MCP

PaperFox provides a Model Context Protocol (MCP) server that allows AI assistants like Claude and ChatGPT to interact with your conference data securely.

With PaperFox MCP, your AI assistant can:

• List conferences you have access to
• Get conference and track statistics
• Monitor review progress
• Query submission data
• Read reviewer scores and comments to help shortlist or triage submissions

All actions respect your existing permissions—the AI can only access data you're authorized to see.

Available Tools

Setup Instructions

PaperFox MCP URL:

https://www.paperfox.ai/mcp

Claude

Set up PaperFox as a custom connector on claude.ai. Remote MCP servers use Connectors; the desktop JSON config file is for local servers.

1. Open Customize → Connectors in Claude.
2. Click + Add → Add custom connector.
3. Enter PaperFox as the name and https://www.paperfox.ai/mcp as the remote MCP server URL, then click Continue.

4. On the authentication screen, select Sign in now and Register automatically under OAuth client. Claude should detect these settings automatically; you do not need to enter a client ID or secret. Click Add.

5. When prompted, sign in to PaperFox, review the requested permissions, and authorize access.

6. Start a conversation with PaperFox enabled and ask: "What conferences do I have access to?"

For more details, see Claude's custom connector guide.

ChatGPT

PaperFox is not yet listed as a public plugin in OpenAI's directory. Connect it by URL using Add custom MCP server. ChatGPT saves this custom connection as a plugin for your account or workspace.

1. Open ChatGPT Plugins in your web browser.
2. Click the + button, then Add custom MCP server.
3. Fill in:
   • Name: PaperFox
   • Connection: leave the toggle on Server URL and enter https://www.paperfox.ai/mcp
   • Authentication: OAuth

4. Review the risk warning, check I understand and want to continue, and click Create as a plugin (the button is enabled once the box is checked). ChatGPT then prompts you to connect PaperFox. Click Connect PaperFox, sign in to PaperFox if needed, and authorize access.

5. Find PaperFox under your Personal plugins (or in the workspace where you created it). If it is not already installed, click Install.
6. Start a conversation, type @, select PaperFox, and ask: "What conferences do I have access to?"

For more details, see OpenAI's Add custom MCP server guide.

Other MCP-Compatible Clients

Any MCP-compatible client can connect using:

• MCP Endpoint: https://www.paperfox.ai/mcp
• Transport: Streamable HTTP
• Protocol revision: 2026-07-28. Clients on the 2025 revisions are still served from the same endpoint, so no client needs to upgrade to keep working.
• Authentication: OAuth 2.1 with PKCE
• Scope: mcp:tools

The client will automatically discover OAuth endpoints via:

• https://www.paperfox.ai/.well-known/oauth-protected-resource
• https://www.paperfox.ai/.well-known/oauth-authorization-server

Usage Examples

Example 1: Discover your conferences

Prompt: "What conferences do I have access to?"

What happens:

• PaperFox returns all conferences where you have a role (chair, reviewer, author, etc.)
• Shows conference name, dates, location, and your specific roles
• Use this to discover conference slugs needed for other tools

Example response:

You have access to 2 conferences:

1. DSW 2027 (Data Science Workshop 2027)
   - June 15–18, 2027 · Boston, MA
   - Your roles: Conference Chair

2. ICMR 2027 (International Conference on ML Research)
   - September 10–12, 2027 · London, UK
   - Your roles: Reviewer

Example 2: Check submission and review statistics

Prompt: "Show me the submission statistics for DSW 2027 and how reviews are going"

What happens:

• PaperFox retrieves conference-level stats: total submissions, status breakdown (submitted, under review, accepted, rejected), and reviewer counts
• Then fetches review progress: assignments completed vs. pending, completion rate, and reviewers with outstanding work
• Requires Conference Chair or Track Chair role

Example response:

DSW 2027 Statistics:
- 142 total submissions (28% of 500 limit)
  - 45 submitted, 67 under review, 18 accepted, 12 rejected
- 24 active reviewers, 89 registrations

Review Progress:
- 216 total assignments, 148 completed (68%)
- 3 reviewers with 5+ pending reviews

Example 3: Look up a specific submission

Prompt: "What's the status of submission dsw2027-42?"

What happens:

• PaperFox retrieves submission details: title, abstract, status, version history
• For chairs: shows authors, reviewer assignments, and review status
• For reviewers: shows submission content (authors hidden in double-blind review)
• For authors: shows their own submission with revision status

Example response:

Submission dsw2027-42: "Efficient Transformer Architectures for Edge Devices"
- Status: Under Review (version 1)
- Track: Main Track
- Submitted: March 1, 2027
- Authors: Jane Smith (MIT), John Doe (Stanford)
- Keywords: transformers, edge computing, model compression
- 2 of 3 reviews completed

Example 4: Shortlist submissions based on reviews

Prompt: "For CSWIM 2026, pick 5 accepted submissions that look strongest based on reviewer signal and content, and give me a short rationale for each."

What happens:

• The AI calls list_submissions once with decision=['ACCEPT','CONDITIONAL_ACCEPT'] and includeReviewSignals=true — this returns abstracts, keywords, and compact review aggregates (review count, avg/min/max scores, recommendation values) for every accepted submission in a single call.
• The AI reasons over content + aggregate scores to pick a shortlist.
• For each shortlisted submission, it calls get_submission_reviews to read the full reviewer comments so its rationale cites specific reviewer language rather than hallucinating one.
• Requires Conference Chair role (or Track Chair role for submissions in your track).

This pattern keeps the big scan cheap (one call, no long comment text) and reserves the expensive full-review reads for the N submissions you actually want to talk about.

Example 5: Read and adjust the review form

Prompt: "What are reviewers actually asked on ICMR 2027's review form? Add a question about ethics approval."

What happens:

• The AI calls get_review_form to read the conference-level default fields — the questions every track inherits. Adding a trackSlug instead returns that track's own extra fields plus the resolved form reviewers actually see.
• To make the change, it calls update_review_form with the complete field list (the existing questions plus the new one). The tool replaces the list it is given, so anything omitted is removed — which is why the AI reads the form first.
• The first call returns a preview with a field-level diff (added / removed / modified / reordered) and writes nothing. The AI shows you that diff; only after you agree does it call again with confirm: true.
• Once reviews have been submitted, the conference-level form becomes append-only: you can add questions and reword existing labels, but existing questions can no longer be removed or reordered. The tool refuses those edits rather than silently invalidating reviews already written.
• A track-level edit can't add a question whose id is already inherited from the conference form — both would write to the same answer. The tool refuses the save and suggests either a different id for the track question or editing the conference form instead (omit trackSlug).

You can make the same edits by hand — see Configure Review Forms →.

Example 6: Work from your own files

Prompt: "Draft a program like this one for ICAI 2027, using this year's accepted submissions" — with last year's program PDF attached.

What happens:

• Your AI assistant reads the attached file itself. PaperFox never receives the file — only the tool calls the AI makes from it.
• The AI calls list_submissions and get_program_schema, drafts a program in the same shape, and shows it to you. import_program_draft creates it as an unpublished draft only after you confirm.
• Other files work the same way: attach a spreadsheet of tracks and ask the AI to create them — it calls create_track once per track.
• Whether you can attach files depends on your AI client, and what the AI reads from them is sent to your AI provider like the rest of the conversation.

Usage Limits

Every authenticated user gets 50 free MCP tool calls per day. After that, each additional call costs 0.5 credits (at $1 = 300 credits, that's about $0.002 per call).

A typical session uses 5–15 tool calls, so most users will stay well within the free tier.

If you run out of free calls and don't have enough credits, the AI will let you know and provide a link to purchase credits. Credit usage appears in your credit history as "MCP Tool Call" entries.

The daily limit resets at midnight UTC.

Authentication Flow

When you first connect an AI assistant to PaperFox:

1. The client discovers the OAuth server automatically
2. A browser window opens for you to sign in to PaperFox
3. You review and approve the permissions requested
4. The client receives secure tokens to access your data
5. Tokens refresh automatically—you won't need to re-authenticate frequently

Your credentials are never shared with the AI assistant. Only secure, scoped access tokens are used.

Security

• OAuth 2.1 with PKCE: Industry-standard secure authentication
• Scoped Access: AI only accesses what you authorize
• Role-Based: Respects your existing PaperFox permissions
• Minimal Write Access: Most tools are read-only; write tools (track and program management) are chair-only
• Confirm-First Writes: Every write tool is two-step — without confirmation it only returns a preview, and nothing is created, changed, or deleted until you approve. MCP can never publish your program or make anything public; programs it creates are unpublished drafts
• Token Rotation: Refresh tokens are rotated for security
• Audience Validation: Every token is bound to PaperFox and rejected anywhere else (RFC 8707)
• Issuer Identification: Authorization responses name PaperFox as the issuer, so your client can prove which server answered before it redeems a code (RFC 9207)

Privacy

PaperFox MCP only provides access to data you already have permission to see in the PaperFox web interface. No additional data is collected or stored when using MCP. See our Privacy Policy for details.

Troubleshooting

"Access Denied" errors

Ensure you have the required role for the tool you're trying to use. For example, get_conference_stats requires Conference Chair access.

Connection fails

1. Check that you're signed in to PaperFox
2. Verify the MCP URL is correct: https://www.paperfox.ai/mcp
3. Try removing and re-adding the server configuration
4. Check your network connection

Token expired

Most clients handle token refresh automatically. If you encounter persistent auth errors, remove the server configuration and re-add it to trigger a fresh authentication.

Claude: "Server not connecting"

Add PaperFox through Customize → Connectors → + Add → Add custom connector, with Sign in now and Register automatically selected. The claude_desktop_config.json file is only for local MCP servers.

ChatGPT: "Add custom MCP server" missing

Open ChatGPT Plugins in a web browser and click +. If Add custom MCP server isn't listed, your workspace permissions or security restrictions block custom MCP servers — check with your administrator.

ChatGPT: PaperFox not shown under @

Find PaperFox under your Personal plugins (or the workspace where you created it). If it is not installed, click Install; if it is not connected, complete the PaperFox sign-in and authorization. Then type @ in a conversation and select PaperFox.