mcp-migrate

mcp-migrate finds and fixes what the MCP 2026-07-28 spec revision breaks in your server: protocol sessions and Mcp-Session-Id removed, initialize / notifications/initialized replaced by server/discover, ping and logging/setLevel gone, resources/subscribe replaced by subscriptions/listen, required resultType and cache metadata on results, server-initiated Sampling/Roots/Elicitation replaced by Multi Round-Trip Requests, and more. The official Python SDK ships no codemod for any of this -- only a manual migration guide. The TypeScript codemod only handles the v1→v2 package rename, not the protocol changes. mcp-migrate fix is the only thing that edits your server's code for you.

uvx mcp-migrate check .
uvx mcp-migrate fix . --write

mcp-migrate check

$ uvx mcp-migrate check tests/fixtures/fixer_roundtrip

mcp-migrate v0.2.0  ->  fixer_roundtrip
2 Python files, 21 rules, spec 2026-07-28

            rule    where         what
breaking    R001    server.py:28  Mcp-Session-Id was removed from the Streamable HTTP transport.
breaking    R001    server.py:29  Mcp-Session-Id was removed from the Streamable HTTP transport.
breaking    R017    errors.py:8   -32002 for resource-not-found is the old code; 2026-07-28 uses -32602.
deprecated  R006    server.py:15  HTTP+SSE transport is deprecated.
deprecated  R006    server.py:24  HTTP+SSE transport is deprecated.
advisory    R004    server.py:32  Tools are returned without an explicit sort.
advisory    R005    server.py:16  Capabilities are declared but `extensions` is absent.
advisory    R010    (project)     This project registers MCP request handlers (tools/resources/prompts)
                                  but has no server/discover implementation anywhere in the project.
advisory    R016    server.py:32  This file implements a list/read handler but neither `ttlMs` nor
                                  `cacheScope` appears in it.

  R001  Uses Mcp-Session-Id, which no longer exists
  SEP-2567 https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2567
  Sessions are gone from the transport. Mint an explicit handle server-side and take it as an
  ordinary tool argument instead.

  [... one block like this per rule that fired ...]

Grade F (26/100)  3 breaking, 2 deprecated, 4 advisory

Add your server to the board:  mcp-migrate entry --repo owner/name

Zero findings prints Grade A and a ready-to-paste badge instead. Add --json for machine-readable, uncapped output (the terminal table caps each rule at 5 rows with a "+N more" line so it stays readable; JSON always has every finding).

--rule R001 (repeatable, e.g. --rule R001 --rule R017) restricts which rules run at all, not just what's printed -- the natural way to adopt this tool on an existing codebase is one rule at a time. An unknown id is a usage error (exit 2), same as --write --dry-run together on fix. A grade computed from part of the rule set isn't a grade, so --rule suppresses it entirely rather than printing one that only ever looked at a subset.

--severity breaking (repeatable, composes with --rule) only changes what gets displayed -- the grade and the exit code are always computed from every finding, so two runs against the same code never disagree just because someone chose to look at a narrower slice of the output.

--fail-on {breaking,deprecated,advisory,never} (default breaking) is the minimum severity that fails the run. A team mid-migration that has decided deprecated blocks a merge sets --fail-on deprecated; a team still adopting the tool that wants findings reported without failing the build yet sets --fail-on never -- which is not the same as check || true: exit 2 ("could not check it") is a refusal, not a severity, and no --fail-on setting suppresses it.

check --json contract

--json emits one JSON object with these always-present top-level keys: tool, version, spec, path, scannable, languages, grade, score, rule_filtered, filters, fail_on, files_scanned, counts, findings, suppressed, and unused_suppressions.

• filters is {"rule": [...], "severity": [...]}, the ids/severities passed to --rule/--severity, empty arrays when neither was given.
• rule_filtered is true whenever --rule was given. counts and findings reflect the requested filters; grade and score are always null when rule_filtered is true.
• fail_on echoes the --fail-on threshold that decided the process exit code, so a consumer doesn't have to re-derive it.

Conditional keys:

• is_sdk and sdk_reason only when the tree is a protocol SDK. In that case grade and score are null.
• reason only when scannable is false. In that case grade and score are null.
• grade and score are also null when a tree was read but is not gradable, for example TypeScript-only trees.

suppressed holds findings silenced by an inline mcp-migrate: ignore[R0NN] comment; they are excluded from findings, from counts, and from the grade, and each carries the reason from its comment.

unused_suppressions holds directives that matched no finding — either it was fixed or the code moved. Each entry has rule, path, line, and reason, one per rule id. It is always present, empty array included: stale suppressions accumulate in CI, and CI is the consumer that reads --json and never sees a console line.

Each finding always has rule, severity, path, line, and message. path and line may be null for project-level findings. fix is present only when the rule has remediation text; it is omitted, never null.

The executable contract is schemas/check-json.schema.json. Pre-1.0, this shape may change in a breaking way; such changes are documented under Changed in CHANGELOG.md, so consumers should pin a version and watch that section.

check --format sarif

SARIF 2.1.0, for GitHub code scanning and anything else that speaks it:

- run: mcp-migrate check . --format sarif > mcp-migrate.sarif
- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: mcp-migrate.sarif

Findings land in the Security tab and annotate the diff, instead of scrolling past in a log. Severities map to SARIF levels as breaking -> error, deprecated -> warning, advisory -> note. deprecated is deliberately not error: code scanning's default gate fails on error alone, and the spec gives deprecated features 12+ months, so blocking a merge today over a change that doesn't break until next year is how an integration gets switched off.

--format {text,json,sarif} is the general flag; --json remains as an alias for --format json. Output is validated against the vendored SARIF 2.1.0 schema in CI.

Suppressing a finding

This tool has open false-positive classes and says so above. When a finding is wrong, or the code is deliberate and not changing, silence that one line rather than the whole rule:

mcp_session_id = req.headers["X-Sid"]  # mcp-migrate: ignore[R001] -- proxy shim, not MCP session state

const mcpSessionId = req.headers["x-sid"];  // mcp-migrate: ignore[R001] -- proxy shim

The rule id is required — a blanket ignore would also silence rules that don't exist yet, and nobody ever revisits it. A reason after -- is expected; check reports directives that lack one.

A suppressed finding doesn't count against the grade — a suppression that still costs you the grade isn't one, and the only move left would be to stop running the tool. That does make the grade partly self-reported, so three things keep it auditable:

• the count is always printed, never behind a flag
• --show-suppressions lists every one with its file, rule and reason
• check reports suppressions that matched nothing, so stale ones don't quietly accumulate

--json carries them under suppressed, each with its reason, and the stale ones under unused_suppressions.

One caveat, because "line-scoped" promises slightly less than it delivers for three rules. R005, R015 and R016 ask a question about a file — does this file declare capabilities without extensions, does it build a JSON-RPC result without resultType, does it implement a list/read handler without cache metadata — so each reports at most one finding per file, on the first line that matches. Suppressing that line therefore silences the rule for the whole file, not just that line. Every other rule reports each occurrence separately, and there suppression really is per line. If you suppress one of these three, you are accepting the rule's verdict on that file.

Exit codes, so it drops straight into CI:

2 means we did not read your code, not "your code is fine" — an empty finding set otherwise conflates "we read it and it's clean" with "we read nothing", and a grade that can't tell those apart is worthless. A tree with nothing readable in it gets silence instead of an A.

Python is graded. TypeScript is scanned but still not graded — and as of R002 landing, that is no longer justified by coverage. All 21 rules read TypeScript. The grade is withheld by a PARTIAL flag, and now that coverage is complete the reason it gives is a decision, not a fraction:

$ mcp-migrate check ./my-ts-server

mcp-migrate v0.2.0  ->  my-ts-server

No grade for this one. Found 1 TypeScript. Every rule reads it now, but
whether it gets graded is still an open decision, not a coverage gap --
see https://github.com/dheerajjha/mcp-migrate/issues/172.

  breaking    R001  server.ts:4  Mcp-Session-Id was removed from the Streamable HTTP transport.
  deprecated  R006  server.ts:1  HTTP+SSE transport is deprecated.
  deprecated  R006  server.ts:5  HTTP+SSE transport is deprecated.

3 finding(s) from the rules that do cover this language. Real, and worth
fixing -- but a letter grade would be a claim about a decision that
hasn't been made.

Whether a language every rule reads should ever get a letter is tracked as #172. Until that is resolved, a TypeScript tree gets findings but no grade, no badge, and no registry entry. Findings are real and complete; only the letter is withheld.

The exit code works on TypeScript, so this drops into CI today regardless: a breaking finding exits 1 whether or not a grade was issued. (Until v0.1.3 it always exited 2, so CI could not fail on a TypeScript regression — that was #98.)

No rule is Python-only any more. Issue #30 is closed on coverage; what remains is the grading decision in #172.

By default, check skips test code: anything under a tests/, test/, testing/, fixtures/, examples/, or docs/ directory, plus test_*.py, *_test.py, and conftest.py files. A backward-compat test that deliberately exercises a deprecated transport, or an integration test that posts a literal {"method": "tools/list"} payload, is evidence your project is well tested -- not evidence the server itself is broken. Pass --include-tests to scan those paths too.

Project config

Everything above is a flag, which means it has to be retyped on every invocation and can't be shared with a team or with CI. Put it in [tool.mcp-migrate] in pyproject.toml instead:

[tool.mcp-migrate]
skip = ["vendor/", "generated/"]
include-tests = false

[tool.mcp-migrate.rules]
R008 = "off"                                          # no reason recorded
R016 = "off -- ttl is enforced by the gateway in front of this"

No pyproject.toml? Which, given this tool targets MCP servers, is most JavaScript and TypeScript projects -- put the same keys in a standalone .mcp-migrate.toml at the project root instead:

skip = ["vendor/"]

[rules]
R008 = "off -- we propagate trace context out of band"

If a pyproject.toml exists at all, it wins, whether or not it actually sets [tool.mcp-migrate] -- the standalone file is only read for projects that have no pyproject.toml in the first place, so the two never end up silently disagreeing about which one applies.

Precedence: a flag beats config, config beats the built-in default. --include-tests on the command line always wins; config can turn it on project-wide when the flag doesn't.

A disabled rule never runs, rather than running and being discarded. It costs nothing against the grade, and it is never mistaken for a pass -- check reports how many rules config switched off (N rule(s) disabled by config, with each rule id and its reason) in every run, not behind a flag, same as suppressions. --json carries the same list under disabled_rules. fix skips the fixer for a disabled rule too.

Malformed config (bad TOML, an unrecognised rule id, rules.R008 = "sometimes") doesn't fail the run -- it falls back to defaults and reports what it couldn't understand, under config_warnings in --json output and as a config warning line in text output.

mcp-migrate fix

$ uvx mcp-migrate fix tests/fixtures/fixer_roundtrip

errors.py
--- a/errors.py
+++ b/errors.py
@@ -5,7 +5,7 @@
 
 def read_resource(handle: str) -> dict:
     if not _exists(handle):
-        return {"code": -32002, "message": "resource not found"}
+        return {"code": -32602, "message": "resource not found"}
     return {"contents": _load(handle)}
 
 
  [R017/safe] line 8: resource-not-found error code -32002 -> -32602

server.py
--- a/server.py
+++ b/server.py
@@ -12,26 +12,29 @@
 from __future__ import annotations
 
 from mcp.server import Server
-from mcp.server.sse import SseServerTransport
+from mcp.server.streamable_http import StreamableHTTPServerTransport
 from mcp.types import ServerCapabilities, Tool, ToolsCapability
 
 server = Server("fixture-server")
 
 capabilities = ServerCapabilities(
     tools=ToolsCapability(list_changed=True),
+    extensions={},
 )
 
-transport = SseServerTransport("/messages")
+# TODO(mcp-migrate): verify no constructor args were lost moving off SSE, see https://modelcontextprotocol.io/specification/draft/changelog
+transport = StreamableHTTPServerTransport()
 
 
 def _session_for(request):
-    mcp_session_id = request.headers.get("Mcp-Session-Id")
+    # TODO(mcp-migrate): replaced by an explicit handle argument, see https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2567
+    # mcp_session_id = request.headers.get("Mcp-Session-Id")
     return mcp_session_id
 
 
 @server.list_tools()
 async def list_tools() -> list[Tool]:
-    return [
+    return sorted([
         Tool(name="zeta", description="Last alphabetically."),
         Tool(name="alpha", description="First alphabetically."),
-    ]
+    ], key=lambda t: t.name)
  [R001/review] line 28: commented out Mcp-Session-Id header access, added TODO
  [R004/safe] line 35: wrapped returned tool list in sorted(key=lambda t: t.name)
  [R005/safe] line 20: added extensions={} to ServerCapabilities(...)
  [R006/review] line 15: import Streamable HTTP transport instead of SSE
  [R006/review] line 25: SseServerTransport(...) -> StreamableHTTPServerTransport(), flagged for review

2 file(s), 6 change(s): 3 safe, 3 flagged for human review
4 finding(s) still need a human after this fix -- run `mcp-migrate check tests/fixtures/fixer_roundtrip` for details.
Dry run -- nothing was written. Re-run with --write to apply.

fix runs in dry-run mode by default (that's what --dry-run names explicitly; it's implied when you pass neither flag) -- it prints the exact unified diff for every file that would change and writes nothing. Pass --write to apply it. --dry-run and --write are mutually exclusive.

Every change is tagged safe or review in the diff output:

• safe -- the fixer is certain the transformation can't change behavior (e.g. wrapping an already-unordered list literal in sorted(), or adding extensions={} where absent already meant "no extensions"). Apply these with confidence.
• review -- the fixer did the mechanical part it can be sure of (renaming an import, commenting out a now-nonexistent header read) and left a # TODO(mcp-migrate): ... exactly where a human has to finish the job, because the rest genuinely requires a judgment call this tool can't make for you (what do you name the new handle argument? what constructor args did the old transport need that the new one doesn't take?).

Pass --safe-only to apply only safe fixers, --rule R006 to restrict to one or more rules (repeatable, same flag check uses), --include-tests to also fix test/fixture paths (skipped by default, same rule as check). Only 19 of the 21 rules ship a fixer at all -- see the table below and mcp-migrate fixers. Fixers are deliberately conservative: when a fixer can't be sure a transformation is correct, it leaves the source untouched rather than guess. A wrong fix that silently corrupts your server is worse than reporting the finding and doing nothing.

Catch it before it is committed

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/dheerajjha/mcp-migrate
    rev: v0.2.0
    hooks:
      - id: mcp-migrate

A breaking finding fails the hook, which is the point. A repository with no readable source does not — the hook runs mcp-migrate-precommit, a thin wrapper that maps check's exit 2 ("could not check it") to 0. pre-commit treats every non-zero exit as a failure, so without that the hook would block every commit in a repo the tool cannot read, which is a reason the user can do nothing about. The message still prints; it just isn't fatal.

The hook scans the project, not the staged files (pass_filenames: false), and that isn't an oversight to optimise away: several rules are whole-project questions — R010 asks whether server/discover exists anywhere — and handed a partial view they fire wrongly. Cost of that choice, measured on 600 files (300 Python + 300 TypeScript): ~0.32 s.

Other commands

mcp-migrate rules                     # list every rule this version ships, with severity and spec ref
mcp-migrate fixers                    # list every fixer, with confidence (safe/review)
mcp-migrate entry --repo owner/name   # print a registry/servers/*.yaml entry for the board

Every rule

Run mcp-migrate rules to see this list for the exact version you have installed, and mcp-migrate fixers for the fixer confidence table. Full spec changelog: https://modelcontextprotocol.io/specification/draft/changelog.

Missed something and want to fill in a fixer, or just document the manual fix? See Contribute below and cookbook/ for every change that doesn't have a recipe yet.

The grade

Every finding costs points, but no single rule can sink your grade by itself: each rule's total contribution is capped, no matter how many times it fires. Real evidence for why this matters: mcp-atlassian's R003 once hit 19 times on the same false-positive pattern, for 475 raw penalty points -- over 4x what one rule is now allowed to cost.

The letter counts kinds of problem, not amount of work. Because every rule is capped, one Mcp-Session-Id read and fourteen of them across four files both score 75 and both grade C. The score moves when a different rule fires, not when the same one fires again.

That is deliberate -- repetition of one systemic issue is one issue -- but it means the grade cannot be used to size a migration. For that, read counts and findings in --json: they carry every occurrence, uncapped. Someone reasonably read the letter the other way and estimated two very different migrations as the same day of work, which is the failure this paragraph exists to prevent.

Score starts at 100 and floors at 0. The letter grade comes from the score:

Get your badge

Run the check, fix what's breaking, then generate your entry and badge in one shot:

uvx mcp-migrate entry --repo owner/name > registry/servers/name.yaml

entry refuses — writing nothing to stdout, so the redirect above leaves no file behind — if it can't read your server, or if your Python is a small minority of a repo that's mostly something else. A board entry is a claim about a whole repo, and this project would rather publish nothing than publish a grade it can't stand behind.

That prints something like:

# registry/servers/name.yaml
name: name
repo: owner/name
language: python
grade: A
score: 100
checked_with: mcp-migrate 0.1.0
spec: "2026-07-28"
status: ready
notes: >-
  Replace this line with one sentence about what your server does.

Edit the notes: line to one sentence about what your server does, then PR it into this repo (steps in CONTRIBUTING.md). CI validates the YAML against registry/schema.yaml and regenerates the board below on merge. Listing is automated and unopinionated: schema passes and the repo exists, it merges -- no maintainer reviews the server itself or argues about your grade.

Drop this in your own README once you know your grade (swap the letter and color using the table above):

[![MCP 2026-07-28](https://img.shields.io/badge/MCP%202026--07--28-A-brightgreen)](https://github.com/owner/name)

The board

16 servers checked (7x A, 7x B, 1x C, 1x D)

All of these were checked by this project, not submitted by the servers' maintainers -- so read it as a survey, not as adoption. If you maintain one of these, submit your own entry and it becomes yours.

Contribute

Three ways in, cheapest first:

• A cookbook recipe (~5 minutes). Pick an open slot in cookbook/ -- markdown, no Python, no tests. See CONTRIBUTING.md.
• A rule (~15 minutes). A Rule subclass -- a check(project) method, a regex or an AST walk, a Finding for every hit. Most rules are under 25 lines. See CONTRIBUTING.md.
• A fixer (~45 minutes). A Fixer subclass that turns one rule's finding into a text-level edit, tagged safe or review. See CONTRIBUTING.md.

The standing principle behind all three: a false positive is worse than a missed finding. When a rule can't be made precise, it ships advisory, or it doesn't ship. Reviews happen within 48 hours; one passing test is enough to merge.

Listing your own server on the board is separate from all three and takes about 60 seconds -- see CONTRIBUTING.md.

Contributors

TypeScript support exists because these people built it. The language backend shipped with two reference ports (R001 and R006). The other twelve were contributed -- in parallel, by seven people who mostly hadn't spoken to each other, each working from a separate issue.

git shortlog -sne is the authoritative list; .mailmap folds the duplicate identities so nobody is counted twice or split in half.

Working principles

A false positive is worse than a missed finding. This tool publishes public grades about other people's projects. A wrong breaking verdict costs a maintainer their badge and their trust in the tool; a missed advisory costs a little visibility and nothing else. When a detection can't be made precise it ships advisory, or it doesn't ship. Fixers follow the same rule: one that can't be certain a transformation is correct leaves the source untouched.

An unearned grade is as damaging as an undeserved finding, and harder to notice. A server that scores A because a rule couldn't see its handlers is a bug, not a pass -- see r010_server_discover_missing.py, where three registration idioms used by real servers went undetected.

No grade is published that hasn't been reproduced. Every entry in registry/servers/ comes from a scan someone ran and read, at that server's own directory. Nothing is estimated, and nothing is carried forward from a previous version of the rules.

Contributing stays cheap. One small file is the unit of contribution, and a submission that passes schema validation and points at a real repository is merged -- there's no usage bar, no curation step, and no reviewer taste test. Reviews happen within 48 hours and one passing test is enough to merge a rule.

The longer form of all of this, with the API you'd actually use, is in CONTRIBUTING.md.

License

Apache-2.0. See LICENSE.

Listing on the board is automated and unopinionated: if registry/servers/*.yaml passes schema validation and the repo it points at exists, it gets merged. No maintainer reviews the server itself.