Consistent Error Response Structure
Tool execution errors now return a standardized response format across all failure types. Previously, thedata field was empty on errors—now it always includes status_code and message, matching the structure of successful responses.
What Changed
All error responses from tool execution now include:data.status_code: HTTP status code (ornullfor non-HTTP errors)data.message: Detailed error messageerror: Same detailed message at the top level
Before vs After
Previous error response:Why This Matters
- Easier parsing: Agents and code can reliably access error details from
data.messagewithout special-casing emptydataobjects - Better debugging: Detailed error messages replace generic HTTP error strings
- Consistent schema: Same response shape whether the tool succeeds or fails
Union Types Preserved in Tool Schemas
Tool schemas now use standard JSON SchemaanyOf for union types, providing accurate type information for LLMs and code generators.
What Changed
Two changes affect how types appear in request/response schemas:Toolkits with multi-type union fields (157 total)
Toolkits with multi-type union fields (157 total)
CRM & Sales: active_campaign, apollo, attio, autobound, capsule_crm, firmao, forcemanager, gong, hubspot, instantly, intercom, kommo, leadfeeder, lever, magnetic, pipedrive, pipeline_crm, salesforce, salesforce_service_cloud, zoominfoMarketing & Email: active_trail, beamer, delighted, dripcel, enginemailer, mailerlite, moosend, mopinion, sendspark, tonedenCommunication & Collaboration: chmeetings, discord, helpdesk, helpwise, missive, slack, textitProductivity & Project Management: basecamp, clicksend, clientary, dart, fibery, monday, notion, onedesk, productboard, rocketlane, todoistDeveloper Tools & APIs: algolia, anonyflow, api_ninjas, api_sports, apify, appdrag, backendless, browserless, bubble, cloudconvert, cloudinary, cloudlayer, convertapi, databricks, datadog, datarobot, deepseek, deployhq, digital_ocean, docmosis, docugenerate, encodian, gitea, gitlab, globalping, groqcloud, hookdeck, hyperbrowser, imgbb, imgix, kibana, neutrino, npm, openai, openrouter, parsera, parseur, phantombuster, pinecone, prismic, procfu, replicate, scrape_do, serpapi, shotstack, snowflake, supabase, tavily, v0, vercel, writer, zenrowsE-commerce & Payments: brex, btcpay_server, coupa, flutterwave, gift_up, lemon_squeezy, quaderno, ramp, shopify, stripe, zoho_invoiceHR & Recruiting: ashby, bamboohr, recruiteeData & Analytics: amplitude, census_bureau, college_football_data, currencyscoop, diffbot, ip2location, mixpanel, nasa, rosette_text_analytics, securitytrails, textrazor, twelve_dataDocuments & Files: carbone, doc_certs, documenso, dropbox, excel, files_com, grist, pdf_co, share_pointDesign & Media: canva, canvas, claid_ai, deepimage, heygen, metatextaiCustomer Support: freshdesk, retently, servicem8, sevdesk, storeganiseCalendar & Scheduling: calendly, deadline_funnel, etermin, googlecalendarSocial Media: facebook, instagram, redditLocation & Maps: addresszen, geoapify, google_maps, mapboxEmail Verification & Validation: clearout, icypeas, neverbounce, zerobounceOther Integrations: bitwarden, canny, cardly, castingwords, confluence, formdesk, getform, habitica, headout, highergov, jira, keen_io, landbot, moonclerk, one_drive, outlook, googlesheets, resend, ritekit, sms_alert, tapfiliate, thanks_io, uptimerobot
Before vs After
For example, theGOOGLECALENDAR_GET_CURRENT_DATE_TIME request schema changes: Previous (only a single type):
page_token in GOOGLECALENDAR_LIST_CALENDARS:
Previous:
Why This Matters
- Accurate schemas: LLMs and code generators see the full set of allowed types
- Better validation: Input validation can now correctly accept all valid types, not just the first one