aftercall

Search and ask questions about your meetings using AI.

After a video call, aftercall automatically saves the summary, the full transcript, and any action items. Then you can ask Claude things like "What did I agree to do for Sarah?" or "What meetings mentioned the budget?" — and get real answers back.

Every to-do from your calls drops into a Notion checklist so nothing falls through the cracks.

Ongoing cost: ~$5/mo for hosting + a fraction of a cent per call for AI processing. Needs a Bluedot account (free trial) to record your calls.

 Bluedot call ──▶ Worker ──▶ D1 + Vectorize + Notion ──▶ Claude.ai (MCP)
   webhook       extract       indexed storage          you ask questions

flowchart LR
    B[Bluedot]
    W{{Cloudflare Worker}}
    D[(D1)]
    V[(Vectorize)]
    N[Notion Followups]
    C{{Claude.ai}}
    U((You))

    B -->|webhook| W
    W -->|upsert| D
    W -->|embed + upsert| V
    W -->|create rows| N

    U -->|"what did I promise Pierce?"| C
    C -->|MCP tools/call| W
    W -->|read| D
    W -->|semantic search| V
    W -->|filter rows| N
    W -->|answer| C
    C -->|response| U

Loading

---

What you get

Every Bluedot recording becomes:

1. A D1 row — transcripts table, idempotent on video_id, holding raw text + structured summary + participants + action items.
2. Embedded chunks in Cloudflare Vectorize (1536d, cosine) for semantic search.
3. A Notion Transcripts row — metadata hub (Date, Participants, Recording URL, link to Bluedot's native summary page). No duplicated summary content — Bluedot's own Notion sync owns that.
4. One Followup row per action item in a Notion inbox — Status = Inbox, linked via a Meeting relation back to the Transcripts row, ready to triage.
5. MCP access from Claude.ai — six tools for semantic search, per-call Q&A (RAG), detail lookup, followups, and owner-scoped action items.

Then you ask Claude.ai things like:

• "What did I commit to do for Pierce in the last week?"
• "In my IT hiring call, what starting compensation did we agree on?"
• "Find every action item assigned to Andy since March."
• "Summarize my open Bluedot followups."
• "What calls mention IronRidge?"

---

Setup (10 minutes)

Prerequisites

Service	Tier	Purpose
Cloudflare	Workers free + Vectorize paid (~$5/mo)	Hosting + D1 + Vectorize + KV
OpenAI	Pay-as-you-go (~$0.001/call)	Extraction (gpt-5-mini) + embeddings (text-embedding-3-small)
Notion + integration	Free	Transcript pages + Followups inbox
Bluedot	Trial	Source of meeting recordings
GitHub OAuth App	Free	MCP auth (optional — only if connecting Claude.ai)
Claude.ai	Free/paid	MCP client (optional)

Run the setup script

git clone https://github.com/jchu96/aftercall.git
cd aftercall
npm install
npx wrangler login
npm run setup

The script walks 10 interactive steps:

1. Verify wrangler auth
2. Provision D1 database (idempotent — reuses existing)
3. Provision Vectorize index
4. Write bindings to wrangler.toml + apply D1 migrations
5. Create (or reuse) Notion Followups + Call Transcripts databases
6. Validate OpenAI API key
7. Write .dev.vars + update wrangler.toml vars
8. Register GitHub OAuth App (opens browser, auto-detects your worker URL) + configure MCP KV + allowlist
9. Push every secret to Cloudflare (no manual wrangler secret put dance)
10. wrangler deploy

Re-runs are safe — the script detects existing values and offers to reuse them instead of recreating.

Set the Bluedot webhook

After deploy, the script prints your worker URL. In Bluedot:

• Settings → Webhooks → Add endpoint
• URL: https://<your-worker>.workers.dev/
• Events: meeting.transcript.created, meeting.summary.created

Bluedot shows a signing secret. Save it:

npx wrangler secret put BLUEDOT_WEBHOOK_SECRET

Connect Claude.ai (optional, MCP)

• Claude.ai → Settings → Connectors → add a custom MCP server
• URL: https://<your-worker>.workers.dev/mcp
• Sign in with GitHub; the Worker checks your username against ALLOWED_USERS and hands Claude a bearer token

Done. Ask Claude about your calls.

Full OAuth + Bluedot + debug walkthrough: docs/auth.md.

---

MCP tools reference

Tool	What it does
search_calls(query, limit?)	Semantic search over all transcripts via OpenAI embeddings + Vectorize
get_call(video_id)	Full details of one call: summary, participants, action items
answer_from_transcript(video_id, question)	RAG over a single call — drill-down Q&A grounded in that meeting's transcript
list_followups(status?, source?, limit?)	Query the Notion Followups DB with select filters
find_action_items_for(person, since?)	All action items assigned to a person (substring match on owner)
recent_calls(days?)	Last N days of calls, newest first

Sample prompts per tool: docs/tools.md.

---

Architecture

Three flows compose the system. Each has its own diagram and runbook:

• Ingestion — Bluedot webhook → OpenAI extraction + embeddings → D1 + Vectorize + Notion (diagram)
• OAuth — Claude.ai → Worker /authorize → GitHub → Worker /auth/github/callback → Claude.ai (diagram)
• Query — Claude.ai → bearer-protected /mcp → tool dispatch → storage → response (diagram)

Deep-dive: docs/architecture.md.

---

Stack

Layer	Tech
Webhook + MCP	Cloudflare Workers + Hono
Transcript store	Cloudflare D1 (SQLite, Drizzle schema)
Embeddings	Cloudflare Vectorize (1536d, cosine, HNSW)
LLM	OpenAI — gpt-5-mini (extraction, json_schema) + text-embedding-3-small
Output	Notion API (direct fetch — not @notionhq/client)
MCP transport	@modelcontextprotocol/sdk Streamable HTTP, stateless mode
MCP auth	@cloudflare/workers-oauth-provider + GitHub OAuth

Not in the stack: Anthropic, Neon, Postgres, Redis, SendGrid, Express. One LLM provider, one Notion integration, one GitHub OAuth App.

---

Testing

npx vitest run       # 110 tests, real D1 via miniflare
npx vitest           # watch mode
npx tsc --noEmit     # typecheck

Tests cover Svix verification, payload normalization, OpenAI extraction (mocked), embeddings chunking, D1 idempotency with a concurrent-retry race test, Vectorize upsert, Notion page builders, the GitHub OAuth handler + allowlist, the full OAuth provider integration (well-known metadata, WWW-Authenticate on 401, bearer revocation), and every MCP tool's business logic against real D1 + mocked OpenAI/Vectorize/Notion.

End-to-end MCP transport (tools/list, tools/call through Streamable HTTP) is validated by live Claude.ai smoke tests — vitest-pool-workers' ESM shim can't resolve the SDK's ajv JSON import.

---

Operating

Task	Command
Tail live logs	npx wrangler tail
Deploy	npx wrangler deploy
Inspect D1	npx wrangler d1 execute aftercall-db --remote --command "SELECT ..."
Query Vectorize	npx wrangler vectorize get-by-ids aftercall-vectors --ids "1-0,1-1"
Reprocess a call	DELETE FROM transcripts WHERE video_id = '...' then refire Bluedot webhook
List KV (OAuth)	npx wrangler kv key list --binding OAUTH_KV
Revoke your bearer	curl -X POST https://.../auth/revoke -H "Authorization: Bearer <token>"

Full runbook: docs/architecture.md#operational-runbook.

---

Environment variables

Full reference (click to expand)

Set via wrangler.toml [vars] for non-secret config and wrangler secret put for secrets. .dev.vars mirrors secrets for local wrangler dev.

Secrets

Variable	Required	Notes
OPENAI_API_KEY	yes	Extraction + embeddings
NOTION_INTEGRATION_KEY	yes	Notion integration token (ntn_...)
BLUEDOT_WEBHOOK_SECRET	yes	Svix signing secret from Bluedot's webhook config
GITHUB_CLIENT_ID	MCP only	GitHub OAuth App client id
GITHUB_CLIENT_SECRET	MCP only	GitHub OAuth App client secret
SENTRY_DSN	optional	Enables error tracking + pipeline tracing. Leave unset to disable Sentry entirely.

Vars (wrangler.toml)

Variable	Default	Notes
OPENAI_EXTRACTION_MODEL	gpt-5-mini	Override to upgrade
NOTION_TRANSCRIPTS_DATA_SOURCE_ID	—	Set by setup script
NOTION_FOLLOWUPS_DATA_SOURCE_ID	—	Set by setup script
BASE_URL	—	Public worker origin; used for OAuth callback URL construction
ALLOWED_USERS	—	Comma-separated GitHub usernames allowed via MCP
SENTRY_ENVIRONMENT	production	Only read when SENTRY_DSN is set
SENTRY_RELEASE	auto	Injected by npm run deploy when Sentry is configured

Bindings

Binding	Type	Purpose
DB	D1	Transcripts
VECTORIZE	Vectorize	Chunk embeddings
OAUTH_KV	KV	OAuth state + tokens (MCP only)

Compatibility flags

Flag	Why
nodejs_compat	Bluedot's Svix package needs Node APIs
global_fetch_strictly_public	Required by @cloudflare/workers-oauth-provider

---

Repo layout

File-by-file (click to expand)

src/
├── index.ts                  # Worker entry — re-exports OAuthProvider
├── handler.ts                # ingestion pipeline orchestration
├── env.ts                    # typed Env (D1, Vectorize, KV, secrets)
├── webhook-verify.ts         # Svix signature verification
├── bluedot.ts                # Bluedot payload normalization
├── extract.ts                # OpenAI structured extraction
├── embeddings.ts             # OpenAI embeddings + chunking
├── d1.ts                     # transcripts table writes
├── vectorize.ts              # Vectorize upserts
├── notion.ts                 # Notion API (transcript pages, followups)
├── schema.ts                 # Drizzle SQLite schema
├── logger.ts                 # structured JSON logging
└── mcp/
    ├── index.ts              # OAuthProvider wiring + Hono default app
    ├── handler.ts            # /mcp API handler (bearer required)
    ├── tools.ts              # McpServer + Streamable HTTP transport
    ├── auth/
    │   ├── github.ts         # /authorize + /auth/github/callback
    │   └── allowlist.ts      # case-insensitive username check
    └── tools/
        ├── search_calls.ts
        ├── get_call.ts
        ├── list_followups.ts
        ├── find_action_items_for.ts
        └── recent_calls.ts

scripts/
├── setup.ts                  # 10-step interactive provisioning
├── smoke-vectorize.ts        # Vectorize round-trip smoke test
└── migrate-from-neon.ts      # historical Neon → D1 migration

drizzle/                      # numbered SQL migrations
docs/
├── architecture.md           # three flows + data model + decisions + runbook
├── tools.md                  # MCP tool reference with sample prompts
└── auth.md                   # GitHub OAuth setup + troubleshooting
test/                         # vitest helpers + ProvidedEnv typing

---

Observability (optional)

The Worker ships with @sentry/cloudflare integration for error tracking + pipeline performance tracing. It's fully optional — leave SENTRY_DSN unset and the SDK is a no-op.

Enable it (your own fork, your own Sentry project):

# 1. Add DSN as a secret
npx wrangler secret put SENTRY_DSN

# 2. (optional) enable source map uploads on deploy
#    - export SENTRY_AUTH_TOKEN from https://sentry.io/settings/account/api/auth-tokens/
#    - override defaults in scripts/deploy.mjs (SENTRY_ORG, SENTRY_PROJECT envs)
SENTRY_ORG=my-org SENTRY_PROJECT=my-project SENTRY_AUTH_TOKEN=sntrys_... npm run deploy

What's instrumented:

• Pipeline spans — bluedot.pipeline.{transcript,summary} wrap the webhook handlers, with child spans for openai.extract, openai.embed, d1.upsert_*, vectorize.upsert, notion.create_transcript_page, notion.create_followup.
• Error capture — every pipeline catch site calls Sentry.captureException with video_id + svix_id tags. Notion failures are non-fatal but still reported.
• MCP + OAuth errors — withSentry wraps the worker entrypoint so any uncaught error in the OAuth or MCP paths auto-reports.
• Source maps — npm run deploy uploads them to Sentry automatically when .sentryclirc or SENTRY_AUTH_TOKEN is present; otherwise the step is skipped with a warning.

Without Sentry configured, npm run deploy just runs wrangler deploy as before.

---

Documentation

Doc	Read when
docs/architecture.md	Understanding the system, adding features, debugging
docs/tools.md	Using MCP tools from Claude.ai, adding a new tool
docs/auth.md	First-time OAuth setup, troubleshooting 401/403/redirect_uri errors
CHANGELOG.md	What changed between versions
CLAUDE.md	Conventions + do-nots when using Claude Code in this repo
conductor/	Product definition, tech stack, workflow, and track specs — the "why" behind the code

---

Roadmap

Not promises — just ideas on the shortlist. File an issue if you want to prioritize one.

• delete_call(video_id) MCP tool — let Claude.ai prune unwanted calls end-to-end:
  • Delete the D1 transcripts row
  • Delete the matching Vectorize chunks (by deterministic {id}-{chunk} IDs)
  • Archive the aftercall Transcripts page + every Followup row related via the Meeting relation (unlink before archive, or Notion will orphan the relation)
  • Leave Bluedot's native Notion summary page alone — it's Bluedot's surface, not ours
  • Needs a confirmation / dry-run guard before landing, since the tool is destructive and Claude can call it without a human in the loop.

---

License

MIT — see LICENSE.