Your AI assistant can work in Synopsule
Search saved transcripts, then opt in to recording, editing, analysis, Library workflows, playback and exports from a compatible AI assistant. The MCP server is built into the Mac app. Connect Claude Desktop, Claude Code or Codex in Settings → Connect to AI Apps; turn on Allow AI app control and keep Synopsule open for actions. Control is off by default. Audio, voiceprints, credentials, calendar contents and live transcript drafts are never returned. The connected client controls where text it reads goes.
The command every client uses:
/Applications/Synopsule.app/Contents/MacOS/Synopsule --mcp
Synopsule → Settings → Connect to AI Apps shows ready-to-paste setup for each client below, with the exact path filled in wherever the app is installed.
On this page
Claude Desktop — two clicks
In Synopsule, click Install… on the Claude Desktop row. Synopsule generates its extension (Synopsule Transcripts.mcpb, saved to Downloads) and opens it in Claude, which shows its Install sheet — click Install, done. The extension contains no binary, just a tiny launcher that runs the Synopsule app in --mcp mode, so it never needs separate updates.
Manual fallback: Copy Config in Synopsule, then Claude → Settings → Developer → Edit Config, and paste under mcpServers:
"synopsule": {
"command": "/Applications/Synopsule.app/Contents/MacOS/Synopsule",
"args": ["--mcp"]
}
Codex (app, CLI, and IDE) — one click
All Codex surfaces share one config file, so this covers the macOS app, the CLI, and the IDE extension at once: click Set Up… on the Codex row and grant access to the .codex folder when asked. Synopsule adds itself to ~/.codex/config.toml; new Codex sessions see it immediately (the desktop app lists it under Settings → Integrations & MCP).
Terminal fallback:
codex mcp add synopsule -- "/Applications/Synopsule.app/Contents/MacOS/Synopsule" --mcp
Claude Code
claude mcp add synopsule -- "/Applications/Synopsule.app/Contents/MacOS/Synopsule" --mcp
ChatGPT
ChatGPT's custom connectors only accept public HTTPS MCP servers — they run from OpenAI's cloud and cannot reach a process on your Mac. Two options:
- No-setup path: turn on Synopsule's optional export folder (Settings → Connect to AI Apps → Export folder) and drag transcript Markdown files into the chat. Works on every plan.
- Power-user path (tunnel): expose the server over a temporary HTTPS tunnel, at the cost of making it reachable outside your Mac while the tunnel is up:
# Terminal 1 — bridge stdio → HTTP (any stdio-to-HTTP MCP bridge works):
npx -y supergateway --stdio "/Applications/Synopsule.app/Contents/MacOS/Synopsule --mcp --read-only" --port 8808
# Terminal 2 — public HTTPS tunnel:
cloudflared tunnel --url http://localhost:8808
Then in ChatGPT: Settings → Connectors (developer mode) → add the https://….trycloudflare.com URL. Anyone with that URL can read your transcripts while the tunnel runs — treat it accordingly and shut it down when done.
From a prompt to a finished workflow
Try “Start a meeting recording”, then “Rename the speakers, add a note at two minutes and export a Word document.” Or “Find the Cedar discussion and create a project note from the decisions.” Existing transcript access still works while the app is closed. App control uses the running Mac app and its existing permission, consent and storage rules.
38 tools: five for transcript retrieval and 33 for app control. File import/export and backup/restore use native file choosers. Deleting a recording or Library item asks for confirmation in the app. macOS permissions, AI destination consent, account/sync choices and purchases stay under your control.
| App-control tool | Purpose |
|---|---|
synopsule_status | Get real capture sources, recording state, analysis progress, current library, and pending user attention. No live transcript text. |
get_operation | Read the result of a previously accepted operation. Never repeat a pending mutation. |
list_recordings | List current library recordings, including analysis state. Returns a bounded page; get_recording supplies revision and editable IDs. |
get_recording | Read finalized recording text, notes, flags, speakers, action items and revision. Excluded/private sources and live drafts are withheld. Page transcript segments with offset/limit. |
start_recording | Start capture using configured audio sources and existing macOS permission/consent flows. Ask the user before recording. Reports actual sources; never returns captured audio. |
pause_recording | Pause a recording that is currently recording. |
resume_recording | Resume a recording that is currently paused. |
stop_recording | Stop capture and finalize locally. Returns operation_id while analysis runs. The normal keep/delete audio prompt remains. |
analyze_recording | Analyze or reanalyze retained local audio, including transcription and speaker separation. Existing app admission and model availability apply. |
summarize_recording | Generate/regenerate a recap with the configured AI destination and optional template. Normal destination consent, subscription and completion-preservation rules apply. Ask before a paid run. |
list_summary_templates | List all recap templates with their IDs and descriptions. |
edit_recording | Edit title/tags/transcript/speaker labels through the app's persistence APIs. Supply fields for exactly one operation; no arbitrary session JSON. |
annotate_recording | Add/edit/delete timestamped notes and flags, or keep/exclude a note as AI context. IDs and revision come from get_recording; timestamp is in seconds. |
set_action_item | Set completion for an action item from get_recording. Persists to the authoritative library. |
delete_recording | Permanently delete one recording and its local audio. Requires confirm=true and a matching revision; the app also asks the user to confirm. |
delete_recording_audio | Delete retained audio while keeping its transcript. Requires confirm=true and matching revision; the app also asks the user to confirm. |
export_recording | Open the native Save dialog for a text/document export (Word, Markdown, Obsidian, text, PDF, subtitles, HTML or Apple Notes). User chooses destination. No audio export. |
play_recording | Play/pause/seek retained audio on this Mac only. Audio bytes and paths never enter the response. Open the reader for visible playback controls. |
search_library | Search eligible notes, files, recordings, people, projects, collections and actions in the current unified Library. Excluded/cloud-private/conflicted content is withheld. |
list_library_items | List eligible authored notes, entities, actions, collections and records. Use get_library_item for revision and fields. |
get_library_item | Read an eligible authored Library item's editable fields and revision. Imported/generated evidence is read through search_library and the app, avoiding raw internal payloads. |
create_library_item | Create an authored note, person/project/topic, action, collection or collection record in the current local Library. Optional synchronization follows the user's existing choice. |
update_library_item | Update an authored Library item's title/text/status with conflict and library checks. Cannot change destination permissions, provenance, source blocks or identities. |
delete_library_item | Delete one authored Library item with matching revision and confirm=true. The app also asks the user. Source deletion uses the Library's tombstone/dependency rules. |
organize_library_item | Run the app's organization for an eligible source. Uses its configured destination, consent and evidence validation. Unavailable/flag-disabled engines stay unavailable. |
get_settings | Read safe capture/transcription/model preferences, installed model choices and automation capabilities. No credentials, calendar contents or voiceprints. |
update_settings | Update capture sources, audio retention, language/output mode, expected speaker count or installed transcription model. Permissions, provider secrets, sync/account choices and automation authorization require the app UI. |
import_file | Ask the user to choose one audio file, document or recording attachment in a native picker, then import it. The original is never moved/deleted. No arbitrary filesystem reads. |
backup_library | Create a portable local ZIP backup at a destination the user chooses in a native Save dialog. Optional audio stays in that user-created archive; no archive/audio bytes or paths are returned to MCP. |
restore_library | Ask the user to select a backup, inspect its manifest, then confirm a verified merge. Existing content/deletion history and selected-library boundaries are preserved. No blind overwrite or automatic origin erasure. |
manage_models | Select, download, cancel downloading or remove a transcription model. Downloads use the app's progress/cancel/free-space checks. Deletion needs confirm=true. No provider credentials are exposed. |
enrich_note | Suggest context for a note using the app's configured AI destination and existing consent. Suggestions remain pending until reviewed; no implicit keep. |
open_synopsule | Open a native app surface. File import, backup/restore, models/downloads, vocabulary, calendar, notifications, speaker recognition, AI/provider consent, iCloud, pairing and purchases remain user-operated flows. Returns requires_user_interaction, not completion. |
Check completion: actions first return accepted and an operation_id. Poll get_operation; acceptance does not mean completion. A result of requires_user_interaction needs the named action in Synopsule. For edits, read the current ID and revision first. Stale changes fail. After a timeout, inspect app state before repeating a change.
Native flows: open_synopsule opens vocabulary, speaker voices, collection/evidence review, calendar, notifications, iCloud/history, pairing and AI settings. These are user-operated flows, not unattended automation. Experimental features retain their existing gates. Library CRUD is limited to authored items; generated/imported evidence is reviewed in the app.
Turn control off any time in Settings. It stops new control requests and result access; work already performed is not undone. Remove the client connection to end its transcript access. Private/excluded sources are withheld from automatic text access. Custom-root and standalone sessions stay read-only; add --read-only to explicitly restrict a client. Any public bridge should be read-only and separately secured; the tunnel example above forces read-only mode.
Transcript retrieval tools
| Tool | Purpose |
|---|---|
list_transcripts |
Recent recordings with id, date, duration, speakers |
search_transcripts |
Full-text search with speaker/timestamp excerpts |
semantic_search_transcripts |
Search by meaning over each recording's AI recap — finds the "pricing tradeoffs" meeting even if those exact words were never said |
get_transcript |
Full Markdown of one recording (auto-chunked when long) |
get_summary |
Just the AI recap — summary, action items, follow-ups |
The two search tools complement each other: search_transcripts matches the exact words in the full transcript text, while semantic_search_transcripts ranks recordings by how close their AI recap is in meaning to the query, using Apple's on-device embedding model — nothing leaves the Mac. If the system embedding asset is unavailable, or no recap meets the similarity threshold, the tool falls back to exact-text search and reports that mode. A query can still return no matches. This tool is separate from the optional in-app search model in Synopsule.
Try: "Search my transcripts for the pricing discussion and summarize what we decided", or "What action items came out of yesterday's standup?"
Where the data lives
- The mirror Synopsule serves is inside the app's own storage and updates as recordings finish or change. The mirror also includes attached-file text and kept, current note context. Drafts, voice embeddings, speaker personas, and audio never enter it.
- The optional export folder (Settings → Connect to AI Apps) keeps an identical, human-browsable copy anywhere you choose — Finder, Obsidian, iCloud Drive, drag-into-chat.
- What an AI app does with transcript text you pull into a chat is governed by that app's privacy policy, not Synopsule's.
iPhone
The iOS app keeps the same Markdown export in Files → On My iPhone → Synopsule (Settings → Storage → Show in Files app). Share a transcript into the ChatGPT/Claude iOS apps from there, or put your Mac export folder in iCloud Drive to read Mac recordings on the go.
Troubleshooting
- The client can't find or start the server. The configured command must point at the Synopsule app binary on this Mac. Open Settings → Connect to AI Apps — every snippet there embeds the app's real installed path. If you've moved or reinstalled the app, run the one-click setup again (or re-paste the config) so the path updates, then restart the AI client.
- The tools answer, but say there are no transcripts. The mirror only contains finalized recordings — record something and let it finish processing. The server's own replies explain how to get transcripts flowing.
- ChatGPT won't add the server. ChatGPT connectors require a public HTTPS URL; a local command can't be added directly. Use the export-folder path or the tunnel recipe above.
- Semantic search reports a substring fallback. The system embedding asset is unavailable, or no recap passes the similarity threshold. The tool searches exact text and reports the fallback. This MCP tool is separate from the optional Search on this device model used inside Synopsule.
- Still stuck? Email [email protected].