other field of the object, apart from server-managed fields, keeps its stored value and no default is written that the caller did not send.

Full text of MWS-UPD-1
MWS-UPD-2⁠⁠MUSTA list edit, move or removal lands on the element the caller read at that address, or is refused
Requirement sentence

A tool that edits, moves or removes one element of a list MUST apply the change to the element that the caller's address denoted in the list as the caller read it, or refuse the call, when elements have been inserted, removed or reordered since that read.

Full text of MWS-UPD-2

CONC Concurrency and preconditions⁠

IDLevelShort formSpecification
MWS-CONC-1⁠⁠MUSTDo not bypass the backend's concurrency control by default
Requirement sentence

A write tool MUST NOT disable or bypass a concurrency control that the backend offers for the operation it performs, unless the caller sets an explicit override argument whose default is off.

Full text of MWS-CONC-1
MWS-CONC-2⁠⁠MUSTRefuse a stale write where the backend can detect it
Requirement sentence

A tool that replaces a whole object, deletes an object or accepts a version MUST refuse the write with a conflict when the object has changed since the state the caller read, wherever the backend can detect the change.

Full text of MWS-CONC-2
MWS-CONC-3⁠⁠MUSTConcurrent writes to different parts do not both succeed with one lost
Requirement sentence

Two concurrent writes through the server to different parts of one object MUST NOT both report success when either change is absent afterwards.

Full text of MWS-CONC-3
MWS-CONC-4⁠⁠MUSTA conflict is an error, is not applied and says so
Requirement sentence

A write refused because of a conflict MUST NOT be applied and MUST be returned as a tool execution error (isError: true) that says the write was not applied.

Full text of MWS-CONC-4
MWS-CONC-5⁠⁠SHOULDThe description discloses stale writes the backend cannot detect
Requirement sentence

The description of a tool that replaces a whole object, deletes an object or accepts a version SHOULD state that stale writes are not detected, or are detected only on a best-effort basis by the server, where the backend cannot detect the change.

Full text of MWS-CONC-5
MWS-CONC-6⁠⁠SHOULDA conflict tells the caller to read again before retrying, or carries the current state
Requirement sentence

A conflict error SHOULD tell the caller to read the object again before it retries, or carry the object's current state.

Full text of MWS-CONC-6

TGT Targeting⁠

IDLevelShort formSpecification
MWS-TGT-1⁠⁠MUSTHonor or reject a scope argument
Requirement sentence

A write tool that accepts a scope argument MUST apply the write within the scope the argument names or reject the call, and MUST NOT apply it anywhere else.

Full text of MWS-TGT-1
MWS-TGT-2⁠⁠MUSTAn empty or match-all selector does not mean everything
Requirement sentence

A tool that deletes or bulk-updates MUST NOT treat an empty, missing or match-all selector, or the path of a container, as selecting every object in the container or collection unless the caller sets an explicit argument that opts in.

Full text of MWS-TGT-2
MWS-TGT-3⁠⁠MUSTA create does not overwrite
Requirement sentence

A create tool MUST NOT overwrite an existing object that has the same name or key unless the caller sets an explicit argument that asks for the overwrite.

Full text of MWS-TGT-3

RES Truthful results⁠

IDLevelShort formSpecification
MWS-RES-1⁠⁠MUSTSuccess only when the state asked for is in effect
Requirement sentence

A write tool MUST report success only when the state the caller asked for is in effect afterwards.

Full text of MWS-RES-1
MWS-RES-2⁠⁠MUSTA failed write is not a success, says what remains applied, and is a tool execution error where the backend or a rule refused it
Requirement sentence

A write that fails in whole or in part MUST NOT be reported as a success; the error MUST identify any change that remains applied, or say that it is not known whether a change was applied, and MUST NOT report a settled outcome as not known, and MUST NOT tell the caller to repeat a write whose outcome is open, except after a read that it names or with the idempotency key or the version that the call carried; and where the failure is a refusal or failure of the backend, or of a rule the tool applies, the error MUST be returned as a tool execution error (isError: true).

Full text of MWS-RES-2
MWS-RES-3⁠⁠MUSTA create returns its identifier, even on later failure
Requirement sentence

A create tool MUST return the identifier of each object it created, including when a later step of the same call fails.

Full text of MWS-RES-3
MWS-RES-4⁠⁠SHOULDA write returns a receipt
Requirement sentence

A write tool SHOULD return a receipt that identifies the object written, the fields changed and, where the resource has a version, its new version.

Full text of MWS-RES-4

ATOM Multi-step and multi-item writes⁠

IDLevelShort formSpecification
MWS-ATOM-1⁠⁠MUSTA partial outcome is not success, and names what was applied
Requirement sentence

A multi-item or multi-step write that is not applied completely MUST NOT report success and MUST name which items or steps were applied and which were not.

Full text of MWS-ATOM-1
MWS-ATOM-2⁠⁠SHOULDMulti-object writes are all-or-nothing
Requirement sentence

A multi-object write SHOULD be all-or-nothing, so that when any part of it fails none of its changes remains applied.

Full text of MWS-ATOM-2

RETRY Retries and replays⁠

IDLevelShort formSpecification
MWS-RETRY-1⁠⁠SHOULDA non-idempotent create accepts an idempotency key
Requirement sentence

A create tool whose repeated call would create a second object SHOULD accept an idempotency key from the caller.

Full text of MWS-RETRY-1
MWS-RETRY-2⁠⁠MUSTAn idempotency key is honored: applied at most once and answered with the first outcome where the server can establish it, new key new write, changed arguments refused
Requirement sentence

A write repeated with the same idempotency key and the same arguments MUST be applied at most once and, where the first call applied it and the server can establish that, MUST be answered with the outcome of that call; a write with a different key MUST be treated as a new write; and a write that reuses a key with different arguments MUST be refused.

Full text of MWS-RETRY-2
MWS-RETRY-3⁠⁠MUSTThe write asked for is not in effect before an input_required result
Requirement sentence

A tool that answers with an input_required result MUST NOT have put into effect, in that call, the write the caller asked for, in whole or in part.

Full text of MWS-RETRY-3

IN Inputs⁠

IDLevelShort formSpecification
MWS-IN-1⁠⁠MUSTNo coercion that changes the meaning of a value
Requirement sentence

A tool MUST NOT coerce an argument to a value whose meaning differs from the value the caller sent.

Full text of MWS-IN-1
MWS-IN-2⁠⁠MUSTUnknown arguments are never dropped silently
Requirement sentence

A tool MUST NOT silently drop an argument that its input schema does not define.

Full text of MWS-IN-2
MWS-IN-3⁠⁠MUSTA call that breaks a declared constraint is rejected
Requirement sentence

A tool MUST reject a call that breaks a constraint its input schema declares with a keyword listed in the applicability condition.

Full text of MWS-IN-3
MWS-IN-4⁠⁠MUSTA validation failure is an error naming the argument and reason
Requirement sentence

A tool MUST report an input validation failure as a tool execution error (isError: true) that names the argument at fault and the reason it was refused.

Full text of MWS-IN-4
MWS-IN-5⁠⁠SHOULDThe schema declares the fixed single-argument constraints that are enforced
Requirement sentence

A tool's input schema SHOULD declare, with the keywords listed in MWS-IN-3, each fixed constraint on the value of a single argument that the server or its backend enforces.

Full text of MWS-IN-5
MWS-IN-6⁠⁠SHOULDWrong types are rejected, not converted or dropped
Requirement sentence

A tool SHOULD reject an argument whose value is not of a type its input schema declares for that argument, rather than convert or drop it.

Full text of MWS-IN-6
MWS-IN-7⁠⁠SHOULDUnknown arguments are rejected
Requirement sentence

A tool SHOULD reject an argument that its input schema does not define.

Full text of MWS-IN-7

CFG Safety settings⁠

IDLevelShort formSpecification
MWS-CFG-1⁠⁠MUSTAn unrecognized safety setting never leaves writes enabled
Requirement sentence

A server MUST NOT accept a write when a read-only or other write-restricting setting, applied at start or with a request, has a value it does not recognize.

Full text of MWS-CFG-1
MWS-CFG-2⁠⁠MUSTAn unknown tool name in a deny-list stops the server or the request, or disables writes
Requirement sentence

A server MUST refuse to start, refuse to serve the request that carries the list, or disable every write tool, when a tool deny-list contains a name that is not the name of a tool the server defines.

Full text of MWS-CFG-2
MWS-CFG-3⁠⁠MUSTRead-only mode refuses every write tool
Requirement sentence

In read-only mode a server MUST refuse every call to a write tool, whatever that tool's annotations declare.

Full text of MWS-CFG-3
MWS-CFG-4⁠⁠SHOULDAn unknown tool name in an allow-list stops the server or the request
Requirement sentence

A server SHOULD refuse to start, or refuse to serve the request that carries the list, when a tool allow-list contains a name that is not the name of a tool the server defines.

Full text of MWS-CFG-4

ANN Annotations⁠

IDLevelShort formSpecification
MWS-ANN-1⁠⁠MUSTreadOnlyHint: true means no state change
Requirement sentence

A tool that declares readOnlyHint: true MUST NOT change state.

Full text of MWS-ANN-1
MWS-ANN-2⁠⁠MUSTdestructiveHint: false means no deletion, and no removal or replacement of data the caller did not name
Requirement sentence

A tool that declares destructiveHint: false MUST NOT delete an object, and MUST NOT remove or replace data that the caller's arguments did not name.

Full text of MWS-ANN-2
MWS-ANN-3⁠⁠MUSTidempotentHint: true means no further effect on repeat
Requirement sentence

A tool that declares idempotentHint: true MUST have no further effect when it is called again with the same arguments.

Full text of MWS-ANN-3
MWS-ANN-4⁠⁠SHOULDreadOnlyHint and destructiveHint are declared explicitly
Requirement sentence

A server SHOULD declare readOnlyHint explicitly on every tool and destructiveHint explicitly on every tool that does not declare readOnlyHint: true.

Full text of MWS-ANN-4
MWS-ANN-5⁠⁠SHOULDdestructiveHint: false means additive only (at risk)
Requirement sentence

A tool that declares destructiveHint: false SHOULD perform only additive updates, as the MCP schema states.

Full text of MWS-ANN-5

SEC Credentials in results⁠

IDLevelShort formSpecification
MWS-SEC-1⁠⁠MUSTNo credentials in tool results, other than one the tool exists to issue
Requirement sentence

A tool result, whether it reports success or an error, MUST NOT contain a credential, including a token, a secret, or a URL that carries a token or a signature.

Full text of MWS-SEC-1

The tools

All six are Apache-2.0, use no model, need no account and send no telemetry. Any tester that follows the specification’s test rules counts equally.
NameWhat it doesStatus
mcp-writecheckCLI: tests your server; 14 requirements in full, 23 in partPreview · Node 22+
mcp-writecheck-actionRuns the CLI in CIDraft
@mcp-write-safety/kitTypeScript helpers0.1.0 · Node 20+
mcp-write-safety-kitPython helpers (mws_kit)0.1.0 · Python 3.10+
Agent skillsOne for building servers, one for calling write tools2 skills
tiny-trackerSample server with defects you can switch on0.1.0

write-result/0.1: an optional result object (Annex B) that says what a write did. No requirement depends on it; both kits return it.

Example
structuredContent
{
  "mws": {
    "shape": "write-result/0.1",
    "outcome": "applied",
    "repeat": "same_version",
    "operation": "update",
    "object": { "type": "item", "id": 42, "version": 8 },
    "changed": ["/title"],
    "read_with": { "tool": "get_item", "arguments": { "id": 42 } },
    "version_argument": "expected_version"
  }
}

Worked example: how writesafe was built to the standard

IsoKron builds both MWS and writesafe, its MCP plugin for Strapi 5: this is our own implementation, not independent evidence, and MWS’s test sample is tiny-tracker. writesafe 1.0.0 is MIT-licensed and on npm.

  • Version checks. A stale field write is refused with a tool error that carries the current values and version and says whether to re-read (MWS-CONC-2, MWS-CONC-4, MWS-CONC-6, MWS-RES-2).
  • Scoped bulk writes. A missing or empty filter is refused; selecting everything needs allow_all: true (MWS-TGT-2).
  • Retry keys. A repeated versioned write returns its first reply; idempotency_key covers other writes (MWS-RETRY-2).

A writesafe receipt:

writesafe receipt
{"ok": true, "documentId": "…", "locale": null, "previous_version": "v2.K3nO0z0n2lq3QW8v5zR7m1x9dYbS4tH6uFjE2cAaB0g", "version": "v2.9Jx1sUe4hTqP0aLmN7cV3bR8kWzY5gD2fH6oQiE1uXs", "changed": ["note"]}

writesafe on GitHub

Where the requirements come from

Every finding MWS 0.9 cites is public and reproducible, and all come from IsoKron’s own audits of five servers, published 9 to 14 November 2026. Eight requirements rest on reasoning rather than a finding; each entry’s Evidence line says which. A finding shows a failure occurs in practice, not the state of a project’s current release.
  • Study · read from source

    Do popular packaged MCP servers say whether a write happened?

    Of 100 servers sampled from the registry’s most-downloaded packaged servers, 44 return a machine-readable success or failure, 17 a safe-to-repeat or replay signal, 7 a machine-readable “unknown” and 8 a structured conflict code.

    Limits: exploratory; read from source, not run; one to three write handlers per server; coded by AI agents only.

  • Study · exploratory pilots

    Conflict errors and an AI agent

    On two servers we built that refuse stale writes, we varied only the conflict error’s text. Told not to re-read, with a label spelled like the current version, the agent lost the other writer’s change in 38 of 48 runs when the error listed the state unquoted, and 0 of 48 when quoted.

    Limits: almost all runs on one model; the quoting result rests on one server, one trap and one prompt; the registered confirmatory study has not been run.

Three of the five audits behind MWS 0.9; all five are in the findings register.
Server and versionOne findingCited underAudit
GitHub MCP server v2.1.0Concurrent issue_write calls from two server processes lost a change in 19 of 30 pairs (Figure 1).MWS-UPD-1, MWS-CONC-3Audit: GitHub MCP server v2.1.0
PostHog MCP serverSimultaneous updates lost a write via the server (7 of 11) and the API (6 of 11): platform behavior.NF-1Audit: PostHog MCP server
Kubernetes MCP server v0.1.2An install sent a chart’s objects to another namespace, 3 of 3 runs.MWS-TGT-1Audit: Kubernetes MCP server v0.1.2

Findings register (CC BY 4.0)

How a result reads

A result is one requirement, one write path, one verdict, printed with its level; several paths or methods give several results. There is no total, score or badge, and no verdict called “pass”.
violated
the forbidden behavior was reproduced.
platform
the platform’s own API does it too; the server passes it on.
held
forced, and handled correctly every time.
not observed
attempted, not forced, not seen; not evidence of meeting it.
not tested
not run, with the reason.
not applicable
the condition is false here.
self-attested
a person’s answer to a review question.

Note: Expect violated rows

We expect most servers today, well-built ones included, to show some violated rows. MWS-RES-2, for one, asks a server to say the outcome is unknown when a backend’s answer is lost; in our study of 100 popular packaged servers, 7 had a machine-readable value for “unknown”, and 20 said a write’s outcome was unknown or partly applied, at least in prose. A violated row describes one forced situation, not a grade. Decide whether it matters for your users, and comment if a requirement asks too much.

To say a server was tested, publish this sentence. The conformance statement it links to lists every requirement’s verdicts, what was not tested and why, and the declaration: a record, not a claim. IsoKron approves, certifies and endorses nothing.

Say you tested
Tested against MWS 0.9 on [date] with [tool and version]; conformance statement at [URL].

Status and the path to 1.0

1.0 has no date. It is published when:

  • every requirement active in 1.0 has an executable test, a sample-server defect and the fixed build;
  • every comment has a published disposition;
  • the suite has run on at least five servers other than the sample server, with results public by agreement with their maintainers: at least two that IsoKron did not audit, written in at least three languages, in front of at least three kinds of backend, from at least four maintainers, none of them IsoKron;
  • a selection rationale is published with the results, and for the servers IsoKron chooses, it is dated before their runs;
  • at least one of those runs is made by a tester outside IsoKron who did not write the suite and is not employed or contracted by IsoKron; the results state any other relation to IsoKron.

The last three are not met in 0.9. Until 1.0, wording and levels may change; IDs are never renumbered or reused.

License, governance and interests

Note: Our interest

IsoKron writes and stewards MWS. We also sell audits and builds of MCP servers, which may report results against MWS: a commercial interest. Every finding in 0.9 comes from our own audits, and we endorse no result, our own included. The text was drafted with AI assistance; its author reviewed every normative sentence. Neither the MCP project nor any cited maintainer has endorsed MWS.

Until a maintainer from outside IsoKron joins, IsoKron decides each change after public comment and records its reasons; after that, each change needs two maintainers, and a change to a MUST needs one from outside IsoKron. We will propose how maintainers are chosen during the window.

  • Licenses: specification CC BY 4.0; software Apache-2.0; the MWS name and mark under the usage policy (TRADEMARKS.md).
  • Name and mark (summary of TRADEMARKS.md): The MWS mark and the name “MCP Write-Safety Requirements” belong to Stormhaven Enterprises LLC, doing business as IsoKron, and are not covered by the CC BY 4.0 license of the text or the Apache License 2.0 of the software. Anyone may use the unmodified mark to refer to the specification: next to a link to it, in an article, talk or list of standards, or in a tool that tests against it. Do not place it on or next to a product, server, package or result, use it as a status icon, combine it with your own name or logo, or alter it, and do not use it to suggest that anything has met MWS or is approved or endorsed by MWS or IsoKron. To say a server was tested, use the sentence in section 15.5 of the specification, without the mark. As an exception, IsoKron’s own MWS implementation packages (the kits, the checker and the Action) may display the mark. IsoKron approves, certifies and endorses nothing.
  • Contributions: DCO sign-off, no CLA, under the license of the file you change.

Contributing

Questions

Is MWS part of MCP?

No. It is an independent layer above the MCP specification and adds nothing to the protocol. 0.9 is checked against MCP revisions 2025-11-25 and 2026-07-28.

How does it relate to MCP’s conformance tests and to scanners?

They test protocol behavior and tool metadata. MWS tests what a tool call does to backend state.

Does a read-only server need MWS?

Mostly not. Only inputs, annotations and credentials in results apply to read tools.

Cite this specification

IsoKron (2026). MCP Write-Safety Requirements (MWS). Version 0.9, 16 November 2026. IsoKron. https://isokron.ai/mws

Attribution the license asks for: MCP Write-Safety Requirements, version 0.9, Joshua Bryan, IsoKron, CC BY 4.0, with a link to the source.

BibTeX
@techreport{isokron2026-mws,
  author      = {{IsoKron}},
  title       = {{MCP} Write-Safety Requirements ({MWS})},
  institution = {IsoKron},
  type        = {Draft specification},
  year        = {2026},
  month       = nov,
  version     = {0.9},
  url         = {https://isokron.ai/mws},
  note        = {Version 0.9, 16 November 2026. Attribution: {MCP} Write-Safety Requirements, version 0.9, Joshua Bryan, {IsoKron}, {CC} {BY} 4.0}
}

Comment on MWS 0.9

Open until 16 December 2026. Each comment is a GitHub issue and gets a recorded disposition; where one leads to a change, IsoKron prepares it. To comment on one requirement, use Comment beside its ID in the requirement index.

No GitHub account? Email comments@isokron.ai: we copy each comment into a public issue with your permission.