Examples
These walkthroughs show what happens when you ask an assistant to work with TalkBar: which tool it calls, with which arguments, and what comes back. Ids, times and text are made up, and long answers are shortened, but every request and response has the shape TalkBar really uses.
Ask about last week's sessions
Setup: turn on MCP and connect Claude Desktop, as in Set up MCP. Reading is on by default, so nothing else is needed.
Then ask Claude: "What did I discuss in my TalkBar sessions last week?"
The assistant lists the sessions that started in that week with talkbar_read. Times are in milliseconds since 1970 (UTC); startedBeforeMs is exclusive.
{
"tool": "sessions.list",
"args": { "startedAfterMs": 1792368000000, "startedBeforeMs": 1792972800000, "limit": 20 }
}
{
"items": [
{
"sessionId": "0b7e4d1a-93c2-4f5e-8a61-2f9d3c7b5e08",
"title": "Weekly sync",
"status": "complete",
"startedAtMs": 1792747800000,
"endedAtMs": 1792749940000,
"projectId": "project-p8d3k2m1q9w7e5r4t6y0u2i1",
"seriesId": null,
"participantCount": 4,
"pinned": false,
"source": "live",
"private": false
},
{
"sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10",
"title": "Pricing review with Acme",
"status": "complete",
"startedAtMs": 1792594800000,
"endedAtMs": 1792597632000,
"projectId": "project-p8d3k2m1q9w7e5r4t6y0u2i1",
"seriesId": null,
"participantCount": 3,
"pinned": true,
"source": "live",
"private": false
}
],
"nextCursor": null,
"hasMore": false,
"totalCount": 2
}
Newest sessions come first. To read the summaries, the assistant batches one sessions.get per session in a single call:
{
"calls": [
{ "tool": "sessions.get", "args": { "sessionId": "0b7e4d1a-93c2-4f5e-8a61-2f9d3c7b5e08" } },
{ "tool": "sessions.get", "args": { "sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10" } }
]
}
The answer is {"results": [...]}, one entry per session, each with the full final summary, its sections, action items and decisions. Claude then writes you a recap.
Find a session and read its transcript in pages
Ask: "Find my pricing review with Acme and show me what Dana said about the discount."
First the assistant searches, to get the session's id:
{ "query": "pricing review Acme", "kinds": ["session"], "limit": 3 }
The first match has "kind": "session" and "id": "6f1c2a9e-…". Then it reads the transcript. A transcript comes in pages of whole turns, 20,000 characters each by default. speaker keeps one person's turns (any part of the name, ignoring case):
{
"tool": "sessions.transcript",
"args": { "sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10", "speaker": "Dana", "maxChars": 4000 }
}
{
"sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10",
"title": "Pricing review with Acme",
"startedAtMs": 1792594800000,
"private": false,
"truncated": true,
"excludedFromAiSpeakers": [],
"nextCursor": null,
"totalTurns": 38,
"turnOffset": 0,
"nextTurnOffset": 14,
"transcript": "[02:15] Dana Lee: We can go to twelve percent if the contract is two years.\n[02:41] Dana Lee: That's the most I can do this quarter.\n…"
}
truncated: true means there is more. The assistant passes nextTurnOffset back as turnOffset for the next page, and stops when nextTurnOffset is null:
{
"tool": "sessions.transcript",
"args": { "sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10", "speaker": "Dana", "maxChars": 4000, "turnOffset": 14 }
}
With "format": "json", the transcript comes back as turns instead, each with turnId, startMs and endMs (milliseconds from the session start), speaker and text. Use that form when you want to bookmark a turn or tie an action item to it.
The same read from a terminal:
talkbar-cli tools call sessions.transcript --args-file read.json
Create a note and an action item
Setup: in Settings → MCP Access, open the Notes row and switch on Create note, then open Action items and switch on Create action item (action_items.create). Leave the other writes off if you don't need them.
Ask: "Write up the follow-ups from the Acme pricing review as a note on that session, and give me an action item to send the revised quote by Tuesday."
The assistant sends both writes in one talkbar_change batch:
{
"calls": [
{
"tool": "notes.create",
"args": {
"title": "Acme pricing follow-ups",
"markdown": "## Follow-ups\n\n- Revised quote at 12% for a two-year term\n- Confirm the start date with legal",
"sessionIds": ["6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10"]
}
},
{
"tool": "action_items.create",
"args": {
"title": "Send the revised quote to Acme",
"sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10",
"itemType": "task",
"dueAtMs": 1793725200000,
"assignees": ["Alex Kim"]
}
}
],
"stopOnError": true
}
{
"results": [
{
"tool": "notes.create",
"ok": true,
"result": {
"noteId": "note-k2v9w8x7c6b5n4m3l2j1h0g9",
"title": "Acme pricing follow-ups",
"pinned": false,
"sessionId": null,
"updatedAt": 1793369525000,
"projectId": null,
"sessionIds": ["6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10"]
}
},
{
"tool": "action_items.create",
"ok": true,
"result": {
"itemId": "outcome-x4c8v2b6n1m9q3w7e5r0t2y8",
"itemType": "task",
"title": "Send the revised quote to Acme",
"status": "open",
"dueAtMs": 1793725200000,
"sessionId": "6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10",
"sessionTitle": "Pricing review with Acme",
"projectId": "project-p8d3k2m1q9w7e5r4t6y0u2i1",
"assignees": [{ "name": "Alex Kim", "contactId": null }],
"origin": "manual",
"sources": [],
"sessionStatus": "complete"
}
}
]
}
The action item response is shortened. A note filed under a session shows on that session's page; sessionId stays null because it is not the session's own note. The action item takes the session's project unless you give another. If TalkBar is open, both show up right away.
Add a bookmark to a live recording
Setup: switch on Add marker (markers.create) under Markers. Recording itself always stays in the app.
While you record, ask: "Bookmark the moment Sam mentioned the budget freeze."
First the assistant finds the session that is recording now:
{ "tool": "sessions.recording" }
{
"recording": {
"sessionId": "c3a9f2e7-1d84-4b6a-9e05-7f2b8d1c6a34",
"title": "Budget planning",
"startedAtMs": 1793368800000,
"elapsedMs": 725000,
"revision": "…",
"appRunning": true,
"stale": false
}
}
recording is null when nothing is recording. stale: true would mean TalkBar is not running, so the session only looks like it is recording.
Then it reads the transcript so far as JSON. A read with cursor: "0:" returns turns in the order they changed, plus a nextCursor to fetch only what is new next time:
{
"tool": "sessions.transcript",
"args": { "sessionId": "c3a9f2e7-1d84-4b6a-9e05-7f2b8d1c6a34", "format": "json", "cursor": "0:" }
}
{
"sessionId": "c3a9f2e7-1d84-4b6a-9e05-7f2b8d1c6a34",
"title": "Budget planning",
"startedAtMs": 1793368800000,
"private": false,
"truncated": false,
"excludedFromAiSpeakers": [],
"nextCursor": "1793369523871:turn-42",
"recording": true,
"turns": [
{
"turnId": "turn-42",
"chunkId": "chunk-17",
"provisional": true,
"startMs": 641000,
"endMs": 652500,
"speaker": "Sam Ortiz",
"isUser": false,
"excludedFromAi": false,
"text": "Heads up, finance just told us there's a budget freeze until January.",
"translatedText": null
}
]
}
It adds the bookmark at that turn, with a short note:
{
"tool": "markers.create",
"args": {
"sessionId": "c3a9f2e7-1d84-4b6a-9e05-7f2b8d1c6a34",
"turnId": "turn-42",
"body": "Budget freeze until January"
}
}
{
"markerId": "…",
"sessionId": "c3a9f2e7-1d84-4b6a-9e05-7f2b8d1c6a34",
"kind": "bookmark",
"body": "Budget freeze until January",
"quote": null,
"atMs": 641000,
"turnId": "turn-42",
"chunkId": null,
"pointId": null,
"color": "",
"updatedAt": 1793369525000,
"sessionStatus": "recording"
}
The bookmark appears on the recording screen right away. Instead of turnId, an assistant can place a bookmark with chunkId, pointId (a timeline point from sessions.timeline), atMs (a moment from the session start) or quote (the words it is about).
Export a session
Setup: add an export destination in the app first, for example a folder in your Obsidian vault; see Exports. Export operations are on by default once MCP is on.
Ask: "Export the Acme pricing review to my Obsidian vault."
The assistant looks up your destinations:
{ "tool": "exports.config" }
{
"destinations": [
{
"destinationId": "5d2e7a10-8c3b-4f69-b1e4-0a9c6d3f2b87",
"workspaceId": "…",
"name": "Obsidian vault",
"format": "obsidian",
"location": "PATH_TO_VAULT/Meetings",
"enabled": true,
"layout": "flat",
"credentialRef": null
}
],
"rules": []
}
Then it queues the export, choosing what to include:
{
"tool": "exports.enqueue",
"args": {
"sessionIds": ["6f1c2a9e-4b7d-4c1e-9a53-0d2e8b7f4c10"],
"destinationIds": ["5d2e7a10-8c3b-4f69-b1e4-0a9c6d3f2b87"],
"selection": { "summary": true, "transcript": true, "metadata": true, "notes": true, "audio": "none" }
}
}
The answer is the new export batch, with its batchId, a status and one entry in deliveries per session and destination. exports.enqueue only queues the work. The TalkBar app delivers queued exports while it runs. To deliver now, the assistant can call exports.run with the batchId, then check with exports.status:
{ "tool": "exports.run", "args": { "batchId": "…" } }
A delivered entry has "status": "exported" and its location. A session that is still recording cannot be queued; that call answers session_recording.
From a terminal, talkbar-cli exports run --workspace WORKSPACE_ID --batch BATCH_ID --wait keeps delivering until the batch is done. See talkbar-cli exports.
Delete safely
Setup: switch on Move note to Trash (notes.trash) under Notes. Keep Allow deletes without confirmation off.
Ask: "Trash the old Acme draft note."
The assistant finds the note with talkbar_find, then calls talkbar_delete:
{ "tool": "notes.trash", "args": { "noteId": "note-a1s2d3f4g5h6j7k8l9z0x1c2" } }
Before anything runs, your AI app shows TalkBar's question:
An assistant wants TalkBar to run:
- notes.trash {"noteId":"note-a1s2d3f4g5h6j7k8l9z0x1c2"}
Confirm to run it, or decline to run nothing.
If you decline, or don't answer within 5 minutes, nothing runs and the assistant gets:
{ "error": { "code": "declined", "message": "The user did not confirm, so nothing ran" } }
If you confirm, the note moves to the Trash:
{ "noteId": "note-a1s2d3f4g5h6j7k8l9z0x1c2", "trashed": true }
Changed your mind? notes.restore brings it back. It runs through talkbar_change, so it needs no confirmation, but you have to switch on Restore note:
{ "tool": "notes.restore", "args": { "noteId": "note-a1s2d3f4g5h6j7k8l9z0x1c2" } }
{ "noteId": "note-a1s2d3f4g5h6j7k8l9z0x1c2", "restored": true }
notes.trash.list lists what is in the notes Trash. Nothing an assistant does deletes a note for good; emptying the Trash stays in the app.
If your AI app cannot show questions from a server, the delete is refused with confirmation_unavailable instead, and nothing runs. See Apps that cannot ask.