Repeto
English
Open app
Documentation

Connect Repeto with MCP

Your learning collection, available safely inside an AI conversation.

Repeto is a remote Model Context Protocol (MCP) server. Add it once to any compatible client, such as Claude or ChatGPT, and the client can work with your Repeto collection when you ask it to. It supports Streamable HTTP, OAuth 2.1 with PKCE, structured tool results, resources, prompts, and portable background job controls.

Connect with OAuth

  1. In your MCP client, add a custom connector or remote MCP server.
  2. Use this server URL: https://api.repeto.ai/mcp
  3. Your browser opens Repeto. Sign in if needed, review the request, and choose Allow access.
  4. Return to the MCP client. Repeto is ready to use.

OAuth is the recommended setup. You do not need to create or paste an access token when your client supports it.

Permissions

Repeto asks for explicit, narrow permissions. Your MCP client may let you approve all of them or connect read-only:

  • collection:read — decks, cards, tags, due counts, and learning state.
  • cards:write — create, edit, move, suspend, or archive cards.
  • decks:write — create decks.

You can review and revoke connected apps at any time in Settings → AI connections.

Read and study tools

  • list_decks — deck totals, due cards, new cards, and suspensions.
  • search_cards — search fields and tags, optionally within one deck.
  • get_card — card content, tags, deck, and learning state.
  • get_deck — one deck's summary and an optional card sample.
  • get_learning_state — new, learning, review, mature, suspended, and due totals.
  • get_weak_concepts — concepts ranked by lapses and FSRS difficulty.
  • get_due_cards — cards due now for conversational study.
  • get_due_summary — due totals grouped by deck.

Create and improve cards

  • create_deck — create a deck by name, or return the existing match.
  • add_cards — add ready-made front/back cards without AI generation.
  • add_terms — generate vocabulary cards with the user's Repeto AI allowance.
  • create_cards_from_context — turn substantive conversation context into drafts.
  • preview_card_changes — preview text, tag, deck, suspension, and archive changes.
  • apply_card_changes — apply an approved, unexpired preview exactly once.
  • undo_mcp_action — restore a reversible MCP change from its activity receipt.

Why edits use preview and apply

  1. The assistant proposes changes with preview_card_changes.
  2. You see the before-and-after result in your MCP client.
  3. Only after you approve does it call apply_card_changes.
  4. Repeto returns an activity ID that can undo the change.

Previews expire after 15 minutes and refuse to overwrite a card that changed after the preview was made.

Background jobs

Large vocabulary batches and conversation-to-card generation return a job ID immediately. Use get_job to read its stage, percentage, result, or error, cancel_job to stop an unfinished job, and resume_job to safely retry from the persisted input. Resume is bounded by the job's retry allowance and waits for active cancellation to finish. These explicit tools work across MCP hosts today, including hosts that do not yet expose native MCP task controls.

Resources

  • repeto://today — today's due and learning-state totals.
  • repeto://decks/{deck} — a deck summary and card sample.
  • repeto://cards/{card_id} — one card and its learning state.
  • repeto://learning/weakest — the current weakest concepts.

Rich study dashboard

MCP Apps-compatible hosts can call render_study_dashboard to show a compact inline view of due cards, new and learning totals, mature cards, and weak concepts. It uses the standard ui.resourceUri contract and a versioned text/html;profile=mcp-app resource; ChatGPT compatibility metadata is included as well. Other MCP clients still receive the complete structured result.

Guided prompts

  • make-cards-from-this-chat — extract durable cards without saving chit-chat.
  • quiz-me — quiz from due cards without inventing review ratings.
  • improve-deck — inspect a deck and use the preview/apply workflow.
  • plan-todays-study — build a realistic plan from today's workload and weak concepts.

Quota

Every account receives 5,000 MCP operations per UTC day. This is a generous abuse-prevention ceiling, not a normal product limit. The current usage and reset time appear in Settings → AI connections.

MCP operations are counted separately from AI-generated cards, ingestion jobs, and research. For example, search_cards uses one MCP operation, while add_terms also uses the normal AI generation allowance for the cards it generates.

Safety and activity

  • Successful MCP writes and jobs appear in your recent activity.
  • Reversible writes include an Undo action in Settings and an MCP undo tool.
  • Activity stores compact identifiers and counts, not copied conversation transcripts.
  • Repeto never exposes a tool that submits review ratings. Official reviews stay user-controlled so FSRS history remains trustworthy.

Try it

After connecting, ask your AI something like:

Create a Repeto deck called “Networking” and add cards for the key ideas from this conversation.

Your MCP client will show you when it wants to call a Repeto tool. Review write requests before approving them. For a deck cleanup, ask: 

Find unclear cards in my “Networking” deck. Preview the improvements, but do not apply them until I approve.

Manual token fallback

If your client cannot open the OAuth browser flow, open Settings → AI connections, create a manual access token, and configure the client to send it as: 

Authorization: Bearer rcl_pat_…

Manual tokens act like passwords. Store them securely, never paste them into a conversation, and revoke unused tokens from the same settings page.

Troubleshooting

The connection or approval request expired

Return to your MCP client, remove the incomplete connection, and add https://api.repeto.ai/mcp again.

The client reports an authorization error

Reconnect with OAuth. If you use a manual token, confirm that the complete token is present in the Authorization header and has not been revoked.

A write tool says a permission is missing

The app was connected with a narrower grant. Revoke it in Settings, reconnect from the MCP client, and approve the required write permission.

The daily MCP ceiling was reached

Requests resume automatically at the next UTC midnight. Check the exact reset time in Settings → AI connections. Repeated polling is rarely necessary: background jobs expose their current progress.