Claude Code's Leak: Every Hardcoded Vendor and Tool

A note on this analysis. In March 2026, Claude Code's full TypeScript source was extracted from publicly available npm source maps. All code referenced here remains the intellectual property of Anthropic. This article is published for educational and research purposes. We do not host or distribute the source code. Our interest is narrow: understanding how a major AI coding agent treats the developer tool ecosystem, so that vendors building for this ecosystem can make informed decisions.

Claude Code's full TypeScript source leaked on GitHub in March 2026. We searched it for vendor names to see if any companies receive special handling. They do. We found hardcoded vendor references in seven different parts of the codebase:

Why does this matter?

Each system gives included providers a concrete advantage. Your MCP output renders cleanly while competitors dump raw JSON. Your docs are readable by the agent without user intervention. Your platform shows up by name in Anthropic's analytics. None of these change which tools Claude recommends. But they change how your product is treated in the terminal millions of developers use daily.

What we found: The collapse allowlist contains 495 typed entries across two Sets, which become 489 unique normalized tool names at runtime. The source comments organize them into 40 label blocks. classifyForCollapse.ts explicitly tags 6 claude.ai-hosted labels, while a separate main.tsx comment suggests Datadog may be an ambiguous seventh. WebFetch contains 89 preapproved host entries (88 unique host/path scopes). env.ts returns 29 explicit deployment labels before unknown fallbacks. The secret scanner ships 36 high-confidence rules across 23 credential families. The API logging layer fingerprints 7 gateways. The tip registry contains 2 plugin tips, only 1 for a third-party vendor.

1. MCP UI allowlist: 489 unique tool names across 40 grouped labels

This is the most explicit UI-specific integration layer in the leaked client. classifyForCollapse.ts contains two hardcoded Set objects, SEARCH_TOOLS and READ_TOOLS. There are 495 typed entries across those two literals, but because a few names repeat, the runtime allowlist resolves to 489 unique normalized tool names. When Claude Code calls a listed tool, the output collapses into a compact summary. When it calls an unlisted tool, you get the generic MCP renderer.

// classifyForCollapse.ts
function classifyMcpToolForCollapse(
  _serverName: string,   // ignored — tool names are stable
  toolName: string,
): { isSearch: boolean; isRead: boolean } {
  const normalized = normalize(toolName)  // camelCase → snake_case
  return {
    isSearch: SEARCH_TOOLS.has(normalized),
    isRead: READ_TOOLS.has(normalized),
  }
}
// Unknown tools don't collapse (conservative default)

The function keys on tool name alone, not server name. The source comment: “Tool names are stable across installs (even when the server name varies, e.g., ‘slack’ vs ‘claude_ai_Slack’).” Unknown tools do not collapse. The allowlist is opt-in, not opt-out.

The file comments group those tool names into 40 label blocks. That is not a clean vendor census: some labels explicitly combine multiple implementations, such as hosted + SDK Slack or hosted + community PubMed. The badge counts below are typed entries within each comment block, not globally unique tools.

Explicitly tagged claude.ai-hosted labels

These six labels are explicitly annotated as claude.ai-hosted inside classifyForCollapse.ts. The transport path itself goes through mcp-proxy.anthropic.com. Users connect through OAuth on claude.ai rather than local MCP configuration. More on the transport in section 2.

search_articlesfind_related_articleslookup_article_by_citationsearch_paperssearch_pubmedsearch_pubmed_key_wordssearch_pubmed_advancedpubmed_searchpubmed_mesh_lookupget_article_metadataget_full_text_articleconvert_article_idsget_copyright_statusdownload_paperlist_papersread_paperget_paper_fulltextget_pubmed_article_metadatadownload_pubmed_pdfpubmed_fetchpubmed_pmc_fetchpubmed_spellpubmed_citepubmed_related

Slack

13 entries · hosted + SDK label

slack_search_publicslack_search_public_and_privateslack_search_channelsslack_search_usersslack_read_channelslack_read_threadslack_read_canvasslack_read_user_profileslack_list_channelsslack_get_channel_historyslack_get_thread_repliesslack_get_usersslack_get_user_profile

gmail_search_messagesgmail_get_profilegmail_read_messagegmail_read_threadgmail_list_draftsgmail_list_labels

gcal_find_my_free_timegcal_find_meeting_timesgcal_find_user_emailsgcal_list_calendarsgcal_list_eventsgcal_get_event

bigquery_querybigquery_schemalist_dataset_idslist_table_idsget_dataset_infoget_table_info

google_drive_searchgoogle_drive_fetchgoogle_drive_export

A separate comment in main.tsxalso lists Datadog among “claude.ai proxy servers.” Because classifyForCollapse.ts does not tag Datadog that way, we treat Datadog as an explicit source conflict rather than a confirmed hosted label.

Vendor-operated or vendor-associated labels

These are the larger vendor-operated or vendor-associated comment blocks in the file, sorted roughly by typed entry count.

GitHub

56 tools · github/github-mcp-server

search_codesearch_repositoriessearch_issuessearch_pull_requestssearch_orgssearch_usersget_meget_team_membersget_teamsget_commitget_file_contentsget_repository_treelist_brancheslist_commitslist_releaseslist_tagsget_latest_releaseget_release_by_tagget_taglist_issuesissue_readlist_issue_typesget_labellist_labelpull_request_readget_gistlist_gistslist_notificationsget_notification_detailsprojects_listprojects_getactions_getactions_listget_job_logsget_code_scanning_alertlist_code_scanning_alertsget_dependabot_alertlist_dependabot_alertsget_secret_scanning_alertlist_secret_scanning_alertsget_global_security_advisorylist_global_security_advisorieslist_org_repository_security_advisorieslist_repository_security_advisoriesget_discussionget_discussion_commentslist_discussion_categorieslist_discussionslist_starred_repositoriesget_issueget_pull_requestlist_pull_requestsget_pull_request_filesget_pull_request_statusget_pull_request_commentsget_pull_request_reviews

Grafana

38 entries · grafana/mcp-grafana

search_dashboardssearch_folderslist_teamslist_users_by_orgget_dashboard_by_uidget_dashboard_summaryget_dashboard_propertyget_dashboard_panel_queriesrun_panel_querylist_datasourcesget_datasourceget_query_examplesquery_prometheusquery_prometheus_histogramlist_prometheus_metric_metadatalist_prometheus_metric_nameslist_prometheus_label_nameslist_prometheus_label_valuesquery_loki_logsquery_loki_statsquery_loki_patternslist_loki_label_nameslist_loki_label_valueslist_incidentsget_incidentlist_sift_investigationsget_sift_investigationget_sift_analysislist_oncall_schedulesget_oncall_shiftget_current_oncall_userslist_oncall_teamslist_oncall_userslist_alert_groupsget_alert_groupget_annotationsget_annotation_tagsget_panel_image

Datadog

30 tools · mcp.datadoghq.com

search_logssearch_spanssearch_rum_eventssearch_audit_logssearch_monitorssearch_monitor_groupsfind_slow_spansfind_monitors_matching_patternaggregate_logslist_spansaggregate_spansanalyze_tracetrace_critical_pathquery_metricsaggregate_rum_eventslist_rum_metricsget_rum_metriclist_monitorsget_monitorcheck_can_delete_monitorvalidate_monitorvalidate_existing_monitorlist_dashboardsget_dashboardquery_dashboard_widgetlist_notebooksget_notebookquery_notebook_cellget_profiling_metricscompare_profiling_metrics

Asana

29 tools · mcp.asana.com

asana_search_tasksasana_typeahead_searchasana_get_attachmentasana_get_attachments_for_objectasana_get_goalasana_get_goalsasana_get_parent_goals_for_goalasana_get_portfolioasana_get_portfoliosasana_get_items_for_portfolioasana_get_projectasana_get_projectsasana_get_project_sectionsasana_get_project_statusasana_get_project_statusesasana_get_project_task_countsasana_get_projects_for_teamasana_get_projects_for_workspaceasana_get_taskasana_get_tasksasana_get_stories_for_taskasana_get_teams_for_workspaceasana_get_teams_for_userasana_get_team_usersasana_get_time_periodasana_get_time_periodsasana_get_userasana_get_workspace_usersasana_list_workspaces

PagerDuty

28 entries · PagerDuty/pagerduty-mcp-server

list_incidentsget_incidentget_outlier_incidentget_past_incidentsget_related_incidentslist_incident_noteslist_incident_workflowsget_incident_workflowlist_servicesget_servicelist_team_membersget_user_datalist_schedulesget_schedulelist_schedule_userslist_oncallslist_log_entriesget_log_entrylist_escalation_policiesget_escalation_policylist_event_orchestrationsget_event_orchestrationlist_status_pagesget_status_page_postlist_alerts_from_incidentget_alert_from_incidentlist_change_eventsget_change_event

Community Atlassian

20 entries · sooperset/mcp-atlassian

confluence_searchjira_searchjira_search_fieldsconfluence_get_pageconfluence_get_page_childrenconfluence_get_commentsconfluence_get_labelsjira_get_issuejira_get_transitionsjira_get_worklogjira_get_agile_boardsjira_get_board_issuesjira_get_sprints_from_boardjira_get_sprint_issuesjira_get_link_typesjira_download_attachmentsjira_batch_get_changelogsjira_get_user_profilejira_get_project_issuesjira_get_project_versions

Sentry

18 entries · getsentry/sentry-mcp

search_docssearch_eventssearch_issue_eventsfind_organizationsfind_teamsfind_projectsfind_releasesfind_dsnswhoamiget_issue_detailsget_issue_tag_valuesget_trace_detailsget_event_attachmentget_docget_sentry_resourcelist_eventslist_issue_eventsget_sentry_issue

Atlassian / Jira

17 entries · mcp.atlassian.com

search_jira_issues_using_jqlsearch_confluence_using_cqllookup_jira_account_idatlassian_user_infoget_accessible_atlassian_resourcesget_visible_jira_projectsget_jira_project_issue_types_metadataget_jira_issueget_transitions_for_jira_issueget_jira_issue_remote_issue_linksget_confluence_spacesget_confluence_pageget_pages_in_confluence_spaceget_confluence_page_ancestorsget_confluence_page_descendantsget_confluence_page_footer_commentsget_confluence_page_inline_comments

kubectl_getkubectl_describekubectl_logskubectl_contextexplain_resourcelist_api_resourcesnamespaces_listnodes_lognodes_toppods_getpods_listpods_list_in_namespacepods_logpods_topresources_getresources_list

Supabase

15 tools · supabase-community

search_docslist_organizationsget_organizationget_costlist_extensionslist_migrationsget_logsget_advisorsget_project_urlget_publishable_keysgenerate_typescript_typeslist_edge_functionsget_edge_functionlist_storage_bucketsget_storage_config

Linear

16 entries · mcp.linear.app

search_documentationlist_commentslist_cyclesget_documentlist_documentslist_issue_statusesget_issue_statuslist_my_issueslist_issue_labelslist_projectsget_projectlist_project_labelslist_teamsget_teamlist_usersget_user

Todoist

16 entries · Doist

find_tasksfind_tasks_by_datefind_completed_tasksfind_projectsfind_sectionsfind_commentsfind_project_collaboratorsfind_activityfind_labelsfind_filtersget_productivity_statsget_overviewfetch_objectuser_infolist_workspacesview_attachment

findsearch_knowledgelist_databaseslist_collectionscollection_indexescollection_schemacollection_storage_sizedb_statsexplainmongodb_logsaggregatecountexport

AWS

13 entries · awslabs/mcp

search_documentationsearch_catalogget_available_servicesread_documentationread_sectionsrecommendanalyze_log_groupanalyze_metricdescribe_log_groupsget_active_alarmsget_alarm_historyget_metric_dataget_metric_metadata

Stripe

13 entries · stripe/agent-toolkit

search_stripe_resourcessearch_stripe_documentationget_stripe_account_inforetrieve_balancelist_customerslist_productslist_priceslist_invoiceslist_payment_intentslist_subscriptionslist_couponslist_disputesfetch_stripe_resources

Playwright

13 tools · microsoft

browser_console_messagesbrowser_network_requestsbrowser_take_screenshotbrowser_snapshotbrowser_get_configbrowser_route_listbrowser_cookie_listbrowser_cookie_getbrowser_localstorage_listbrowser_localstorage_getbrowser_sessionstorage_listbrowser_sessionstorage_getbrowser_storage_state

Notion

5 tools · mcp.notion.com

searchfetchget_commentsget_usersget_self

Elasticsearch

4 tools · elastic

list_indicesget_mappingsesqlget_shards

MCP SDK reference labels

These ship under the @modelcontextprotocol namespace. Slack and Google Drive already appear above because the source comments join their hosted and SDK variants under one tool family.

Slack

13 entries · shared hosted + SDK label

Filesystem

10 tools · @mcp/server-filesystem

search_filesread_fileread_text_fileread_media_fileread_multiple_fileslist_directorylist_directory_with_sizesdirectory_treeget_file_infolist_allowed_directories

Git

7 tools · mcp-server-git

git_statusgit_diffgit_diff_unstagedgit_diff_stagedgit_loggit_showgit_branch

Memory

3 tools · @mcp/server-memory

search_nodesread_graphopen_nodes

Google Drive

3 tools · @mcp/server-gdrive

google_drive_searchgoogle_drive_fetchgoogle_drive_export

SQLite

3 tools · @mcp/server-sqlite

read_querylist_tablesdescribe_table

Postgres

1 tool · @mcp/server-postgres

query

Puppeteer

1 tool · @mcp/server-puppeteer

puppeteer_screenshot

Search & crawl labels

In the leaked build, the hardcoded search-provider labels are Exa, Firecrawl, Perplexity, Tavily, and Brave Search.

web_search_exaweb_search_advanced_exapeople_search_exalinkedin_search_exadeep_search_exaget_code_context_exacompany_research_exacrawling_exadeep_researcher_check

firecrawl_searchfirecrawl_scrapefirecrawl_mapfirecrawl_crawlfirecrawl_check_crawl_statusfirecrawl_extract

perplexity_searchperplexity_search_webperplexity_askperplexity_researchperplexity_reason

tavily_searchtavily_extracttavily_crawltavily_maptavily_research

brave_web_searchbrave_local_search

Other / community labels

Obsidian

9 tools · MarkusPfundstein

obsidian_simple_searchobsidian_complex_searchobsidian_list_files_in_vaultobsidian_list_files_in_dirobsidian_get_file_contentsobsidian_batch_get_file_contentsobsidian_get_periodic_noteobsidian_get_recent_periodic_notesobsidian_get_recent_changes

search_memoriesfind_memories_by_namelist_instancesget_instance_detailsget_instance_by_name

search_recordslist_recordslist_basesget_record

search_modulessearch_providerssearch_policies

Figma

2 tools · GLips/Figma-Context-MCP

get_figma_datadownload_figma_images

Figma's entry is a community server, not an official Figma product. Atlassian is split into two distinct comment blocks (official and community), not one merged total. GitHub's 56 typed entries make it the largest single block in the file.

If you are included

Your tool output collapses into clean one-liners. In a multi-step debugging session, your results read as a narrative while unlisted competitors dump walls of raw JSON. That UX gap compounds over dozens of tool calls. Developers gravitate toward the tools that keep their terminal readable.

If you are not

Your MCP server still works. But the output is uncollapsed: every field, every nested object, rendered in full. Functional but visually noisy. Developers are more likely to disconnect a server that clutters their workflow.

2. claude.ai proxy connectors: 6 explicit labels, Datadog ambiguous

The leaked source directly shows a claude.ai-managed MCP path. It fetches org-configured connectors, converts them to transport type claudeai-proxy, and routes calls through a dedicated proxy at mcp-proxy.anthropic.com, rather than a user-managed local MCP process.

The architecture, visible across three source files:

// oauth.ts — proxy infrastructure
MCP_PROXY_URL: 'https://mcp-proxy.anthropic.com'
MCP_PROXY_PATH: '/v1/mcp/{server_id}'

// claudeai.ts — fetches org-configured servers
const url = `${baseUrl}/v1/mcp_servers?limit=1000`
// requires OAuth scope: 'user:mcp_servers'

// Each server becomes:
configs[name] = {
  type: 'claudeai-proxy',
  url: server.url,
  id: server.id,
  scope: 'claudeai',
}

When a Claude.ai subscriber connects a claude.ai MCP connector, the flow is:

User authenticates via OAuth on claude.ai with the scope user:mcp_servers
Claude Code fetches the org's server list from api.anthropic.com/v1/mcp_servers
Tool calls route through mcp-proxy.anthropic.com/v1/mcp/{server_id}

What the code proves: six label blocks in classifyForCollapse.ts are explicitly marked claude.ai-hosted: Slack, Gmail, Google Drive, Google Calendar, PubMed, and BigQuery. A separate main.tsx comment about skipping claude.ai proxy servers also lists Datadog. Because those two sources disagree, Datadog is ambiguous in the leaked build rather than counted as confirmed hosted.

Slack receives unusually specific treatment in the leaked build. Beyond the proxy transport, Claude Code has three Slack-specific code paths:

1. Compact UI rendering

trySlackSendCompact() in UI.tsx renders message sends as one-liners with #channel names and archive links instead of verbose JSON.

2. Custom OAuth error handling

Slack returns HTTP 200 for error responses, violating RFC 6749. Claude Code normalizes non-standard codes like invalid_refresh_token and token_expired to standard invalid_grant in auth.ts.

3. Deduplication logic

If both a local Slack MCP server and the Claude.ai-hosted version are configured, the local server takes precedence. Handled in config.ts.

PubMed is the outlier in the explicit hosted set. Its label covers 24 typed entries across hosted and community implementations. That does not prove Anthropic's selection criteria, but it does show the proxy path is not limited to workplace chat and calendar tools.

If you are included

Zero-config for users. OAuth managed by Anthropic, server fetched automatically, calls proxied through their infrastructure. No local MCP process, no API keys to rotate, no token refresh to debug. Users click “Connect” and it works. Only 6 vendors have this.

If you are not

Users must configure your MCP server locally: install a package, set environment variables, manage API keys, handle token refresh. Every step is a drop-off point. For enterprise teams, hosted servers win by default because they require no IT support.

3. WebFetch preapproval: 89 host entries, 88 unique scopes

Claude Code's WebFetch tool normally requires the user to provide a URL. But tools/WebFetchTool/preapproved.ts lists 89 preapproved host entries Claude Code can fetch on its own. That resolves to 88 unique host/path scopes because learn.microsoft.com appears twice, and some entries are path-scoped rather than whole-domain allowlists (for example github.com/anthropics and vercel.com/docs). This is a security allowlist for GET requests only. the source warns that it is deliberately not inherited by the sandbox system because some domains allow file uploads.

The list is broad and organized by category. The badges below normalize a few www. prefixes for readability, but the exact allowlist entries come from preapproved.ts.

Anthropic

Claude Platform

platform.claude.com

Claude Code

code.claude.com

MCP Protocol

modelcontextprotocol.io

Agent Skills

agentskills.io

Anthropic GitHub

github.com/anthropics

Programming languages

Python

docs.python.org

C/C++

Java

docs.oracle.com

C# / .NET

MDN

developer.mozilla.org

go.dev

Go Packages

pkg.go.dev

PHP

php.net

Swift

docs.swift.org

Kotlin

kotlinlang.org

Ruby

ruby-doc.org

Rust

doc.rust-lang.org

TypeScript

typescriptlang.org

Web frameworks & libraries

React

react.dev

Angular

angular.io

Vue.js

vuejs.org

Next.js

nextjs.org

Express

expressjs.com

Node.js

nodejs.org

Bun

bun.sh

jQuery

jquery.com

Bootstrap

getbootstrap.com

Tailwind

tailwindcss.com

D3.js

d3js.org

Three.js

threejs.org

Redux

redux.js.org

Webpack

webpack.js.org

Jest

jestjs.io

React Router

reactrouter.com

Python ecosystem

Django

docs.djangoproject.com

Flask

flask.palletsprojects.com

FastAPI

fastapi.tiangolo.com

Pandas

pandas.pydata.org

NumPy

numpy.org

TensorFlow

tensorflow.org

PyTorch

pytorch.org

scikit-learn

scikit-learn.org

Matplotlib

matplotlib.org

Requests

requests.readthedocs.io

Jupyter

jupyter.org

PHP, Java, .NET

Laravel

laravel.com

Symfony

symfony.com

WordPress

wordpress.org

Spring

docs.spring.io

Hibernate

hibernate.org

Tomcat

tomcat.apache.org

Gradle

gradle.org

Maven

maven.apache.org

ASP.NET

asp.net

.NET

dotnet.microsoft.com

NuGet

nuget.org

Blazor

blazor.net

Mobile

React Native

reactnative.dev

Flutter

docs.flutter.dev

Apple Dev

developer.apple.com

Android Dev

developer.android.com

ML & data science

Keras

keras.io

Spark

spark.apache.org

Hugging Face

huggingface.co

Kaggle

kaggle.com

Databases

MongoDB

mongodb.com

Redis

redis.io

PostgreSQL

postgresql.org

MySQL

dev.mysql.com

SQLite

sqlite.org

GraphQL

graphql.org

Prisma

prisma.io

Cloud & DevOps

AWS

docs.aws.amazon.com

Google Cloud

cloud.google.com

Azure

learn.microsoft.com

Kubernetes

kubernetes.io

Docker

docker.com

Terraform

terraform.io

Ansible

ansible.com

Vercel

vercel.com

Netlify

docs.netlify.com

Heroku

devcenter.heroku.com

Testing, game dev, and other tools

Cypress

cypress.io

Selenium

selenium.dev

Unity

docs.unity.com

Unreal

docs.unrealengine.com

Git

git-scm.com

Nginx

nginx.org

Apache

httpd.apache.org

If you are included

Claude Code can pull your docs on its own while reasoning about which tool to recommend. Your product is more legible to the agent during real-time decision-making. When Claude is comparing options, your documentation is one fetch away.

If you are not

Claude Code can still read your docs, but only when the user explicitly pastes a URL. The agent will not proactively visit your site to check compatibility or read your API reference. Your docs are less likely to be visited opportunistically. Your visibility depends on model memory and web search alone.

4. Deployment environment detection: 29 explicit labels

utils/env.ts returns 29 explicit deployment labels before falling back to unknown-darwin, unknown-linux, unknown-win32, or unknown. This is telemetry, not a user-facing feature.

Cloud dev environments

CODESPACES

Gitpod

GITPOD_WORKSPACE_ID

REPL_IDREPL_SLUG

PROJECT_DOMAIN

Cloud platforms

VERCEL

Railway

RAILWAY_ENVIRONMENT_NAME

RAILWAY_ENVIRONMENT_NAMERAILWAY_SERVICE_NAME

RENDER

NETLIFY

DYNO

FLY_APP_NAMEFLY_MACHINE_ID

CF_PAGES

Deno Deploy

DENO_DEPLOYMENT_ID

AWS Lambda

AWS_LAMBDA_FUNCTION_NAME

AWS Fargate

AWS_ECS_FARGATE

AWS_EXECUTION_ENV=AWS_ECS_FARGATE

AWS_EXECUTION_ENV=AWS_ECS_EC2

AWS EC2

/sys/hypervisor/uuid

/sys/hypervisor/uuid → ec2*

K_SERVICE

GOOGLE_CLOUD_PROJECT

Azure App Service

WEBSITE_SITE_NAME

WEBSITE_SITE_NAMEWEBSITE_SKU

Azure Functions

AZURE_FUNCTIONS_ENVIRONMENT

DigitalOcean

ondigitalocean.app

APP_URL contains ondigitalocean.app

Hugging Face Spaces

SPACE_CREATOR_USER_ID

CI/CD

GitHub Actions

GITHUB_ACTIONS

GITLAB_CI

CIRCLECI

BUILDKITE

Container / orchestration

Kubernetes

KUBERNETES_SERVICE_HOST

/.dockerenv file exists

If you are included

Anthropic knows how many Claude Code sessions run on your platform. That data could shape future integration priorities. If 40% of sessions run on your infra, that is a strong argument for deeper partnership. Detection is not a feature for users. It is a signal that feeds Anthropic's product roadmap.

If you are not

Your usage shows up as “unknown-linux” in Anthropic's telemetry. You are invisible in the data that could drive their partnership and integration roadmap. If you want Anthropic to invest in your platform, they first need to see you in their numbers.

5. Secret scanner: 36 rules across 23 credential families

services/teamMemorySync/secretScanner.ts scans content client-side before upload to keep secrets out of team memory. The file says it uses a curated subset of high-confidence Gitleaks rules. In the leaked build that is 36 explicit rules, including multi-rule vendors and a private-key catchall.

The badges below show representative credential families covered by the rules. They are not a one-rule-per-vendor list.

aws-access-token

gcp-api-key

azure-ad-client-secret

anthropic-api-keyanthropic-admin-api-key

openai-api-key

huggingface-access-token

github-patgithub-fine-grained-patgithub-app-tokengithub-oauthgithub-refresh-token

GitLab

glpat-* / deploy token

gitlab-patgitlab-deploy-token

slack-bot-tokenslack-user-tokenslack-app-token

twilio-api-key

sendgrid-api-token

stripe-access-token

shopify-access-tokenshopify-shared-secret

npm-access-token

pypi-upload-token

databricks-api-token

hashicorp-tf-api-token

pulumi-api-token

postman-api-token

Grafana

glsa_ / glc_ token

grafana-api-keygrafana-cloud-api-tokengrafana-service-account-token

sentry-user-tokensentry-org-token

DigitalOcean

dop_v1_ / doo_v1_

digitalocean-patdigitalocean-access-token

HashiCorp TF

atlasv1.* token

hashicorp-tf-api-token

Private Key

BEGIN PRIVATE KEY

private-key

If you are included

Claude Code actively prevents your API keys from leaking into team memory. A safety feature, not a partnership, but it means Anthropic considers your credentials common enough to warrant dedicated regex patterns. Your users are protected by default.

If you are not

Your API keys could end up in team memory if a developer accidentally shares them. Generic detection may still catch them, but there is no vendor-specific regex for your key format. Actionable tip: if your keys have a distinctive prefix, submit a pattern to the public Gitleaks rule set. Claude Code's scanner derives from it, so accepted rules flow downstream.

6. API gateway detection: 7 gateways

services/api/logging.ts detects third-party AI API gateways via response headers or hostnames. Purely observational for analytics. no behavior change.

LiteLLM

x-litellm-* headerslitellm_* hostname

Helicone

helicone-* headers

Portkey

x-portkey-* headers

Cloudflare AI

cf-aig-* headers

Kong

x-kong-* headersvia: kong/* header

Braintrust

x-bt-* headers

Databricks

*.databricks.com hostname

If you are included

Anthropic can see how much traffic flows through your proxy. That visibility matters for pricing, rate-limiting, and partnership conversations. If a large share of traffic routes through your gateway, you have leverage.

If you are not

Your proxy traffic is invisible to Anthropic's analytics. You can operate without scrutiny, but you also have no visibility signal when Anthropic evaluates which gateways to formally support. Being known is the prerequisite for partnership conversations.

7. Plugin tips and special integrations

tipRegistry.ts contains two plugin tips. contextual suggestions to install official plugins from claude-plugins-official:

Vercel

triggered by vercel.json / vercel CLI

Frontend Design

triggered by .html / .css files

Vercel is the only third-party vendor with a dedicated plugin tip. When Claude Code detects a vercel.json or the vercel CLI, it suggests /plugin install vercel@claude-plugins-official. The other tip is Anthropic's own frontend-design plugin, triggered by HTML and CSS files.

Anthropic also maintains a commercial MCP registry at api.anthropic.com/mcp-registry/v0/servers. Servers on this registry have their tool names logged for analytics (rather than being redacted as PII). In practice,inclusion on this registry likely acts as a trust or approval signal.

And Claude Code supports two MCP extension metadata fields any server can use: anthropic/alwaysLoad (tool appears in prompt without deferred loading) and anthropic/searchHint (custom description for tool discovery). These are available to everyone; In practice,servers that populate them are easier for Claude Code to surface and describe cleanly.

If you are included

Claude Code proactively suggests your plugin when it detects relevant project signals. You are part of the install flow, not something the developer has to discover. Vercel is the only third-party vendor with this treatment today. Your tool is recommended before the developer even starts working.

If you are not

Your plugin must be manually discovered and installed. Claude Code will never suggest it. The developer must already know about you, find your server on npm or GitHub, and configure it themselves. Anthropic controls who gets a tip.

8. The complete vendor map

Vendors that appear in at least two systems. The Hosted column marks only explicitly tagged claude.ai-hosted labels; Datadog's hosting ambiguity is noted below instead of encoded as a yes/no.

Showing vendors that appear in 2+ systems. Datadog is omitted here because the leaked build contains a hosted-connector conflict rather than a clean second confirmed system.

9. Conclusion

AI coding agents are becoming real distribution infrastructure. The vendors hardcoded in Claude Code today have a head start that compounds with every session: cleaner output, easier onboarding, docs the agent can read on its own, visibility in telemetry. These advantages are small individually but they stack.

The MCP allowlist is entirely manual. Someone typed each tool name into a Set object by hand. That does not scale. Either Anthropic moves to convention-based classification, or the current list hardens into a de facto ranking of which integrations they consider worth maintaining. For vendors not yet on it, the window to get added narrows as the file grows.

Vercel shows there is more than one path in. It has no MCP UI support but it has a plugin tip, preapproved docs, and environment detection. Different distribution channel, same destination: being recognized by name inside the product.

If you ship developer tools, the question is no longer whether your product works with Claude Code. It is whether Claude Code knows your product exists.

Where does your tool show up?

If you ship developer tools and want to understand your position in the AI agent stack, we can help.