AI Agent Connector
Snipish exposes a remote Model Context Protocol server so an agent can capture, upload, inspect, annotate, and share screenshots in your workspace.
Connector URL
https://snipish.mfdi.me/mcpConnect Claude
- Open Claude Settings, then Connectors.
- Add a custom connector and enter the connector URL above.
- Select Connect and sign in to Snipish.
- Review the requested permissions and select Allow access.
Claude registers itself automatically. Snipish uses OAuth 2.0 authorization code flow with PKCE and rotates refresh tokens, so no API key needs to be pasted into Claude.
Give the agent one prompt
Run the built-in snipish-agent-setup MCP prompt once. It gives the agent the ingest decision tree, annotation rules, verification step, and publish-only-when-asked guardrail.
Choose the lowest-token ingest path
| Public web page | snipish_capture_url |
| Public or presigned image URL | snipish_import_screenshot_url |
| Local or private image file | snipish_create_screenshot_upload then raw HTTP PUT |
| Client cannot perform HTTP PUT | snipish_upload_screenshot base64 fallback |
Binary upload URLs expire after ten minutes and work once. Their JSON response contains the preview resource URI, collection item ID, annotation revision, and the exact next action.
Choose how the agent annotates
| Editable layers | Upload the clean image with annotationState: "unannotated", inspect it, then add normalized rectangles, arrows, and text through MCP. |
| Already annotated | Let the agent draw with its own image tool, verify the result, then upload it with annotationState: "flattened". Snipish will not tell the agent to recreate the same marks. |
Snipish stores this state with the screenshot so later agents know whether markup is baked into the pixels. The annotate-screenshot MCP prompt teaches both modes, including contrast, label placement, coordinate conversion, and visual verification.
Agent workflows
- Capture a public URL into the reusable AI Captures collection.
- Import image URLs or upload raw binary files without putting base64 in the MCP call.
- Read model-sized previews, originals, collection manifests, and annotation state.
- Add editable layers, or upload locally drawn markup as a verified flattened image.
- Publish a finished collection and return its clean share URL only when the user asks.
Annotation writes use a revision number. An agent must read the latest item before changing it, which prevents silent overwrites when a person and agent edit at the same time.
Permissions
media:read | View screenshots and their metadata |
media:write | Upload screenshots to your workspace |
capture:create | Capture public web pages for storage in Snipish |
collections:write | Create collections and place screenshots |
annotations:write | Add, update, and delete annotations |
Bearer-token fallback
For an MCP client without OAuth, create an API key under Dashboard, Settings, Integrations and send it as Authorization: Bearer snp_.... Treat that key as a secret; current API keys grant the complete MCP permission set.