Build a Zoom MCP your AI can actually use.

<task>
Build a local Zoom MCP server and an optional indexed memory layer for me.
Work with me until the connection is tested end to end.
</task>

<operating_rules>
- Inspect my machine, current folder, Python version, and available MCP clients first.
- Never ask me to paste Zoom credentials, tokens, or secrets into chat.
- Keep secrets in a local .env file that is excluded from Git.
- Use a dedicated Zoom Server-to-Server OAuth app.
- Request only scopes used by implemented tools.
- Start read-only unless I explicitly approve scheduling or update access.
- Do not add deletion scopes without a separate explanation and confirmation.
- Stop before changing Zoom permissions, rotating credentials, or submitting forms;
  put the correct screen in front of me and tell me exactly what will change.
- Use current official Zoom and MCP documentation when an API detail may have changed.
</operating_rules>

<tools_to_build>
Read tools:
- list upcoming meetings
- get one meeting
- list cloud recordings by explicit date range
- get recording files
- list AI meeting summaries by explicit RFC 3339 date range
- get one summary using the meeting-instance UUID
- get and download one retained transcript

Optional write tools, only after I approve them:
- create a meeting
- update a meeting
</tools_to_build>

<zoom_setup>
Guide me through creating and activating the Server-to-Server OAuth app.
Map every requested granular scope to one implemented tool.
Also verify these Zoom settings:
- Cloud recording
- Create audio transcript
- Meeting Summary with AI Companion
- retained meeting transcripts for use by other AI services
Explain that cloud recordings, AI summaries, and retained transcripts are separate API lanes.
</zoom_setup>

<implementation>
- Use Python and the official stable MCP Python SDK.
- Use a small, readable project with .env.example, requirements.txt, and README.md.
- Cache one-hour Zoom access tokens safely.
- Encode meeting UUIDs correctly for API paths.
- Return useful errors without printing credentials or private meeting content.
- Register the server with my chosen MCP client using absolute paths.
</implementation>

<verification>
Test each lane separately and report status codes plus counts, not private content:
1. token request
2. meeting list
3. summary list for the last seven days
4. summary detail for the newest meeting UUID
5. retained transcript metadata and download character count
6. cloud recordings for an explicit date range
If an endpoint returns zero results, inspect its date window and data type before
concluding that permissions are broken.
</verification>

<memory_layer>
After the MCP works, ask me to choose one destination:
1. SQLite with normalized tables and FTS5 search
2. OpenBrain with provenance and entity links
3. Markdown with YAML frontmatter and monthly indexes

Build an idempotent ingestion job that stores meeting UUID, meeting number, topic,
time, people/projects, summary, transcript location or content, decisions, actions,
source URL, content hash, and sync time. The AI should search this local index first
and call Zoom only for fresh, missing, or action-oriented data.
</memory_layer>

<finish_line>
The work is complete only when the MCP reconnects cleanly, the new tools appear in
the client, a real summary and transcript can be retrieved, the chosen local index
can answer a search query, and all created files and retest commands are documented.
</finish_line>