Connect your account, work with your boards, and use supported image, video and speech models—without moving files between apps.
Connect
Follow one setup path and approve access. You do not need to open a board to connect.
Manual setup: If the link does not open the form, go to Customize → Connectors → + → Add custom connector, or the equivalent in your version. Enter FlashBoards and the server URL above. Leave optional client ID, client secret, and headers empty for automatic registration.
If tools do not appear after approval, check the connector state and retry native tool discovery before restarting Claude.
Connect through ChatGPT Plugins. This creates a private custom MCP connection, not a public directory listing or an OpenAI-reviewed plugin. Availability and allowed actions depend on your account and workspace policy; FlashBoards sign-in and any required administrator permission still apply.
Seeing the legacy Apps interface instead? Follow the controls shown in your account, such as Apps → Create, or, in managed workspaces, Workspace settings → Apps → Create and require admin-granted access. Stop if your account lacks permission; workspace publication is not needed for this private connection.
Previously failed to connect? An old FlashBoards connector created before the OAuth registration fix may need to be deleted and recreated for fresh registration. Only reset that failed connector—not a working connection. Never reuse a failed authorization code or paste credentials into chat.
Tool availability and in-chat displays depend on your ChatGPT account and client. Verify access in the chat you intend to use. A successful connection does not verify paid generation, live-board edits, token refresh, or native desktop support.
For direct video generation, use generate_video with a compatible Seedance model. For Topaz Starlight source-video enhancement, use estimate_video_enhancement, then enhance_video with one videoFile or inputReceipt and a stable generationAttemptId. ChatGPT attachments can be forwarded directly; if forwarding is unavailable, open show_media_upload_picker and pass its opaque receipt. Upload alone does not generate or charge; the board flow is optional.
Official ChatGPT connection guide ↗ · Workspace permissions ↗
claude mcp add --transport http flashboards --scope user https://flashboards.yaroflasher.com/api/mcpSecret storage — replace the placeholder locally, not in chat:
FLASHBOARDS_API_KEY=YOUR_FLASHBOARDS_API_KEYServer configuration — merge with existing servers:
mcp_servers:
flashboards:
url: "https://flashboards.yaroflasher.com/api/mcp"
headers:
Authorization: "Bearer ${FLASHBOARDS_API_KEY}"This is the API-key path, not a claim that Hermes only supports API keys. The agent guide includes an interactive OAuth recipe; do not layer a Bearer header over an existing OAuth connection.
Using Cursor, VS Code, Codex, or interactive Hermes? Read the client-specific agent guide ↗. Those recipes are documentation-based, not a claim that every client has been live-tested.
Available to active FlashBoards accounts. Client availability and workspace permissions may differ. Already connected? Use your existing connection—no need to add it again.
Once connected
Download user skills for characters, video, audio, advertising and upscaling or discover them here with list_skills and get_skill.
For reusable local-board behavior, call list_skills and get_skill with flashboards-sdk, or read /skills/flashboards-sdk/SKILL.md. Follow the live describe schemas. An optional panel can visualize headless behavior; a live eligible board tab and human Settings consent still gate writes and invocation.
Say “flashboards” to open the start menu: pick a mode (Reels, Music video, Cartoon, Ad, Character, Animate, Film scene, Upscale, Avatar, YouTube & podcast, Music) and then what to make, and the agent continues with that mode's method. MCP App hosts such as Claude and ChatGPT show the menu as a card; other clients get the same choices as a numbered list. In Claude Code, run /flashboards:start. Motion design drawn by code is offered to clients that run local commands. Opening the menu is free.
Your photos, videos and audio reach FlashBoards as short-lived receipts that generation, finishing, enhancement and audio tools accept — never as bytes in the conversation. In ChatGPT, attach the file: import_chat_file (or a tool's own file field) takes it. Elsewhere the agent calls create_upload_link for a one-time link: Claude Code, Codex and Claude's code sandbox upload the file themselves with curl; otherwise you open the link's upload page, drop the file, and the agent picks it up with get_upload. MCP App hosts also offer the in-chat picker. Images up to 50 MiB (larger photos are downscaled to 16 megapixels; HEIC is not supported), video up to 100 MiB, audio up to 15 MiB and 5 minutes. Uploading is free.
Browse published shared boards, inspect a live board, and apply precise edits. Keep your FlashBoards board tab open when using live-board tools.
Use generate_audio for ElevenLabs v3 or v4 speech without an open board; estimate_generation_cost with kind audio and narration text quotes the charge before running. Discover supported models, check inputs and prices, start image, video and speech generations, and follow their status. ChatGPT hosts with file input support can pass imageFile to generate_image, imageFile or videoFile to compatible Seedance generate_video models, and a file to describe_media; MCP App hosts can instead pass one show_media_upload_picker image receipt in referenceImageReceipts to a compatible generate_image model. These paths need no open board. For Topaz Starlight enhancement, estimate_video_enhancement and enhance_video accept one verified videoFile up to 8 MiB or an uploaded inputReceipt up to 100 MiB; this is source-video enhancement, not prompt generation. validate_chat_media checks one host file without upload or paid work. Seedance also accepts one local image attachment or caller-bound catalog preview alongside supported image URLs. MiniMax H3 and H3 Max use the board path for attached images and previews: add_media_to_board, prepare_board_generation with the returned itemId and discovered slot, then generate_on_board after consent. Catalog previews use their preview rendition only. Local MP4 board import accepts a data URL up to 10 MiB encoded when the caller can access its bytes; catalog previews for image generation also use the board path.
For requests like 'show me ten sky references', use the MCP-only show_reference_gallery tool to display images together on compatible MCP App hosts, with text links as fallback. The default is six; the agent can set limit to any integer from 1 to 15 to match the requested count. Fewer matches may be returned, and preview failures are shown per card. All active authenticated accounts can also use search_references for prompt search, search_references_by_image for validated prepared JPEG/PNG/WebP bytes up to 327,680 bytes, and get_reference_preview for one on-demand native preview. When configured, previews include signed HTTPS links valid for 24 hours: anyone with a link can read the preview without cookies until expiry, independently of the source handle or issuing connection. Final rendering is client-controlled. Run the search again after expiry; the gallery never renews links automatically. These tools need no open board or browser session, return no private source URLs or originals, and share the Website's per-account catalog limits.
All active authenticated accounts can browse the video reference library in FlashBoards or use list_video_reference_categories, search_video_references, and get_video_reference through MCP. Search matches titles and English/Russian labels; category, term, and MP4/animated WebP filters are available. get_video_reference returns a cookie-free HTTPS playback link valid for up to 24 hours when delivery is configured. Anyone with the link can play it until expiry, subject to current Report exclusions and the issuer's active account. Client video rendering varies. Report remains limited to admins and testers.
Review OAuth access and revoke connected apps or API keys from Settings → Developer. Connections act as your account.
Simple billing
Connecting and working with boards is free. AI image, video and speech generation uses your FlashBoards balance. Your AI client's own subscription or API charges are separate.
Safety
Troubleshooting
Check the connection is enabled and selected in the chat. Retry native discovery or the client's targeted refresh before restarting.
Check Plugins → plus button → Add custom MCP server, account eligibility, and workspace permissions. Use the official instructions linked in the ChatGPT section.
For a pre-fix failed registration, recreate only that connector and complete fresh OAuth. Do not reset a working integration.
An initial OAuth challenge can be normal; persistent 401 needs authentication diagnosis. Use Streamable HTTP—not legacy SSE. Opening the endpoint in a browser is not a connection test.
Connect once. Verify access. Then choose what to create.
Open FlashBoards →