Claude or ChatGPT for Identity Workflows: What Actually Differs

Key takeaways

• Claude and ChatGPT can connect to the same Didit Model Context Protocol (MCP) service, but they do not receive the same public tool contract.
• Claude’s standard hosted connection at https://mcp.didit.me/mcp exposes 115 tools. Didit’s OpenAI app surface uses https://mcp.didit.me/mcp/console, a restricted 101-tool catalogue.
• The difference is deliberate endpoint policy, not a general MCP limitation in ChatGPT. The console profile removes raw personal-data submission and correction tools while preserving most investigation, workflow, case, transaction, and reporting operations.
• The ChatGPT catalogue also removes portrait_image from session creation and document_number from Anti-Money Laundering screening. The tools remain available, but those inputs are not advertised or accepted through that profile.
• Both hosted paths use OAuth (Open Authorization) 2.1 with Proof Key for Code Exchange (PKCE) and Dynamic Client Registration. Neither asks the user to paste a hosted-server application key into chat.
• Choose Claude when an authorized operator needs the broader hosted catalogue. Choose ChatGPT when the 101-tool reviewed surface covers the workflow and tighter data-input boundaries are desirable.

“Does it work in Claude or ChatGPT?” sounds like a compatibility question. For identity workflows, it is really a question about which catalogue, inputs, and release controls the client receives. Both clients can speak to a remote MCP server. The practical difference is the contract Didit exposes to each surface.

This comparison focuses on that decision. It does not repeat the click-by-click setup in the Claude installation guide, the ChatGPT-specific walkthrough in the Know Your Customer (KYC) with ChatGPT guide, or every operation in the Didit MCP tools reference.

The choice in one table

Question	Claude	ChatGPT / OpenAI app surface
Connection	Native remote custom connector in Claude web and desktop; remote HTTP registration in Claude Code	Remote MCP connection through ChatGPT Developer Mode; reviewed app/plugin distribution for public availability
Didit endpoint	https://mcp.didit.me/mcp	https://mcp.didit.me/mcp/console for the reviewed OpenAI surface
Visible tools	115	101
Sensitive-input policy	Broader hosted operator surface	Fourteen additional tools removed; two retained tools have restricted parameters stripped
Authentication	OAuth 2.1, PKCE, Dynamic Client Registration	OAuth 2.1, PKCE, Dynamic Client Registration
Best fit	Broader authorized operations, including direct document and biometric checks	Investigation, orchestration, and operational workflows that fit a public-app review boundary

The important distinction: ChatGPT is not inherently limited to 101 Didit tools. Didit deliberately points its reviewed OpenAI app surface at a restricted catalogue. A private Developer Mode connection should use that same console endpoint when the goal is to reproduce the public app contract.

How Claude connects in practice

Claude supports remote custom connectors as a native account feature. Add the Didit connector in Claude web or Claude Desktop and the connection is brokered through Anthropic’s cloud. The same authorized connector can then be enabled for a conversation. For the shortest path, use the Didit connector deep link.

Claude Code can register a remote HTTP server with one command-line interface (CLI) command:

claude mcp add --transport http didit https://mcp.didit.me/mcp

Authorization then happens through the client’s MCP connection flow. That command is the useful difference for developers who want identity operations available beside source code, logs, and test fixtures. The detailed authentication and scope steps belong in the installation guide rather than in this comparison.

Through the standard hosted path, Claude sees 115 tools. Examples include didit_context_get for organization and application discovery, didit_session_create and didit_session_get_decision for verification sessions, didit_verify_aml for Anti-Money Laundering (AML) screening, didit_verify_kyb_search for Know Your Business (KYB) registry search, and didit_transaction_screen_wallet for Know Your Transaction (KYT) wallet screening.

The broader catalogue matters when the operator intentionally needs direct checks that accept identity-document or biometric inputs. It does not grant extra organizational authority: every call still runs as the signed-in Didit user, and backend roles continue to decide what that person may read or change.

How ChatGPT connects in practice

ChatGPT can connect to a public Streamable HTTP MCP endpoint in Developer Mode. The current OpenAI flow adds the server from ChatGPT’s plugin/app settings, discovers its tool metadata, and lets the developer test tool selection, arguments, results, and approval behavior. For public distribution, OpenAI scans the server and stores the reviewed metadata as a versioned snapshot.

Didit’s reviewed OpenAI surface uses https://mcp.didit.me/mcp/console. It is the same service and authorization system, but the route selects the console profile. That profile exposes 101 tools and enforces its restrictions both when tools are listed and when calls are dispatched. A client cannot recover an excluded operation by guessing its name.

This design has two consequences. First, a team can test the public contract in Developer Mode before review by connecting to the console route. Second, catalogue or schema changes for a published OpenAI app are release events: they must be scanned, reviewed, and published as a new metadata version. That is more controlled than assuming every live schema change should appear immediately in a public app.

Where the 14-tool difference comes from

The full self-hosted Didit catalogue contains 121 tools. The normal hosted OAuth catalogue removes six before either public surface is considered: four unauthenticated account-bootstrap operations, live application-credential revelation, and digital-credit top-up. Public app review does not permit digital-credit checkout flows, and live credential exposure is inappropriate for a hosted conversational catalogue. That leaves the 115-tool hosted surface used by Claude.

The OpenAI console profile removes 14 more tools. They fall into three practical groups:

• Direct document and biometric checks: didit_verify_id, didit_verify_poa, didit_verify_passive_liveness, didit_verify_face_match, didit_verify_face_search, didit_verify_age, and didit_verify_database.
• Contact challenge operations: didit_verify_email_send, didit_verify_email_check, didit_verify_phone_send, and didit_verify_phone_check.
• Raw upload and reviewer correction operations: didit_lists_entry_upload_face, didit_session_update_data, and didit_session_update_poa_data.

The pattern is more informative than the count. ChatGPT’s app surface keeps tools suited to conversational investigation and orchestration, while removing operations that invite raw biometric, document, contact-challenge, or reviewer-correction data directly through chat.

The two parameter-level restrictions

Removing entire tools would be too coarse for two important workflows, so the console profile applies field-level restrictions:

• didit_session_create stays available, but its portrait_image property is removed. ChatGPT can create a workflow-backed verification link without accepting a portrait through the app contract.
• didit_verify_aml stays available, but its document_number property is removed. ChatGPT can screen a person or company using fields such as name, entity type, date of birth, and nationality without submitting that government identifier.

These are not cosmetic schema changes. The server removes the properties from the advertised input schemas and drops them if they appear in a console-profile call. This makes the restriction enforceable even when a model or user supplies an unexpected argument.

What both clients can still do

The 101-tool surface is not a read-only demo. Both clients can discover workspace context, create workflow-based sessions, retrieve decisions, perform AML screening within the available fields, search and select KYB records, inspect transactions, screen wallets, work with cases, manage workflows and webhooks, and query analytics. Exact availability still depends on the connected user’s Didit role.

For example, both surfaces retain didit_context_get, didit_session_create, didit_session_get_decision, didit_verify_aml, didit_verify_kyb_search, didit_verify_kyb_select, didit_transaction_screen_wallet, didit_case_manage, didit_workflow_create, didit_webhook_create, and didit_analytics. See the tool documentation for the canonical schemas and the workflow examples for composed operations.

Both paths use the hosted, stateless, POST-only Streamable HTTP service. Both rely on delegated user authorization rather than a pasted hosted-server secret. Both expose tools only, not MCP resources or prompt templates. The server itself is free; any paid verification or fraud operation retains its normal Didit price.

How to choose

Choose Claude when the broader operator surface is required

Claude is the better fit when an authorized user needs direct document, biometric, proof-of-address, database-validation, contact-verification, or reviewer-correction operations from the assistant. It is also a natural choice when Claude Code is already the working environment and the developer wants one remote endpoint beside the repository.

Choose ChatGPT when the reviewed contract already covers the job

ChatGPT is a strong fit for session orchestration, decision review, AML and KYB investigation, transaction and wallet analysis, case operations, workflow inspection, webhooks, and analytics. The 101-tool profile is especially useful when a team wants a public app boundary that minimizes raw personal-data entry without reducing the experience to passive lookup.

Choose the application programming interface for deterministic production automation

Neither conversational client should replace backend code for a product flow that must execute predictably on every user request. Use Didit’s REST application programming interface (API) or software development kits for application-controlled automation, and use the Business Console for governed human configuration. MCP is most valuable when a signed-in person wants an AI client to investigate or execute bounded work on their behalf.

A practical verdict

Pick the client after mapping the workflow to the available contract. If it needs any of the 14 raw-input or correction tools, use Claude’s 115-tool hosted catalogue or move the deterministic operation into application code. If the job fits the 101-tool console profile, ChatGPT provides the same core identity and fraud service with a narrower conversational data boundary.

The implementation is inspectable in the public MIT-licensed GitHub repository. For transport and authorization details, read the Didit MCP overview and authentication guide. For the wider developer surface, visit Didit MCP for developers.