Context Diet: the 25 heaviest MCP tool lists, and how to slim them

Every MCP server you enable sends its full tool list (names, descriptions and input schemas) to the model before you type anything. We took the 25 largest tool lists that qualify for our most-context size list (a token-cost estimate, not deprecated, and not a later failed run whose stored estimate is from an earlier capture) and asked one question: if your agent only does one job, how much of that list could it leave out?

Together these 25 servers list 2,180 tools and cost 1,115,608 tokens (cl100k_base estimate). If each one were trimmed to the slim set we suggest below, the total would be 275,511 tokens, a cut of 840,097 tokens (75%). 2 of them, Microsoft 365 MCP (Softeria) and SmartBear MCP (official), are each larger than a 200k-token context window on their own.

Slim sets are our suggestion; we have not tested a slimmed configuration in a client. We worked them out from the measured per-tool token counts. Token counts are estimates (cl100k_base).

Method

• Source. The tools/list response we captured for each server (the files behind /data/token-cost.json). No credentials were used. A few servers change their tool list with setup, and those rows say so. Descriptions in the site snapshots have whitespace normalised (newlines turned into spaces). Non-ASCII characters are kept as-is.
• Token cost. We use tiktoken cl100k_base over the compact JSON (no spaces, non-ASCII kept) of [{name, description, inputSchema}] for every tool. That is the same figure as /best/most-context-mcp-servers/. It is an estimate: other models use their own tokenizers, and each client adds its own wrapping.
• Slim total. We re-count the kept tools as a single list with the same method. The saving is the full figure minus the slim figure, so it is exact for that method and is not a sum of rounded per-tool numbers.
• Destructive share. The share of tools whose own MCP annotations set destructiveHint: true. We count only an explicit true. We do not infer it from a missing hint. When most of a server's tools carry no hints, we write "not annotated".
• Write share. The share of tools our permission card marks as write (state-changing). That covers destructiveHint: true and readOnlyHint: false first, then the tool's own description, then words in the tool name. A tool with no signal stays "unclassified". Most "read-only agent" suggestions below drop exactly these write tools.
• Ranking status. Each row carries the same status as /best/most-context-mcp-servers/. Deprecated servers are left out of the 25 and listed separately. REA is named in a left-out note, because it is not on that size list.
• Tool filters. Where a maintainer README for the pinned version documents a built-in tool-filter flag or env var, that row links to the README. Otherwise the row says it needs a client-side tool filter.

Summary

Slim suggestions are ours, not tested in a client. Token counts are estimates (cl100k_base). Data: context-diet.json.

Server by server

1. Microsoft 365 MCP (Softeria) (@softeria/ms-365-mcp-server)

• Status on /best/most-context-mcp-servers/: Ranked #1.
• Token cost: 238,259 tokens (cl100k_base estimate), 119.1% of a 200k window.
• Tools: 202.
• Destructive share (destructiveHint: true): 57% (116 of 202); 5 tools not annotated.
• Write share (permission card): 58% (118 of 202); 4 unclassified.
• Slim version (our suggestion, not tested in a client): Mail-only assistant: keep the tools with 'mail' in the name plus list-accounts and get-current-user; drop calendar, OneDrive, Excel, OneNote, Planner, To Do, contacts and subscriptions. That keeps 35 of 202 tools and saves 190,070 tokens (80%), for a slim total of 48,189.
• Alternative, read-only cut: dropping all 118 write tools leaves 84 tools at 47,109 tokens (saves 191,150).
• Tool filter: The README documents a built-in tool filter: The README documents --enabled-tools (or ENABLED_TOOLS), --enabled-tools-file (or ENABLED_TOOLS_FILE), --preset and --read-only (or READ_ONLY) to limit the tool list. README (version 0.178.0).
• Tools dropped in this suggestion (167): accept-calendar-event, add-calendar-event-attachment, add-excel-table-rows, cancel-calendar-event, clear-excel-range, copy-drive-item, create-calendar, create-calendar-event, create-calendar-event-attachment-upload-session, create-contact-child-folder, create-contact-folder, create-contact-in-folder, create-drive-item-preview, create-drive-item-share-link, create-excel-chart, create-excel-table, create-focused-inbox-override, create-forward-draft, create-my-calendar-permission, create-onedrive-folder, create-onenote-notebook, create-onenote-page, create-onenote-section, create-onenote-section-page, create-outlook-category, create-outlook-contact, create-planner-bucket, create-planner-plan, create-planner-task, create-planner-task-message, create-reply-all-draft, create-reply-draft, create-specific-calendar-event, create-subscription, create-todo-checklist-item, create-todo-linked-resource, create-todo-task, create-todo-task-list, create-upload-session, decline-calendar-event, delete-calendar, delete-calendar-event, delete-contact-folder, delete-drive-item-permission, delete-excel-range, delete-excel-table-row, delete-focused-inbox-override, delete-my-calendar-permission, delete-onedrive-file, delete-onenote-page, delete-outlook-contact, delete-planner-bucket, delete-planner-task-message, delete-specific-calendar-event, delete-subscription, delete-todo-checklist-item, delete-todo-linked-resource, delete-todo-task, delete-todo-task-list, dismiss-calendar-event-reminder, and 107 more (full list in context-diet.json).
• Measured: 2026-11-11 (version 0.178.0).

2. SmartBear MCP (official) (@smartbear/mcp)

• Status on /best/most-context-mcp-servers/: Not ranked: Not promoted: no passing tool call, under 50 stars.
• Token cost: 230,507 tokens (cl100k_base estimate), 115.3% of a 200k window.
• Tools: 316.
• Destructive share (destructiveHint: true): 9% (27 of 316).
• Write share (permission card): 39% (123 of 316).
• Slim version (our suggestion, not tested in a client): BugSnag error-triage agent: keep the bugsnag tools; drop the QMetry, QTM4J, Zephyr, contract-testing, Swagger, Reflect, BearQ and Collaborator tools. That keeps 19 of 316 tools and saves 219,391 tokens (95%), for a slim total of 11,116.
• Alternative, read-only cut: dropping all 123 write tools leaves 193 tools at 115,792 tokens (saves 114,715).
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (297): bearq_chat_with_qa_lead, bearq_delete_test_cases, bearq_expand_application_model, bearq_get_task, bearq_get_task_status, bearq_get_workflow_run, bearq_list_environments, bearq_list_workflows, bearq_run_regression_tests, bearq_run_test_cases, bearq_run_tests_in_functional_areas, bearq_run_workflow, bearq_stop_task, bearq_stop_workflow_run, bearq_wait_for_task, collaborator_create_remote_system_configuration, collaborator_create_review, collaborator_delete_remote_system_configuration, collaborator_edit_remote_system_configuration, collaborator_find_review_by_id, collaborator_get_reviews, collaborator_reject_review, collaborator_reviewservice_action, collaborator_test_remote_system_configuration_connection, collaborator_update_remote_system_configuration_webhook, contract-testing_add_label_to_pacticipant, contract-testing_admin_add_role_to_user, contract-testing_admin_create_role, contract-testing_admin_create_system_account, contract-testing_admin_create_team, contract-testing_admin_create_user, contract-testing_admin_delete_role, contract-testing_admin_delete_team, contract-testing_admin_delete_user, contract-testing_admin_get_role, contract-testing_admin_get_system_account_tokens, contract-testing_admin_get_team, contract-testing_admin_get_team_user, contract-testing_admin_get_user, contract-testing_admin_invite_users, contract-testing_admin_list_permissions, contract-testing_admin_list_roles, contract-testing_admin_list_team_users, contract-testing_admin_list_teams, contract-testing_admin_list_users, contract-testing_admin_patch_team_users, contract-testing_admin_remove_role_from_user, contract-testing_admin_remove_user_from_team, contract-testing_admin_reset_roles, contract-testing_admin_set_team_users, contract-testing_admin_set_user_roles, contract-testing_admin_update_role, contract-testing_admin_update_team, contract-testing_admin_update_user, contract-testing_can_i_deploy, contract-testing_check_pactflow_ai_entitlements, contract-testing_create_environment, contract-testing_create_pacticipant, contract-testing_create_secret, contract-testing_create_webhook, and 237 more (full list in context-diet.json).
• Measured: 2026-11-11 (version 0.43.0).

3. Korean Tax Calc MCP (Upstage) (korean-tax-calc-mcp)

• Status on /best/most-context-mcp-servers/: Not ranked: Held: no passing tool call yet.
• Token cost: 78,979 tokens (cl100k_base estimate), 39.5% of a 200k window.
• Tools: 85.
• Destructive share (destructiveHint: true): 0% (0 of 85); 12 tools not annotated.
• Write share (permission card): 0% (0 of 85); 12 unclassified.
• Slim version (our suggestion, not tested in a client): VAT-only agent: keep the tools with 'vat' in the name plus find_tool and common_input_tax; drop income, corporate, inheritance, gift and withholding calculators. That keeps 13 of 85 tools and saves 68,391 tokens (87%), for a slim total of 10,588.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Note: None of this server's tools is marked as a write tool, so a read-only cut saves nothing here. Splitting by tax area is the only way to slim it.
• Tools dropped in this suggestion (72): assessment_limitation, asset_grant_reserve, bad_debt_adjustment, bad_debt_allowance, bundled_supply_split, business_car_expense, capital_gains_income, capital_gains_tax, child_tax_credit, corporate_deemed_rent_income, corporate_tax, daily_worker_withholding, deemed_bonus_resettlement, deemed_dividend_bonus_shares, deemed_dividend_capital_reduction, deemed_interest, depreciation_adjustment, disaster_loss_credit, dividend_received_deduction, dividend_tax_credit, dominant_shareholder, donation_limit, earned_income_gross_pay, entertainment_adjustment, entertainment_limit, estimated_business_income, export_workbook, extract_relations, financial_income_aggregate, foreign_tax_credit, gift_tax, housing_deemed_rent_income, housing_rental_income, income_adjustment_total, income_loss_offset, income_tax, inheritance_gift_valuation, inheritance_tax, input_tax_recalculation, inventory_valuation_adjustment, invoice_penalty, kinship_check, late_payment_penalty, liquidation_income, loss_carryforward_limit, minimum_tax, missing_receipt_disallowance, nonbusiness_interest_disallowance, nonresident_withholding, other_income_aggregate, ownership_ratio, pension_account_withdrawal, pension_income, personal_deduction, precheck_schema, related_judge, related_provision_at, religious_income, retirement_allowance_adjustment, retirement_income_tax, and 12 more (full list in context-diet.json).
• Measured: 2026-11-11 (version 0.15.1).

4. BlindPay MCP (official) (@blindpay/mcp)

• Status on /best/most-context-mcp-servers/: Not ranked: Not promoted: no passing tool call, under 50 stars.
• Token cost: 60,359 tokens (cl100k_base estimate), 30.2% of a 200k window.
• Tools: 93.
• Destructive share (destructiveHint: true): 9% (8 of 93).
• Write share (permission card): 52% (48 of 93).
• Slim version (our suggestion, not tested in a client): Read-only payments lookup agent: drop the 48 tools our permission card marks as write (state-changing); keep the 45 read tools. That keeps 45 of 93 tools and saves 54,512 tokens (90%), for a slim total of 5,847.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (48): DeleteV1InstancesCustomersBankAccountsById, DeleteV1InstancesCustomersBlockchainWalletsById, DeleteV1InstancesCustomersById, DeleteV1InstancesCustomersWalletsById, DeleteV1InstancesMembersByUserId, DeleteV1InstancesPartnerFeesById, DeleteV1InstancesPayablesById, DeleteV1InstancesWebhookEndpointsById, PatchV1InstancesCustomersBlockchainWalletsById, PostV1EInstancesTos, PostV1InstancesCreateAssetTrustline, PostV1InstancesCustomers, PostV1InstancesCustomersBankAccounts, PostV1InstancesCustomersBankAccountsOfframpWallets, PostV1InstancesCustomersBankAccountsPlaid, PostV1InstancesCustomersBlockchainWallets, PostV1InstancesCustomersLimitIncrease, PostV1InstancesCustomersRfi, PostV1InstancesCustomersVirtualAccounts, PostV1InstancesCustomersWallets, PostV1InstancesCustomersWalletsYieldDisable, PostV1InstancesCustomersWalletsYieldEnable, PostV1InstancesMintUsdbSolana, PostV1InstancesMintUsdbStellar, PostV1InstancesPartnerFees, PostV1InstancesPayables, PostV1InstancesPayinQuotes, PostV1InstancesPayinQuotesFx, PostV1InstancesPayinsCardPayment, PostV1InstancesPayinsEvm, PostV1InstancesPayoutsDocuments, PostV1InstancesPayoutsEvm, PostV1InstancesPayoutsSolana, PostV1InstancesPayoutsStellar, PostV1InstancesPayoutsStellarAuthorize, PostV1InstancesPrepareDelegateSolana, PostV1InstancesQuotes, PostV1InstancesQuotesFx, PostV1InstancesRfi, PostV1InstancesTransferQuotes, PostV1InstancesTransfers, PostV1InstancesWebhookEndpoints, PostV1Presign, PutInstancesCustomersBankAccountsOfframpWalletsById, PutV1InstancesById, PutV1InstancesCustomersById, PutV1InstancesCustomersVirtualAccountsById, PutV1InstancesMembersByUserId.
• Measured: 2026-11-11 (version 1.10.0).

5. CloudBase MCP (@cloudbase/cloudbase-mcp)

• Status on /best/most-context-mcp-servers/: Ranked #2.
• Token cost: 56,457 tokens (cl100k_base estimate), 28.2% of a 200k window.
• Tools: 42.
• Destructive share (destructiveHint: true): 40% (17 of 42).
• Write share (permission card): 52% (22 of 42).
• Slim version (our suggestion, not tested in a client): Read-only environment inspection agent: drop the 22 tools our permission card marks as write (state-changing); keep the 20 read tools. That keeps 20 of 42 tools and saves 38,950 tokens (69%), for a slim total of 17,507.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (22): auth, callCloudApi, deployApply, deployBuild, downloadTemplate, envDomainManagement, manageAgents, manageAppAuth, manageApps, manageCloudRun, manageEnv, manageFunctions, manageGateway, manageHosting, manageLogs, manageMysqlDatabase, managePermissions, managePgDatabase, manageStorage, modifyDataModel, writeNoSqlDatabaseContent, writeNoSqlDatabaseStructure.
• Measured: 2026-11-11 (version 2.37.5).

6. Google Workspace MCP (taylorwilsdon) (workspace-mcp)

• Status on /best/most-context-mcp-servers/: Ranked #3.
• Token cost: 46,166 tokens (cl100k_base estimate), 23.1% of a 200k window.
• Tools: 119.
• Destructive share (destructiveHint: true): 33% (39 of 119).
• Write share (permission card): 54% (64 of 119).
• Slim version (our suggestion, not tested in a client): Gmail-only assistant: keep the tools with 'gmail' in the name plus start_google_auth; drop Docs, Sheets, Slides, Forms, Drive, Calendar, Chat, Contacts, Tasks, Apps Script and search tools. That keeps 15 of 119 tools and saves 39,783 tokens (86%), for a slim total of 6,383.
• Alternative, read-only cut: dropping all 64 write tools leaves 55 tools at 13,963 tokens (saves 32,203).
• Tool filter: The README documents a built-in tool filter: The README documents --tools to cherry-pick services, --tool-tier (core, extended, complete), --disabled-tools to subtract named tools, and --read-only. README (version 2.3.0).
• Tools dropped in this suggestion (104): append_table_rows, batch_update_doc, batch_update_form, batch_update_presentation, check_drive_file_public_access, copy_drive_file, create_calendar, create_doc, create_drive_file, create_drive_folder, create_form, create_presentation, create_reaction, create_sheet, create_spreadsheet, create_table_with_data, debug_table_structure, download_chat_attachment, export_doc_to_pdf, find_and_replace_doc, format_sheet_range, generate_trigger_code, get_contact, get_contact_group, get_doc_as_markdown, get_doc_content, get_drive_file_content, get_drive_file_download_url, get_drive_file_permissions, get_drive_shareable_link, get_events, get_form, get_form_response, get_messages, get_page, get_presentation, get_script_activity, get_script_project, get_script_version, get_search_engine_info, get_spreadsheet_info, get_task, get_task_list, import_to_google_doc, import_to_google_sheets, import_to_google_slides, insert_doc_elements, insert_doc_image, inspect_doc_structure, list_calendars, list_contact_groups, list_contacts, list_docs_in_folder, list_document_comments, list_drive_items, list_form_responses, list_presentation_comments, list_script_deployments, list_sheet_tables, list_spaces, and 44 more (full list in context-diet.json).
• Measured: 2026-11-11 (version 2.3.0).

7. OneUptime MCP (official, remote) (oneuptime-remote)

• Status on /best/most-context-mcp-servers/: Ranked #4.
• Token cost: 45,985 tokens (cl100k_base estimate), 23.0% of a 200k window.
• Tools: 351.
• Destructive share (destructiveHint: true): 47% (165 of 351).
• Write share (permission card): 47% (165 of 351).
• Slim version (our suggestion, not tested in a client): Read-only monitoring and incident lookup agent: drop the 165 tools our permission card marks as write (state-changing); keep the 186 read tools. That keeps 186 of 351 tools and saves 25,349 tokens (55%), for a slim total of 20,636.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (165): acknowledge_alert, acknowledge_incident, add_alert_note, add_incident_note, create_alert, create_alert_internal_note, create_alert_severity, create_alert_state, create_alert_state_timeline, create_bi_dashboard, create_bi_tool, create_check_setting, create_cloud_budget, create_custom_data_type, create_data_asset, create_data_check, create_data_flow, create_data_observability_setting, create_data_pipeline, create_data_product, create_data_reconciliation, create_data_security_identity, create_data_security_issue, create_data_security_scan, create_data_security_setting, create_data_store, create_data_store_access_entry, create_data_store_finding, create_incident, create_incident_alert, create_incident_internal_note, create_incident_public_note, create_incident_severity, create_incident_state, create_incident_state_timeline, create_label, create_monitor, create_monitor_status, create_monitor_status_event, create_on_call_policy, create_postmortem_status, create_residency_rule, create_scheduled_maintenance_event, create_scheduled_maintenance_state, create_scheduled_maintenance_state_timeline, create_security_case, create_security_case_evidence, create_security_case_note, create_security_case_priority, create_security_case_status, create_security_case_task, create_security_detection, create_security_playbook, create_status_page, create_status_page_announcement, create_team, create_warehouse_job, delete_alert, delete_alert_internal_note, delete_alert_severity, and 105 more (full list in context-diet.json).
• Measured: 2026-11-11 (version n/a).

8. Porkbun MCP (@porkbunllc/mcp-server)

• Status on /best/most-context-mcp-servers/: Not ranked: Not promoted: no passing tool call, under 50 stars.
• Token cost: 40,842 tokens (cl100k_base estimate), 20.4% of a 200k window.
• Tools: 120.
• Destructive share (destructiveHint: true): 30% (36 of 120).
• Write share (permission card): 54% (65 of 120).
• Slim version (our suggestion, not tested in a client): Read-only domain and DNS lookup agent: drop the 65 tools our permission card marks as write (state-changing); keep the 55 read tools. That keeps 55 of 120 tools and saves 29,220 tokens (72%), for a slim total of 11,622.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (65): buy_closeout, cancel_transfer, configure_auto_topup, confirm_email_verification, confirm_phone_verification, connect_domains_to_cloudflare, create_cloudflare_record, create_dns_record, create_dnssec_record, create_email_forward, create_glue_record, create_hosting, create_mailbox, create_sandbox_key, create_url_forward, create_webhook, create_wp_credentials, delete_cloudflare_record, delete_dns_record, delete_dnssec_record, delete_email_forward, delete_glue_record, delete_hosting, delete_hosting_file, delete_url_forward, delete_webhook, delete_wp_credentials, deploy_site, disconnect_cloudflare, edit_cloudflare_record, import_dns_records, make_hosting_dir, prepare_transfer, register_domain, renew_domain, renew_mailbox, request_api_key, resend_webhook, restore_dns_zone, retrieve_api_key, retry_cloudflare_domain, rollback_cloudflare_domain, rotate_webhook_secret, sandbox_reset, sandbox_topup, sandbox_trigger_webhook, send_email_verification_code, send_phone_verification_code, set_cloudflare_proxy, set_cloudflare_zone_settings, set_email_password, setup_mailbox, start_transfer, test_webhook, top_up_account_credit, top_up_with_card_mpp, top_up_with_usdc, transfer_domain, update_auto_renew, update_contacts, and 5 more (full list in context-diet.json).
• Measured: 2026-11-11 (version 0.55.0).

9. Keboola MCP (official) (keboola-mcp-server)

• Status on /best/most-context-mcp-servers/: Not ranked: Held: no passing tool call yet.
• Token cost: 36,363 tokens (cl100k_base estimate), 18.2% of a 200k window.
• Tools: 47.
• Destructive share (destructiveHint: true): 28% (13 of 47).
• Write share (permission card): 49% (23 of 47).
• Slim version (our suggestion, not tested in a client): Read-only data exploration agent: drop the 23 tools our permission card marks as write (state-changing); keep the 24 read tools. That keeps 24 of 47 tools and saves 23,512 tokens (65%), for a slim total of 12,851.
• Tool filter: The README documents a built-in tool filter: The README documents HTTP headers X-Allowed-Tools, X-Disallowed-Tools and X-Read-Only-Mode to limit which tools the server exposes. README (version 1.112.2).
• Tools dropped in this suggestion (23): add_config_row, create_conditional_flow, create_config, create_flow, create_oauth_url, create_project, create_python_js_data_app_git_credential, create_sql_transformation, create_stream, delete_python_js_data_app_draft, deploy_data_app, link_shared_bucket, modify_flow, modify_python_js_data_app, modify_streamlit_data_app, run_job, run_sync_action, update_config, update_config_row, update_descriptions, update_flow, update_project_description, update_sql_transformation.
• Measured: 2026-11-11 (version 1.112.2).

10. Toggl MCP (@togglhq/mcp)

• Status on /best/most-context-mcp-servers/: Not ranked: No public repo.
• Token cost: 26,860 tokens (cl100k_base estimate), 13.4% of a 200k window.
• Tools: 54.
• Destructive share (destructiveHint: true): 43% (23 of 54).
• Write share (permission card): 67% (36 of 54).
• Slim version (our suggestion, not tested in a client): Read-only time reporting agent: drop the 36 tools our permission card marks as write (state-changing); keep the 18 read tools. That keeps 18 of 54 tools and saves 22,964 tokens (85%), for a slim total of 3,896.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (36): ai, alerts, auth, clients, custom-fields, customreports, designations, goals, logout, milestones, org-invitations, org-members, org-time-off-me, organization, profile-remove, profile-switch, projects, publicsharedreports, rates, recurring-budget-suggestions, reportschedules, saved-views, shared-working-hours, sharedreports, statuses, subscriptions, tags, tasks, time-entries, time-off, timesheets, track-team-invites, users, working-hours, workspace-settings, workspace-switch.
• Measured: 2026-11-11 (version 1.13.121).

11. GrowSurf MCP (@growsurfteam/growsurf-mcp)

• Status on /best/most-context-mcp-servers/: Ranked #5.
• Token cost: 24,272 tokens (cl100k_base estimate), 12.1% of a 200k window.
• Tools: 66.
• Destructive share (destructiveHint: true): 35% (23 of 66).
• Write share (permission card): 53% (35 of 66).
• Slim version (our suggestion, not tested in a client): Read-only referral reporting agent: drop the 35 tools our permission card marks as write (state-changing); keep the 31 read tools. That keeps 31 of 66 tools and saves 16,425 tokens (68%), for a slim total of 7,847.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (35): growsurf_add_participant, growsurf_bulk_delete_participants, growsurf_cancel_delayed_referral, growsurf_capture_referral_flow_screenshots, growsurf_clone_campaign, growsurf_create_campaign, growsurf_create_campaign_reward, growsurf_create_campaign_webhook, growsurf_create_mobile_participant_token, growsurf_create_program_resource, growsurf_delete_campaign_reward, growsurf_delete_campaign_webhook, growsurf_delete_program_resource, growsurf_email_participant, growsurf_prepare_program_resource_file, growsurf_record_sale, growsurf_refund_transaction, growsurf_remove_campaign_custom_domain, growsurf_request_participant_payout_destination_confirmation, growsurf_request_team_verification, growsurf_resend_team_owner_verification_email, growsurf_set_campaign_custom_domain, growsurf_test_campaign_webhook, growsurf_trigger_referral, growsurf_update_campaign, growsurf_update_campaign_design, growsurf_update_campaign_emails, growsurf_update_campaign_installation, growsurf_update_campaign_options, growsurf_update_campaign_reward, growsurf_update_campaign_webhook, growsurf_update_participant, growsurf_update_program_resource, growsurf_update_team, growsurf_verify_campaign_custom_domain.
• Measured: 2026-11-11 (version 0.19.30).

12. Grafana MCP Server (official) (grafana/mcp-grafana)

• Status on /best/most-context-mcp-servers/: Not ranked: Held: no passing tool call yet.
• Token cost: 20,513 tokens (cl100k_base estimate), 10.3% of a 200k window.
• Tools: 79.
• Destructive share (destructiveHint: true): 11% (9 of 79).
• Write share (permission card): 24% (19 of 79).
• Slim version (our suggestion, not tested in a client): Read-only incident investigation agent: drop the 19 tools our permission card marks as write (state-changing); keep the 60 read tools. That keeps 60 of 79 tools and saves 7,043 tokens (34%), for a slim total of 13,470.
• Tool filter: The README documents a built-in tool filter: The README documents --enabled-tools to choose tool categories (some categories stay off until listed) and --disable-write to drop write tools. README (version 2.1.1).
• Tools dropped in this suggestion (19): add_activity_to_incident, alerting_routing_write, alerting_rules_write, alerting_silences_write, create_annotation, create_datasource, create_folder, create_incident, create_snapshot, delete_annotation, delete_snapshot, generate_deeplink, grafana_api_request, install_plugin, update_alert_group, update_annotation, update_dashboard, update_datasource, update_incident.
• Measured: 2026-11-11 (version 2.1.1).

13. dbt MCP (official) (dbt-mcp)

• Status on /best/most-context-mcp-servers/: Ranked #6.
• Token cost: 18,871 tokens (cl100k_base estimate), 9.4% of a 200k window.
• Tools: 39.
• Destructive share (destructiveHint: true): 0% (0 of 39).
• Write share (permission card): 8% (3 of 39).
• Slim version (our suggestion, not tested in a client): Read-only analytics agent: drop the 3 tools our permission card marks as write (state-changing); keep the 36 read tools. That keeps 36 of 39 tools and saves 1,221 tokens (6%), for a slim total of 17,650.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (3): cancel_job_run, retry_job_run, trigger_job_run.
• Measured: 2026-11-11 (version 2.10.1).

14. Power BI Modeling MCP (@microsoft/powerbi-modeling-mcp)

• Status on /best/most-context-mcp-servers/: Ranked #7.
• Token cost: 17,921 tokens (cl100k_base estimate), 9.0% of a 200k window.
• Tools: 22.
• Destructive share (destructiveHint: true): 82% (18 of 22).
• Write share (permission card): 82% (18 of 22).
• Slim version (our suggestion, not tested in a client): Read-only DAX query agent: drop the 18 tools our permission card marks as write (state-changing); keep the 4 read tools. That keeps 4 of 22 tools and saves 16,614 tokens (93%), for a slim total of 1,307.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (18): calculation_group_operations, calendar_operations, column_operations, culture_operations, database_operations, function_operations, measure_operations, model_operations, named_expression_operations, object_translation_operations, partition_operations, perspective_operations, query_group_operations, relationship_operations, security_role_operations, table_operations, transaction_operations, user_hierarchy_operations.
• Measured: 2026-11-11 (version 1.0.1).

15. Trigger.dev MCP (official) (trigger.dev)

• Status on /best/most-context-mcp-servers/: Ranked #8.
• Token cost: 17,869 tokens (cl100k_base estimate), 8.9% of a 200k window.
• Tools: 55.
• Destructive share (destructiveHint: true): 42% (23 of 55).
• Write share (permission card): 42% (23 of 55).
• Slim version (our suggestion, not tested in a client): Read-only run monitoring agent: drop the 23 tools our permission card marks as write (state-changing); keep the 32 read tools. That keeps 32 of 55 tools and saves 9,604 tokens (54%), for a slim total of 8,265.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (23): cancel_run, close_agent_chat, create_alert, create_dashboard, create_project_in_org, create_prompt_override, deploy, generate_webhook_secret, initialize_project, promote_prompt_version, reactivate_prompt_override, remove_prompt_override, replay_run, replay_webhook_delivery, resolve_alert, send_agent_message, set_dashboard_widgets, set_webhook_secret, start_agent_chat, submit_feedback, trigger_task, update_prompt_override, write_session_channel.
• Measured: 2026-11-11 (version 4.8.3).

16. JustLend MCP (@justlend/mcp-server-justlend)

• Status on /best/most-context-mcp-servers/: Ranked #9.
• Token cost: 17,553 tokens (cl100k_base estimate), 8.8% of a 200k window.
• Tools: 104.
• Destructive share (destructiveHint: true): 27% (28 of 104).
• Write share (permission card): 39% (41 of 104).
• Slim version (our suggestion, not tested in a client): Read-only market data agent: drop the 41 tools our permission card marks as write (state-changing); keep the 63 read tools. That keeps 63 of 104 tools and saves 7,816 tokens (45%), for a slim total of 9,737.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (41): approve_jst_for_voting, approve_liquidator_token, approve_moolah_proxy, approve_moolah_vault, approve_underlying, borrow, buy_energy_direct, cast_vote, claim_moolah_mining_period, claim_rewards, claim_strx_rewards, claim_v1_mining_period, connect_browser_wallet, deposit_jst_for_votes, enter_market, exit_market, moolah_borrow, moolah_liquidate, moolah_repay, moolah_supply_collateral, moolah_vault_deposit, moolah_vault_redeem, moolah_vault_withdraw, moolah_withdraw_collateral, rent_energy, repay, return_energy_rental, set_active_wallet, set_network, set_wallet_mode, stake_trx_to_strx, supply, transfer_trc20, transfer_trx, unstake_strx, unwrap_trx, withdraw, withdraw_all, withdraw_votes_from_proposal, withdraw_votes_to_jst, wrap_trx.
• Measured: 2026-11-11 (version 1.1.3).

17. Browserless MCP (official) (@browserless.io/mcp)

• Status on /best/most-context-mcp-servers/: Not ranked: Held: no passing tool call yet.
• Token cost: 17,023 tokens (cl100k_base estimate), 8.5% of a 200k window.
• Tools: 19.
• Destructive share (destructiveHint: true): 26% (5 of 19).
• Write share (permission card): 26% (5 of 19).
• Slim version (our suggestion, not tested in a client): Read-only scraping and session lookup agent: drop the 5 tools our permission card marks as write (state-changing); keep the 14 read tools. That keeps 14 of 19 tools and saves 10,294 tokens (60%), for a slim total of 6,729.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (5): browserless_agent, browserless_function, browserless_link_checkout, browserless_link_connect, browserless_live_url.
• Measured: 2026-11-11 (version 1.43.1).

18. Files.com MCP (official) (files-com-mcp)

• Status on /best/most-context-mcp-servers/: Not ranked: Not promoted: no passing tool call, under 50 stars.
• Token cost: 16,781 tokens (cl100k_base estimate), 8.4% of a 200k window.
• Tools: 67.
• Destructive share (destructiveHint: true): 30% (20 of 67).
• Write share (permission card): 40% (27 of 67).
• Slim version (our suggestion, not tested in a client): Read-only file lookup agent: drop the 27 tools our permission card marks as write (state-changing); keep the 40 read tools. That keeps 40 of 67 tools and saves 4,123 tokens (25%), for a slim total of 12,658.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (27): Begin_Delete_File, Copy_File, Create_Bundle, Create_Bundle_Notification, Create_Bundle_Recipient, Create_Folder, Create_Group, Create_Permission, Create_User, Delete_Bundle, Delete_Bundle_Notification, Delete_File, Delete_Group, Delete_Permission, Delete_User, Download_File_to_Local, Gpg_Decrypt_File, Gpg_Encrypt_File, Move_File, Transform_File, Unzip_File, Update_Bundle, Update_Bundle_Notification, Update_Group, Update_User, Upload_File_from_Local, Zip_File.
• Measured: 2026-11-11 (version 1.0.105).

19. Azure DevOps MCP (official) (@azure-devops/mcp)

• Status on /best/most-context-mcp-servers/: Ranked #10.
• Token cost: 16,641 tokens (cl100k_base estimate), 8.3% of a 200k window.
• Tools: 44.
• Destructive share (destructiveHint: true): 32% (14 of 44).
• Write share (permission card): 41% (18 of 44).
• Slim version (our suggestion, not tested in a client): Read-only work item and repo lookup agent: drop the 18 tools our permission card marks as write (state-changing); keep the 26 read tools. That keeps 26 of 44 tools and saves 7,548 tokens (45%), for a slim total of 9,093.
• Tool filter: The README documents a built-in tool filter: The README documents a -d domains flag on the local server (core, work, work-items, search, test-plans, repositories, wiki, pipelines, advanced-security) to load only chosen tool groups. README (version 2.11.0).
• Tools dropped in this suggestion (18): pipelines_artifact, pipelines_write, repo_create_branch, repo_file_write, repo_pull_request_thread_write, repo_pull_request_write, testplan_test_case_write, testplan_test_plan_write, testplan_test_suite_write, wiki_upsert_page, wit_backlog, wit_work_item_attachment_link, wit_work_item_attachment_upload, wit_work_item_comment_write, wit_work_item_link_write, wit_work_item_write, work_capacity_write, work_iteration_write.
• Measured: 2026-11-11 (version 2.11.0).

20. Notion (official) (@notionhq/notion-mcp-server)

• Status on /best/most-context-mcp-servers/: Ranked #11.
• Tool list changes with setup: tool list can depend on the connected account.
• Token cost: 16,383 tokens (cl100k_base estimate), 8.2% of a 200k window.
• Tools: 24.
• Destructive share (destructiveHint: true): 50% (12 of 24).
• Write share (permission card): 50% (12 of 24).
• Slim version (our suggestion, not tested in a client): Read-and-search agent: keep the read tools plus API-post-search and API-query-data-source (both annotated as not read-only by the server, but needed to find pages); drop page, block, comment and data-source writes. That keeps 14 of 24 tools and saves 7,334 tokens (45%), for a slim total of 9,049.
• Alternative, read-only cut: dropping all 12 write tools leaves 12 tools at 7,284 tokens (saves 9,099).
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Note: A strict read-only cut (dropping every tool the server annotates as not read-only) would also remove search and data-source queries. That leaves 12 tools at 7,284 tokens, but the agent could not find pages.
• Tools dropped in this suggestion (10): API-create-a-comment, API-create-a-data-source, API-delete-a-block, API-move-page, API-patch-block-children, API-patch-page, API-post-page, API-update-a-block, API-update-a-data-source, API-update-page-markdown.
• Measured: 2026-11-11 (version 2.6.1).

21. Mapbox (official) (@mapbox/mcp-server)

• Status on /best/most-context-mcp-servers/: Ranked #12.
• Token cost: 15,978 tokens (cl100k_base estimate), 8.0% of a 200k window.
• Tools: 31.
• Destructive share (destructiveHint: true): 0% (0 of 31).
• Write share (permission card): 3% (1 of 31).
• Slim version (our suggestion, not tested in a client): Search and directions agent: drop the offline geometry helpers (area, bbox, bearing, buffer, centroid, convex, destination, difference, distance, intersect, length, midpoint, nearest point, nearest point on line, points within polygon, simplify, union) and feedback_tool. That keeps 13 of 31 tools and saves 5,139 tokens (32%), for a slim total of 10,839.
• Alternative, read-only cut: dropping all 1 write tools leaves 30 tools at 15,246 tokens (saves 732).
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (18): area_tool, bbox_tool, bearing_tool, buffer_tool, centroid_tool, convex_tool, destination_tool, difference_tool, distance_tool, feedback_tool, intersect_tool, length_tool, midpoint_tool, nearest_point_on_line_tool, nearest_point_tool, points_within_polygon_tool, simplify_tool, union_tool.
• Measured: 2026-11-11 (version 0.17.0).

22. MongoDB (official) (mongodb-mcp-server)

• Status on /best/most-context-mcp-servers/: Ranked #13.
• Tool list changes with setup: tool list can depend on the connected database.
• Token cost: 14,277 tokens (cl100k_base estimate), 7.1% of a 200k window.
• Tools: 27.
• Destructive share (destructiveHint: true): 22% (6 of 27).
• Write share (permission card): 33% (9 of 27).
• Slim version (our suggestion, not tested in a client): Read-only query agent: drop the 9 tools our permission card marks as write (state-changing); keep the 18 read tools. That keeps 18 of 27 tools and saves 3,814 tokens (27%), for a slim total of 10,463.
• Tool filter: The README documents a built-in tool filter: The README documents MDB_MCP_DISABLED_TOOLS or --disabledTools to turn off named tools, operation types (create, update, delete) or categories (atlas, mongodb). README (version 3.3.0).
• Tools dropped in this suggestion (9): create-collection, create-index, delete-many, drop-collection, drop-database, drop-index, insert-many, rename-collection, update-many.
• Measured: 2026-11-11 (version 3.3.0).

23. Azure MCP (Microsoft) (@azure/mcp)

• Status on /best/most-context-mcp-servers/: Ranked #14.
• Token cost: 13,912 tokens (cl100k_base estimate), 7.0% of a 200k window.
• Tools: 73.
• Destructive share (destructiveHint: true): not annotated (68 of 73 tools carry no hint; 0 flagged).
• Write share (permission card): 10% (7 of 73); 60 unclassified.
• Slim version (our suggestion, not tested in a client): App team on Storage and Key Vault: keep storage, keyvault, monitor, group_list, group_resource_list, subscription_list, documentation and get_azure_bestpractices; drop the other service namespaces. Most tools here carry no read/write signal, so a read-only cut is not possible. That keeps 8 of 73 tools and saves 12,190 tokens (88%), for a slim total of 1,722.
• Alternative, read-only cut: dropping all 7 write tools leaves 66 tools at 12,033 tokens (saves 1,879).
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Note: Azure MCP groups each Azure service behind one namespace tool. Its average tool is about 190 tokens, so the size comes from the number of services. Keep the namespaces you use. The keep list here is one example.
• Tools dropped in this suggestion (65): acr, adme, advisor, aks, appconfig, applens, applicationinsights, appservice, arm, azd, azurebackup, azuremigrate, azureterraform, azureterraformbestpractices, bicepschema, cloudarchitect, communication, compute, confidentialledger, containerapps, cosmos, datadog, deploy, deviceprovisioning, deviceregistry, eventgrid, eventhubs, extension_cli_generate, extension_cli_install, fileshares, foundry, foundryextensions, functionapp, functions, grafana, insights, iothub, iotoperations, kusto, loadtesting, managedlustre, marketplace, mysql, netappfiles, optimization, policy, postgres, pricing, quota, redis, resiliency, resiliencyagent, resourcehealth, role, search, servicebus, servicefabric, signalr, speech, sql, and 5 more (full list in context-diet.json).
• Measured: 2026-11-11 (version 4.0.0-beta.6).

24. yutu YouTube MCP (@eat-pray-ai/yutu)

• Status on /best/most-context-mcp-servers/: Ranked #15.
• Token cost: 13,439 tokens (cl100k_base estimate), 6.7% of a 200k window.
• Tools: 74.
• Destructive share (destructiveHint: true): 20% (15 of 74).
• Write share (permission card): 70% (52 of 74).
• Slim version (our suggestion, not tested in a client): Read-only channel analytics agent: drop the 52 tools our permission card marks as write (state-changing); keep the 22 read tools. That keeps 22 of 74 tools and saves 8,485 tokens (63%), for a slim total of 4,954.
• Tool filter: Needs a client-side tool filter (turning tools off in the client, or an allowlist or proxy in front of the server).
• Tools dropped in this suggestion (52): abuseReport-insert, caption-delete, caption-download, caption-insert, caption-update, channel-update, channelBanner-insert, channelSection-delete, comment-delete, comment-insert, comment-markAsSpam, comment-setModerationStatus, comment-update, commentThread-insert, liveBroadcast-bind, liveBroadcast-delete, liveBroadcast-insert, liveBroadcast-insertCuepoint, liveBroadcast-transition, liveBroadcast-update, liveChatBan-delete, liveChatBan-insert, liveChatMessage-delete, liveChatMessage-insert, liveChatMessage-transition, liveChatModerator-delete, liveChatModerator-insert, liveStream-delete, liveStream-insert, liveStream-update, playlist-delete, playlist-insert, playlist-update, playlistImage-delete, playlistImage-insert, playlistImage-update, playlistItem-delete, playlistItem-insert, playlistItem-update, subscription-delete, subscription-insert, thirdPartyLink-delete, thirdPartyLink-insert, thirdPartyLink-update, thumbnail-set, video-delete, video-insert, video-rate, video-reportAbuse, video-update, watermark-set, watermark-unset.
• Measured: 2026-11-11 (version 0.11.0).

25. AntV Chart (@antv/mcp-server-chart)

• Status on /best/most-context-mcp-servers/: Ranked #16.
• Token cost: 13,398 tokens (cl100k_base estimate), 6.7% of a 200k window.
• Tools: 27.
• Destructive share (destructiveHint: true): 0% (0 of 27).
• Write share (permission card): 0% (0 of 27).
• Slim version (our suggestion, not tested in a client): Everyday charts agent: keep area, bar, column, line, pie and scatter chart tools; drop maps, diagrams, spreadsheets and specialist charts. That keeps 6 of 27 tools and saves 10,305 tokens (77%), for a slim total of 3,093.
• Tool filter: The README documents a built-in tool filter: The README documents a DISABLED_TOOLS environment variable (a comma-separated list of tool names to disable). README (version 0.9.10).
• Tools dropped in this suggestion (21): generate_boxplot_chart, generate_district_map, generate_dual_axes_chart, generate_fishbone_diagram, generate_flow_diagram, generate_funnel_chart, generate_histogram_chart, generate_liquid_chart, generate_mind_map, generate_network_graph, generate_organization_chart, generate_path_map, generate_pin_map, generate_radar_chart, generate_sankey_chart, generate_spreadsheet, generate_treemap_chart, generate_venn_chart, generate_violin_chart, generate_waterfall_chart, generate_word_cloud_chart.
• Measured: 2026-11-11 (version 0.9.10).

Left out: REA: Reverse Engineer Anything

REA: Reverse Engineer Anything is not in this 25 because it is not on the most-context size list: it failed its latest published run (the stored estimate is from an earlier capture). The stored estimate is 179,483 tokens for 207 tools, measured 2026-11-11. See its listing.

Left out: deprecated servers

Deprecated servers are not part of the 25. None of the 8 deprecated servers we have measured would have made the top 25 (the cutoff is 13,398 tokens). The largest is CircleCI (official) at 10,761 tokens and 13 tools (measured 2026-11-11). The others are GitHub (legacy reference) 3,359, Browser (agent-infra) 1,549, Browserbase 1,044, Puppeteer (legacy reference) 518, Redis (legacy reference) 214, AWS KB Retrieval (legacy reference) 99, and Postgres (legacy reference) 32. See /deprecated/.

Related: Which MCP servers use the most context? · Lowest context cost MCP servers · How much context do your MCP servers cost? · Tool history · Methodology.

Published 2026-11-12 by the testedmcp team. Figures rebuilt 2026-11-12 from live captures.