Skip to content

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[0.3.8] - 2026-10-21

Security

  • Pin certifi>=2026.10.19 for the eMudhra / emSign CA incident (GHSA-6q48-6c7w-9gw2): this server's outbound TLS — Drupal sites, OAuth token and JWKS endpoints — verifies against certifi's bundle via httpx, and a pre-incident bundle still trusts the roots the eight compromised intermediates chain through. Fresh installs already resolve the patched bundle; the floor makes pip install -U drupal-mcp-server fix existing environments too.

[0.3.7] - 2026-10-20

Added

  • Drupal-side identity for networked deployments: DRUPAL_OAUTH_DELEGATE=1 serves ADR 0005's proxy delegation. MCP clients keep the standard flow against this server (discovery, dynamic registration, PKCE) while authorize/token requests are forwarded to Drupal's simple_oauth server, and every presented token is validated per request — locally via JWKS (no per-request hop, endpoint derived from DRUPAL_OAUTH_TOKEN_URL by the /oauth/token convention) or by RFC 7662 introspection when DRUPAL_OAUTH_INTROSPECTION_URL is set (hop per cold token, cached). A token URL that does not follow the convention, or missing upstream credentials, refuses to start with the override settings named. Optional: DRUPAL_OAUTH_AUTH_ENDPOINT, DRUPAL_OAUTH_JWKS_URI, DRUPAL_OAUTH_AUDIENCE, DRUPAL_OAUTH_REQUIRED_SCOPES (empty accepts any valid token) and MCP_BASE_URL for public redirects. The default posture is unchanged: the in-process authorization server, which needs no Drupal reachable. Pinned offline against generated keys and a stub upstream; the live simple_oauth round trip is not yet exercised (recorded gap, like the outbound client-credentials leg).

[0.3.6] - 2026-10-20

Added

  • MCP_READ_ONLY=1 exposes only read-only tools: a registrar drops every writing tool at decoration time, so it never enters the registry — tools/list cannot advertise it and a call by name fails as an unknown tool, not a refused call. The classifier is a naming contract (query prefixes/suffixes plus a vetted extras list), fail-closed so a future tool stays hidden until its name reads as a query, pinned by a census test at 185 of 414 tools. One deliberate exception: the search_ prefix heads the search_api_* mutation family, so those eight names (create/reindex/clear/enable/disable) are denied explicitly while the family's readers pass on their own suffixes.
  • MCP_RATE_LIMIT=N puts a token bucket of N requests/minute per client on every HTTP route — MCP endpoints, custom routes, and the OAuth token endpoint alike — answering 429 with Retry-After past the budget and reporting X-RateLimit-Limit/X-RateLimit-Remaining on the way through. Keyed by bearer token when the request carries one (a credential being guessed burns its own budget, not the shared address of every client behind it), else by client IP. 0 — the default — disables it. The default surface is unchanged: 414 tools without either flag.

[0.3.5] - 2026-10-20

Fixed

  • Every scaffold writer now emits files Drupal can actually load. The live sweep only proves a tool answers; offline tests over each writer found and pinned the defects it never could:
  • module_create named its controller with a literal {machine_name.capitalize()} (missing f) and reported the config schema at the module root though it writes config/schema/;
  • module_service_create pasted $this->x = $x at class scope — a PHP ParseError whenever arguments were passed — and kept dots from service ids in variable names ($entity_type.manager);
  • test_scaffold_phpunit and module_test_create wrote shell-style \$this escapes into raw PHP and pluralized the namespace (KernelTests/UnitTests), so the namespace never matched the directory the test landed in and PSR-4 discovery never loaded it;
  • module_plugin_create never closed the plugin docblock, gave formatters the widget's formElement instead of the abstract viewElements (fatal on load), skipped imports its signatures name, and joined the PHP namespace fragment into the OS path — one directory literally named Plugin\Block on POSIX, so no plugin file was ever in the PSR-4 tree;
  • the form, controller and event-subscriber writers reported file paths with no module root; EntityForm scaffolding is now refused with the reason (entity forms route via _entity_form and need a type this tool does not take), instead of shipping a form base class that cannot route;
  • module_event_subscriber_create subscribed methods it never generated — one handler stub per subscribed event now ships;
  • module_hook_implement offered hook_config_import_create, a hook Drupal has never invoked; replaced with hook_config_presave;
  • module_info_generate and module_dependencies_analyze "detected" dependencies that are not modules (form, entity, container, config) via dead string checks — both now infer real module names from Drupal\<module>\ class usage;
  • theme_create dumped library definitions into info.yml's libraries key, which takes a list of names to attach: with libraries passed the theme's info.yml was uninstallable;
  • generate_module_migration wrote a class its filename disagreed with on case (PSR-4 never autoloads it on a case-sensitive filesystem), never imported the MigrateSourcePlugin annotation it documents, and dropped migrations/, info.yml and .module inside src/Plugin/ while reporting them at the module root.
  • theme_settings_form generated a class that cannot load on any core: ConfigFactoryInterface and ContainerInterface used unimported, and a constructor override that dropped D11's required TypedConfigManagerInterface while its services.yml entry passed only @config.factory — the container could not build the form.
  • theme_debug_enable ran three config:set twig.config … calls (one malformed argument string, against a config object that does not exist, results discarded) and reported twig_debug: True regardless — on every site it ever ran against, nothing changed. It now refuses honestly with the settings.php recipe and runs nothing.
  • theme_template_override silently created a half-theme for a theme that does not exist, raised an uncaught FileNotFoundError for nested names like views/views-view-table.html.twig, and let an absolute template_name write anywhere; theme_breakpoint_create had no theme-exists guard at all.
  • theme_asset_optimize reported success: True regardless of what drush answered.
  • migrate_taxonomy's create paths were dead code: the client raises (LookupError) for a missing resource instead of returning an error dict, so a vocabulary or term absent from the target site was reported as an error and never created, on every call. Vocabulary updates also forwarded the caller's terms list as an attribute (a 400 on any real update), term payloads sent the read-only tid, and the inert source_jsonapi parameter was accepted, checked twice and never used — it is gone.
  • generate_migration_yaml accepted destination and ignored it (the plugin was hardcoded to entity:node), hand-rolled its YAML so any value containing ": " produced an unparseable document, and interpolated an unvalidated migration id into config Drupal's migrate storage must load.
  • sync_media_files returned a silent synced: 0 for a source path that does not exist, raised SameFileError when a directory was synced onto itself, and abandoned the whole run on the first unreadable file.

Testing

  • 91 offline tests across four batches (now 675): every scaffold writer and every migration tool is pinned — guards, generated YAML/PHP content, PSR-4 namespace↔directory agreement, verbatim delegation of the plan-service wrappers, and that every path a report claims exists on disk. Red run first, each failure fixed in the same commit.
  • The two intentional sweep-verdict changes are pre-recorded in both baselines: theme_debug_enable OK→REJECTED (the honest refusal) and sync_media_files OK→BLOCKED (a fresh runner has no synthesized source directory; it now says so). The other 26 tools in those two modules hold their verdicts, predicted offline with the harness's own synthesize/classify.

Documentation

  • The tool-docs status column labels offline coverage honestly (🧪 unit vs ❔ no offline tests); all 414 tools' rows regenerated — module, theme and migration writers now carry the unit badge.

[0.3.4] - 2026-10-20

Added

  • Dynamic client registration (POST /register, advertised as registration_endpoint): an MCP client can now complete the whole flow over a live socket — discovery → registration → PKCE authorization → token exchange → authenticated call. Before this, the client registry could never be populated and /token rejected every client as invalid_client, so no wire client could obtain a token at all.
  • Authorization codes expire ten minutes after issue (RFC 6749 §10.5), enforced by the token endpoint.

Fixed

  • load_authorization_code() and load_refresh_token() returned the raw dicts the tokens were stored in; the token handler reads .client_id, .expires_at, .redirect_uri and .code_challenge off the result, so every over-the-wire exchange died on AttributeError and answered 500. Both loaders now return the models the SDK validates — the same failure shape as 0.3.2's load_access_token dict fix, the third in the family.
  • Both token exchanges only accepted raw strings (str() on a pydantic model yields its repr), which worked for the unit tests and broke on the models the SDK hands over the wire; each now accepts either.

Documentation

  • ADR 0005 records where MCP OAuth tokens are validated: the served authorization server is in-process, the credentials split between the guard's intent switch and the outbound client-credentials leg, and introspection / proxy delegation / JWT verification are the recorded options for Drupal-side identity.
  • The README, configuration guide and migration guide now say what the guard actually checks (DRUPAL_OAUTH_CLIENT_ID), that MCP tokens are issued in-process rather than delegated to Drupal, and that self-service registration proves client-ness, not user identity.

Testing

  • A real-socket test drives the installed console script through discovery, dynamic registration, PKCE authorize, exchange, an authenticated call, code replay and refresh; unit tests pin scope enforcement, the outbound client-credentials grant, and the model-not-dict loaders.

[0.3.3] - 2026-10-19

Security

  • Refuse to start an http, sse or streamable-http transport unless an OAuth client (DRUPAL_OAUTH_CLIENT_ID) is set. The tool surface executes drush, composer and shell commands, so an unauthenticated listener was remote code execution for anyone who could reach the bind address; MCP_ALLOW_INSECURE_HTTP=1 opts back in deliberately.
  • MCP_HOST now defaults to 127.0.0.1 instead of 0.0.0.0: serving to the network should be a decision, not a default.

Added

  • DRUPAL_HTTP_RETRIES (default 2): idempotent GET requests retry transient failures — timeouts, connection resets, and 502/503/504 — with a short backoff, so one dropped read under host load no longer fails the tool outright. Mutations are never retried: a timed-out POST may already have landed, and re-issuing it could duplicate a node or double a deletion.
  • DRUPAL_HTTP_TIMEOUT (60) and DRUSH_TIMEOUT (120) make the previously hardcoded request budgets configurable.

Testing

  • A weekly and on-demand CI job (.github/workflows/live-sweep.yml) provisions a fresh DDEV site, bootstraps it through the MCP tools, sweeps all 414 tools and gates the result with tests/live/compare_sweeps.py: DEAD or a baseline-OK tool turning REJECTED/UPSTREAM fails the run, while timing-only slips warn. The committed ddev baseline is now produced by that job on a pristine site, so the gate compares like with like. The job's first runs surfaced and fixed eight setup defects — swallowed ddev returncodes, an install gate that trusted the public URL and could re-run site:install on a live site, a nested .ddev/config.yaml, a PSR-4 case bug, a missing commerce_shipping require, and missing layout/paragraph fixtures among them.
  • CI matrix extended to Python 3.13 and 3.14, and all GitHub Actions bumped off the deprecated Node 20 runtime.

Documentation

  • docs/agents.md and both READMEs referenced content_list, a tool that does not exist in the package — the documented spot check failed with "no such tool". Examples now use list_vocabularies and inventory_content_types, both verified against a live site.
  • Document the retry/timeout settings and the HTTP transport security posture ("Serving over HTTP" in the configuration guide).

[0.3.2] - 2026-10-18

Fixed

  • ddev_start, ddev_stop, ddev_exec, lando_start, lando_stop and lando_exec accepted project_name and then dropped it, so every call ran in the MCP server's own working directory — ddev_start against another site started the wrong project or failed. The parameter is now forwarded as the subprocess working directory, the directory semantics are documented, and a path that does not exist returns a clean error.
  • _path_map() cached an empty mapping when /jsonapi was unreachable (JSON:API not enabled yet, or the server restarting), permanently degrading every bare-type lookup onto the shorthand path that core does not serve. Failures are now left uncached, so the next call recovers as soon as the resource index answers.
  • lando_setup wrote a top-level webroot: key, which Lando 3 drops silently — the generated site then 404'd on every request. The docroot is now nested under config:, where Lando reads it.

Documentation

  • Add a generated catalog of all 414 tools (docs/tools/all.md): one page of per-module tables with each tool's description, drafted use case, example call, and per-environment live-verification badges (DDEV and Lando sweeps) plus unit-test coverage. scripts/generate_tool_docs.py builds it from the tool source, the hand-maintained docs/tools/_tool_overrides.yaml and the sweep results, and CI fails when the page drifts from the code.

[0.3.1] - 2026-10-17

Fixed

  • Make uvx drupal-mcp-server actually run the server. The package shipped only a drupal-mcp executable, and current uv refuses to guess when the tool name differs from the package name ("Use uvx --from drupal-mcp-server drupal-mcp instead"), so the documented no-install command — including the PATH fallback the installer writes into client configs — failed to start. A drupal-mcp-server console-script alias now runs the same entry point, and the installer's fallback, help text, README and AI-agents guide name the unambiguous uvx --from drupal-mcp-server drupal-mcp form, which works on every uv version.

[0.3.0] - 2026-10-17

Added

  • Wave 5 of the drupal-scout-mcp port — 2 local security tools (414 total, 41 modules): code_security_scan runs scout's zero-dependency pattern engine (XSS, SQL injection, command injection, path traversal, hardcoded secrets, access control, CSRF, deprecated API — 8 families, category/severity filters, per-finding file/line/ snippet/fix recommendation) with its Drupal-aware false-positive table (EntityQuery conditions, stream wrappers, escapeshellarg, example secrets, comment lines), so a site without PHPStan gets concrete findings; counts stay complete when max_findings truncates the list. route_access_audit is the routing half of scout's anonymous-exploit scan done properly per route: every *.routing.yml route classified as anonymous / permission / role / entity / custom / a named underscore requirement / an alias (whose definition may live in a PHP route provider this YAML pass cannot read) / denied / missing, with its controller mapped to the declaring file. Targets confine to DRUPAL_ROOT and prune vendor, tests, and contrib from the default scan; symlinks are skipped. Two accuracy corrections over scout's registry: t( is a live D8+ function (scout flagged every translation call as a Drupal 7 leftover) and scout's CSRF $request pattern was escaped so it could never match PHP. Scout's verify_vulnerability was not ported — it returns static advice rather than site-specific data — and the optional tree-sitter AST layer is skipped: the pattern pass carries the signal. Live testing on D11.4.10's core node module (115 files) caught five accuracy bugs before release: JS template literals matched the PHP backtick-shell pattern, Drupal's {table} / ->...Table() placeholder concatenation read as SQL injection, unrecognized underscore requirements and alias-only entries were both reported as missing access (node went from five "missing" routes to zero real ones), and a zero-file scan read as a clean bill of health.
  • Wave 4 of the drupal-scout-mcp port — 8 read-only theme analysis tools (412 total, 41 modules): theme_active (default/admin/enabled themes from system.theme + pm:list), theme_info (pm:list metadata merged with a safe YAML parse of the theme's info.yml — regions in declaration order, libraries, dependencies, and the base-theme chain followed upward), theme_regions (regions with their block placements and counts, empty regions visible, placements in regions the theme no longer defines flagged), theme_blocks (every block.block.* placement with theme/region/search filters — search matches labels and plugins, folding scout's separate search_blocks into one tool), theme_templates (pruned *.html.twig walk with sizes), theme_template_overrides (one docroot-wide index splits templates that shadow same-named core/module files from theme originals), template_suggestions (the Twig suggestion chain per entity type, each marked with whether the default theme already ships that file), and view_template_info (the chain for a view display, looked up in the JSON:API view collection by its machine id — Views config names differ between Drupal generations, view.* on older sites and views.view.* on Drupal 11.4+, so the entity collection is the stable source — with an unknown display_id erroring with the display list and an unknown view listing the ones the site has). Every file path resolves inside DRUPAL_ROOT before a read — symlinks and .. segments that leave the docroot are rejected — and no caller input reaches a shell; theme blocks come from JSON:API rather than scout's interpolated drush ev
  • The shared JSON:API helpers (jsonapi_collect, jsonapi_attrs, jsonapi_bundle, jsonapi_rel_ids), flag_label, and the export path/write helpers moved from the exports module into _utils so both waves page collections and confine destinations the same way
  • Wave 3 of the drupal-scout-mcp port — 4 export tools (404 total, 40 modules): export_nodes, export_users, export_media, export_taxonomy write CSV (UTF-8 with BOM) or Excel files (openpyxl, new runtime dependency) and return only the path, row/column counts, size, and a three-row preview. Destination paths are confined to the Drupal root, the system temp directory, or the home directory; data is paged per bundle over JSON:API (never through the shared merged offset, which skips rows), and taxonomy usage counts come from one validated GROUP BY query per field storage targeting taxonomy terms, with unreadable tables reported in a note instead of failing the export (and a safe_to_delete verdict withheld entirely whenever the counts are incomplete). Entries with no internal id — Drupal's uid-0 anonymous user serves only a display_name — are skipped and counted as skipped_without_id, and an unreadable status stays empty in the CSV rather than being presented as "blocked"
  • sql_output/sql_rows moved from the analysis module into _utils so the export tools parse drush sql:query text the same way
  • Wave 2 of the drupal-scout-mcp port — 8 tools (400 total, 39 modules):
  • Status section (4): status_report (drush core:requirements, severity filter), database_updates (read-only pending DB updates), available_updates (composer outdated with dev filtering and a security-only mode intersected with composer audit), health_check (Drush, database, JSON:API, and module probes with a fix per failure)
  • Analysis: find_entity_references (bundles whose reference fields target an entity type, via JSON:API field config), search_entities (escaped LIKE search across six entity tables), find_entity_by_path (route or path-alias resolution to an entity)
  • Config: config_status (read-only active vs sync comparison)
  • run_composer gained a tail argument so tools that parse full JSON documents (composer outdated) are not truncated

[0.2.0] - 2026-10-16

Security

  • Fix command injection vulnerabilities in code audit and security audit tools
  • Fix SQL injection in database query tools
  • Fix path traversal in media upload tool
  • Add input validation for URLs, queue names, and file paths
  • Add authentication check to /mcp/tools/list endpoint

Added

  • Add drupal-mcp install, a self-installer that registers the server with local MCP clients (Claude Code, Claude Desktop, Cursor, Codex CLI, OpenCode, Windsurf, Pi, Hermes Agent, OpenClaw), with --print/--json for scripted setup and an AI Coding Agents guide
  • Add module_issues_search, module_popular, and module_recommendation (client-side issue-queue relevance scoring, usage-ranked module lists by category, local-index + drupal.org recommendations), plus the DRUPAL_ROOT and DRUPAL_MODULES_PATH settings for local codebase analysis
  • Add a local codebase index with 12 codebase_* tools: functionality search, module inventory/descriptions, unused and redundant module detection, hook implementation lookup, and safe module file/tree/function reading (all reads confined to DRUPAL_ROOT)
  • Add moderation_transition_delete, giving moderation_transition_create the counterpart it never had. The tool count is now 367
  • Add Lando support to SiteManager
  • Add subprocess timeout to DrushClient
  • Add OAuth token fetching to AuthManager
  • Create shared tools/_utils.py with ensure_dict, validate_pagination, build_pagination_params
  • Add pagination validation to list tools
  • Add input validation to accessibility and performance tools
  • Add LICENSE, CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md
  • Add GitHub issue and PR templates

Changed

  • Correct what the sweep asks for, not just how it reads the answer. Every parameter ending in id had been answered with the fixture node's id, so view_id, menu_id, webform_id and eighty-odd others were all sent one node number and answered a lookup miss: 84 of 174 refusals in the run before this were the harness asking the wrong question, not tools refusing, and a miss says nothing about whether a tool works. Of 367 tools, 192 declare no id-shaped parameter, 93 read-only ones now get a real id for their own resource — with the tool's own name deciding where one parameter covers several, since config_id is a field config, a metatag config or a search index depending on who is asking — and 82 mutating ones deliberately keep junk ids, because views_update with a real view id would save the view back with the harness's synthesised arguments and every later views_execute would then run against a view this run broke. The remaining 8 read-only tools keep junk because the site has no such entity to name. The fixture node also re-checks and re-provisions itself: delete_node runs mid-sweep and took the node with it, which alone 404'd revision_list, revision_get, revision_revert, revision_diff, translation_status, translation_add, moderation_transition_execute, layout_builder_override_remove and accessibility_aria_validate
  • Keep feeds_log_list an explicit refusal rather than implement it. Feeds 3.x has no log table, no feeds* service and no log entity to read, so there is nothing behind it to query; the tool names imports_import_status, which does answer, instead of returning an empty list that would read as "no log entries"
  • Definitive full re-sweep (OK 195 / REJECTED 156 / BLOCKED 7 / UPSTREAM 4 / DEAD 0 / SKIPPED 5, 367 total); no tool that used to work regressed and nothing is dead. Harness follow-up fixes (entity_id and revision-id handling) pushed six more tools past their lookup so they reach their own logic, where they report honestly.

  • Re-sweep all 366 tools against the provisioned site. BLOCKED fell from 71 to 5, OK rose from 156 to 181, and nothing is dead. 25 tools that could not be reached at all now run; 41 reach Drupal and are turned away with a reason instead of stopping at a 404. The five still BLOCKED are two feeds_feed tools needing a feed type with a feeds_item field, two moderation_state tools for a resource Drupal never exposes, and split_config needing a ./config/sync directory. The one UPSTREAM is revision_delete, since JSON:API has no revision collection endpoint

  • Correct the sweep's own accounting. Six tools — translation_add, translation_update, the three interface_translation_* tools and config_translation_add — were reported OK on every run while never being called once. When arguments fail FastMCP's validation the answer is prose rather than JSON, and nothing in the classifier matched it, so the fall-through said pass; the server logged missing_argument for each and the report showed a green result. A refused call now counts as refused, through the SDK's own isError signal rather than a guess at wording. The sweep also stopped filling langcode with a junk value (JSON:API write-protects that field on a multilingual site whatever it says, so create_term answered 403 about a permission instead of running), started giving ddev_exec a command the container has, gave drush_watchdog the enum and the number drush wants, and names a queue that exists
  • Provision the live test site with the modules the tools assume are there: media, comment, contact, language, config_translation, workflows, content_moderation and commerce plus its submodules from core and contrib, and token, pathauto, redirect, metatag, paragraphs, search_api, facets, webform, feeds and graphql from drupal.org, with phpstan, drupal-check, phpcodesniffer and coder for the code-quality tools. Drupal's entity types went from 26 to 78 and the JSON:API index from 27 resources to 76, which turned most of the sweep's BLOCKED bucket into a real verdict. The site is also multilingual now -- English and French -- because a single-language site cannot exercise translation behaviour, and that is what surfaced the langcode bug below
  • Add three live suites. tests/live/install_modules.py provisions the modules the sweep needs using the server's own composer_require and drush_command tools, so bootstrapping the site is also a test of them; resource_check.py asks Drupal for the entity types it knows and flags any resource name our tools ask for that cannot exist, which is how the views_view and moderation_state bugs were found; mcp_call.py is a one-call driver for operating the site through the server instead of a shell

  • Gate CI coverage on the code testable offline (non-tool core ≥70%) instead of the total: 86% of statements live in the tool modules, which only a live sweep can exercise, so a single total threshold is unreachable by construction. coverage.xml is now actually generated, which the Codecov step already required with fail_ci_if_error: true

  • Extend layout_builder_override_remove with an optional entity_id, since an override belongs to an entity and the old signature could never name one
  • Document what live testing found, and how a failure arrives, in tests/live/README.md and the README's Development section
  • Extract _ensure_dict from 35+ tool modules into shared utility
  • Remove duplicate service implementations (ai.py, migration.py, site.py)
  • Clean up unused imports across all tool modules
  • Update ARTIFACT.md to reflect actual project state

Fixed

  • Make an issued OAuth token worth issuing. HTTP transport with OAuth enabled refused everyone: load_access_token returned the dict the token was stored in where the declared type is FastMCP's AccessToken, so the first thing the bearer middleware did with a good token -- auth_info.expires_at -- raised AttributeError: 'dict' object has no attribute 'expires_at' and the request was refused exactly as a forged one would be. That made the defect invisible by construction, because the fail-closed path and the fail-for-everyone path are the same 401: the only OAuth behaviour anyone had ever observed was the reassuring one, and the 401 that read as security working was also masking an unusable transport. expires_at was stored as a float from time.time() while AccessToken.expires_at is an int, so pydantic would have refused it regardless. Verified three ways -- unit tests over authorize/exchange/load, an ASGI test against FastMCP's real HTTP app (401 unauthenticated, 200 for the issued token, 401 for a forged one), and a replay of the old implementation that reproduces the AttributeError
  • Implement twelve tools that refused or had never been written. The five moderation_transition_* tools all read $wf->get("transitions"), a key the workflow entity has never had — transitions live on Workflow::getTypePlugin()->getTransitions() — so list answered count: 0 against a five-transition workflow, which reads as a pass because an empty list is not an error, get reported "No transition with id 7" for a transition present by name, and execute, create and the new delete were unreachable behind the same mistake. The four interface_translation_* tools read TranslationString::getLanguage() and getTranslation(), methods that class does not have; they now go through locale.storage, and a live round trip verified the list (13666 French rows, with search and context filters), add, update and delete, including a source string carrying a non-ASCII character and a $. Writing a translation also dispatches LocaleEvents::SAVE_TRANSLATION through _locale_refresh_translations(), which invalidates the locale cache tag every LocaleLookup row depends on — invalidateByPrefix and cache_clear_all do not exist in Drupal 11. The three webform_handler_* tools answered absent_entity because they had never been written; they now use plugin.manager.webform.handler with Webform::addWebformHandler / updateWebformHandler / deleteWebformHandler, take webform_id to disambiguate a handler id that is unique per webform rather than site-wide, replace configuration keys whole so omitted ones keep their current values, and locate handlers through getHandlers()->getInstanceIds() instead of Webform::getHandler(), which throws PluginNotFoundException for an unseen id rather than returning NULL. A contact form started this with two handlers and ended with two after a full create/update/delete — the point of a round trip
  • Hand structured values to php:eval as base64 rather than as JSON, which is not PHP: it emits \uXXXX and PHP has no such escape, so non-ASCII mangles silently; $ interpolates into the code; and {"k":1} is a parse error. _utils.php_text() and php_json() carry a string as base64_decode(...) and an array as json_decode(base64_decode(...), TRUE)
  • Stop chasing the login redirect to a URL the site invented. Drupal answers a good login with 303 to the account page, built from the scheme and host it saw the request arrive on, so when the site reported itself as plain http while the client was on HTTPS the redirect was fetched as HTTP against a TLS port, nginx answered 400, and the session cookie that same response had just issued was discarded — leaving only a startup warning. Every write then fell back to basic auth, which Drupal's mutating endpoints reject, on any site whose self-issued URL disagrees with the client's behind a proxy or a misconfigured base_url. The POST is no longer followed, because only its status and its cookie were ever wanted; bad credentials still fail, since Drupal refuses them with 200 and no cookie
  • Make the sweep follow the site's links on the origin it was given. Drupal answers http:// when reached through a port-mapped HTTPS endpoint, and the harness followed those verbatim — plain HTTP to a TLS listener, which nginx refuses with 400 — so a whole run died before calling a single tool. filter_check.py hardcoded its base URL while every other live script honours DRUPAL_BASE_URL, so it could not be pointed at all
  • Ask whether a node exists in a way JSON:API can answer. The check built the resource name into the path (/jsonapi/node--article/7, where the resource lives at /jsonapi/node/article) and asked for an internal id where rows are addressed by uuid, so it 404'd for nodes that exist and reported every fixture as gone — which would have had the harness re-provisioning on every check instead of only when delete_node removed one
  • Make three tools answer with the reason instead of a bare error. views_execute could never have run: it asked for a views.views service that does not exist (the executable is built from the view's config entity by views.executable), read the row's entity through a magic property that isset() always reports as absent, and took no care over the display. It now loads the view, checks the display against the ones the view actually has -- reporting the rest when the caller's is wrong -- and reads _entity. graphql_query and graphql_schema_get returned a raw 404 whenever the site had no GraphQL server, which is the state a plain install leaves it in, since the module ships no default; they now say that, and name the config entity and the permission that fix it. paragraphs_item_list returned a bare 404 for a bundle that does not exist, which reads as "paragraphs are missing" rather than "that type is not configured here"; it now lists the types the site does have
  • Fix eleven hard-coded JSON:API filters that Drupal rejected outright, each returning an empty result on every call on every site. commerce_product, commerce_promotion and commerce_shipping_method filtered on store where the field is the plural stores, and commerce_order on store where it is the store_id reference; all four needed the .id specifier that entity references require. webform_submission filtered on webform where the field is webform_id, also a reference. paragraphs_item_list filtered on paragraphs_type, which is not a field at all -- a paragraph's type is its bundle, so it now addresses paragraph--{type}. pathauto_pattern_list filtered on entity_type and bundle, which live in a pattern's id as {entity_type}--{bundle}, so it now filters on id when it has both. comments_thread_list filtered on pid for top-level comments, but pid is a reference to the parent comment, so it 400'd -- and pid.id = 0 would have returned nothing, since no comment has id 0; the whole thread is returned now, with each row's pid available for building the tree. facets_facet_list filtered on index, but a facet attaches to a facet source rather than to an index, so it lists facets and reports their facet_source_id
  • Stop sending langcode: "en" in create payloads. Four tools did this. It is harmless on a single-language site, but once a second language exists the field is writable only with translation permission, so every create failed 403 "the current user is not allowed to POST the selected field (langcode)" -- including an ordinary authenticated user creating a term in the site's own language. The field is now sent only when the caller asks for a language
  • Fix drush_command option rendering, which made every confirmation-answering call fail. Boolean options were emitted as --key=value, but drush flags take no value, so options={"yes": True} -- the documented way to install a module -- was rejected with "The --yes option does not accept a value" and nothing was installed while the tool reported success. A False was worse: it emitted the bare flag, so asking not to confirm silently suppressed confirmation on destructive commands. Zero-valued options were also mangled, --count=0 becoming a bare --count. Booleans are now bare flags, False omits the option, and everything else keeps its value
  • Return an error from drush_command when Drush exits non-zero. It handed back the raw result with no error key, so the generic escape hatch -- the tool an agent reaches for first -- still presented a failed command as an empty success, even though ensure_dict already fixed this for the 224 specialised call sites
  • Split several packages in composer_require. They arrived as one argv token, so composer read drupal/pathauto as a version constraint on drupal/token and refused the whole request with "Could not parse version constraint"
  • Ask for view, not views_view: the entity type has been view since Drupal 8, so all six views tools 404'd on a site with Views installed and healthy, and read as "module not installed"
  • Ask for content_moderation_state, not moderation_state, in the moderation tools. The node field is still called moderation_state and is unchanged
  • Ask for facets_facet, not search_api_facet, in the facet tools
  • Rewrite config_translation_list and config_translation_add, which targeted three mechanisms that do not exist between them: a config_translation JSON:API resource, a config:translation:add Drush command (Drush has no translation command category at all), and $entity->addTranslation() on a config entity, which implements no TranslatableInterface. The list reports the mappers this site has. The add resolves {entity_type}/{entity_id} through the config translation mapper to the config object the entity is stored as -- node_type and article become node.type.article -- and writes a language config override for the target language, which is where Drupal keeps config translations. Verified live: the French name and description land on the override while the base config keeps its own
  • Drop a dead get_resource("jsonapi") call from nextjs_helpers_generate. It requested /jsonapi/jsonapi, a path that does not exist, and the 404 aborted the whole tool for a value the generator never used

  • Implement the six layout_builder_* tools against the real API; they asked JSON:API for a resource that has never existed and returned an opaque 404 on every call, so reading, writing and reverting layouts now go through Drush to entity_view_display and OverridesSectionStorage

  • Fix DrushClient._flatten shlex-splitting list arguments, which made db_tables, db_query, db_count, drush_php_eval and every ev tool fail with "Too many arguments"
  • Report a failed Drush run as an error: ensure_dict passed {returncode: 1, data: {}, stderr} through untouched, byte for byte an empty success, so 224 call sites presented broken commands as empty results
  • Return real tool schemas from /mcp/tools/list, which reported input_schema: {} for all 366 tools
  • Fix cron_list (called Drupal 7's drupal_get_cron_jobs(), removed in Drupal 8), queue_create (a shell escape in PHP that never passes through a shell, so PHP died with a ParseError) and queue_delete (a second argument on all three paths when queue:delete accepts exactly one)
  • Route queue_create and queue_delete around drush queue:delete, which resolves only registered queue services and answered "Queue not found" for a queue this server had just created
  • Fix permission_audit and menu_link_list filtering on fields Drupal does not expose (filter[name] on user_role, filter[menu] on menu_link_content), a 400 on every call on every site
  • Validate whole statements in db_query; a SELECT 1; DROP TABLE x passed the startswith("SELECT") guard and dropped the table
  • Stop stacking log handlers on repeated setup_logging calls

Removed

  • Remove dead register_all_tools function from tools/init.py
  • Remove duplicate _find_tool_path definition in security_audit.py

[0.1.0] - 2026-10-13

Added

  • Initial release with 35+ MCP tools across all major Drupal categories
  • FastMCP server with stdio and HTTP transports
  • JSON:API and GraphQL clients with authentication
  • Drush command execution via subprocess
  • Multi-provider AI support (Ollama, OpenAI, Anthropic, NVIDIA NIM)
  • Content management, user management, taxonomy, views, paragraphs tools
  • DevOps tools for DDEV/Lando, Composer, database queries
  • Migration tools for D7 to D10 upgrades
  • Security audit and code quality tools