Everything you follow, as an MCP server
Adviserry is an MCP server over your own subscriptions. It ingests the newsletters in your Gmail and the YouTube channels you follow, then exposes them as tools inside Claude Desktop, Claude Code, Cursor, and Gemini CLI.
Your assistant stops guessing from the open web and answers from the people you chose to listen to, with the creator, title, and publish date attached to every passage it quotes.
The facts, in one place
- What it is
- A hosted MCP server, one per account, speaking JSON-RPC 2.0 at https://adviserry.com/api/mcp.
- Who it is for
- Solo founders, consultants, and operators who follow more newsletters and channels than they can read.
- Sources
- Newsletters from Gmail (read-only) or forwarded to a private address, YouTube channels with full transcripts, and your own uploaded documents.
- Works in
- Claude Desktop, Claude Code, Cursor, Gemini CLI, OpenClaw. ChatGPT connectors are not supported yet.
- Price
- $49/month after a 7-day free trial. Team is $149/month for 5 seats, each with its own token.
- Auth
- A personal bearer token from Settings, sent in an Authorization header. Revoke or regenerate it any time.
- Limits
- 500 requests per rolling hour, per user.
What can you ask it?
Once the server is connected you type these into your normal chat and the client picks the tool. Here is what an operator actually asks, and which tool answers it.
“What have my sources said about pricing experiments in the last month?”
Uses
search_newsletters“Find the bit in that podcast where they talk about hiring a first salesperson.”
Uses
search_newsletters“Catch me up on the last week across my Growth panel.”
Uses
get_recent_issues“What should I be working on this week? Show me the full draft.”
Uses
get_actions, then get_action“Check my strategy doc against what my sources are saying about onboarding.”
Uses
search_context + search_newsletters“Add the October Team plan launch to what I am working on.”
Uses
update_context
How do I set it up?
Three steps, about five minutes, and most of that is waiting for your first Gmail sync.
- Connect Gmail or add YouTube channels, so there is something to search. Adviserry detects which senders are newsletters and only ingests those.
- Copy your MCP token from Settings, under MCP Server Access.
- Paste the config below into your client, restart it, and ask it something.
Claude Desktop Configuration
Add this to your claude_desktop_config.json
{
"mcpServers": {
"adviserry": {
"url": "https://adviserry.com/api/mcp",
"headers": {
"Authorization": "Bearer YOUR_MCP_TOKEN"
}
}
}
}macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Restart Claude Desktop after saving changes.
What tools does the server expose?
13 tools, discovered automatically by your client the moment it connects. Parameters and defaults for each one are in the MCP documentation.
| Tool | What it does |
|---|---|
search_newsletters | Semantic search across every newsletter issue and video transcript in your archive. Returns the matching passage, the creator, and the publish date. |
list_panels | Lists your panels and the sources inside each one. |
get_recent_issues | Everything that arrived in the last N days, optionally scoped to one panel. |
get_newsletter_summary | The full text and summary of one issue, by id. |
get_topics | The topics your sources have been covering lately. |
search_context | Searches your own uploaded documents: PDFs, decks, spreadsheets, notes. |
get_insights | Proactive observations your panels generated, including cross-source connections. |
search_actions | Finds a past drafted action by topic. Returns the full draft, never truncated. |
get_actions | The most recent drafted actions, newest batch first. |
get_action | One action in full, with every source it was built from and why it was drafted now. |
update_action_status | Marks an action shipped, dismissed, read, or unread. Every value is reversible. |
get_user_context | Reads your working context: role, projects, challenges, goals, notes. |
update_context | Writes to that context. Removals need a second confirming call and are archived, not deleted. |
How does it compare?
Measured against the ways people actually solve this today. Where a category varies too much between products to state a fact, the cell says so rather than guessing.
| Pasting into ChatGPT by hand | Your own MCP server over Gmail | A generic web-search MCP | A read-it-later app with an MCP | Adviserry | |
|---|---|---|---|---|---|
| Auto-ingestion | No. You paste each issue, every time | Yes, once you have built and hosted it | No. It fetches the public web on demand | Varies. Usually what you saved by hand | Yes. Gmail and YouTube sync on a schedule |
| Full-transcript search | Only what fits in the message you pasted | You would have to build transcript fetching | No | Varies | Yes. Chunk-level search over the whole transcript |
| Citations back to the source issue | Whatever you remembered to paste | Up to your implementation | Links to public pages | Varies | Creator, title, publish date, and issue id on every result |
| Works inside Claude Desktop | You are already there, but pasting | Yes, if you host and maintain it | Yes | Varies | Yes, plus Claude Code, Cursor, Gemini CLI, OpenClaw |
| Keeps your project context | Only inside one chat | You would have to build it | No | Varies | Yes, and the assistant can update it with update_context |
| Setup time | None, and it never ends | OAuth, storage, embeddings, hosting | A few minutes | Varies | Connect Gmail, paste one config block |
How much of my context window does it use?
A search returns passages, not documents. Each result comes back as one line of metadata (creator, title, publish date, issue id) plus the region of the text that actually matched your query, capped at roughly 1,200 characters, with [...] marking where text was skipped.
You control the count. search_newsletters defaults to 5 results and accepts up to 10, so the worst case is about 12,000 characters of body text and a default search is closer to 6,000. That is a fraction of what pasting five newsletters costs, and it is why retrieval runs over roughly 1,000-token chunks of each issue instead of over whole documents.
One deliberate exception: an action comes back as its complete draft, never truncated. A summary of a deliverable is not a deliverable.
Is my data sent to Anthropic or OpenAI?
Some of it, and here is the exact split. Your newsletter and transcript content lives in Supabase (Postgres plus object storage) behind row-level security, so only your account can read it. Gmail and YouTube tokens are encrypted with AES-256-GCM before they are stored, and the Gmail scope is read-only.
OpenAI sees content when Adviserry generates summaries and embeddings. Anthropic sees subjects and summaries when it drafts your actions. Neither provider trains on API data, per their published policies.
MCP responses go to whichever client you connected. If that is Claude Desktop, the passages become part of your Claude conversation and are handled under your own account with Anthropic. Adviserry has no visibility into what happens after the response leaves our server. That is the trade you make when you wire any MCP server into an assistant, and it is worth knowing rather than assuming.
Full detail is in the privacy policy.
Frequently asked questions
Which MCP servers integrate best with Claude Desktop?
The ones that carry data Claude cannot already reach. A web-search server duplicates what the model does anyway. Adviserry exposes your own newsletters, YouTube transcripts, and uploaded documents over MCP, so Claude Desktop can answer from sources it has no other way to see. Setup is one config block with a bearer token.
Does the Adviserry MCP server work with ChatGPT?
Not yet. ChatGPT custom connectors require OAuth 2.1 with dynamic client registration, which Adviserry does not implement today. Adviserry works with Claude Desktop, Claude Code, Cursor, Gemini CLI, and OpenClaw right now, all using a bearer token. OAuth support is on the roadmap, and the ChatGPT tab in our docs will update when it ships.
Can an MCP server search my YouTube subscriptions?
Adviserry's can. Add the channels you follow and every new long-form video gets transcribed, chunked, and embedded. The search_newsletters tool then searches inside the transcript, not just the title or a summary, so a claim made 35 minutes into an episode is findable. Results come back with the channel name, video title, and publish date.
How much does the Adviserry MCP server cost?
MCP access is included in Adviserry Pro at $49 per month, after a 7-day free trial with a card up front. There is no separate MCP fee and no per-request charge. Team is $149 per month for 5 seats, each with its own token. Every plan gets 500 MCP requests per rolling hour.
Can the assistant update my context through MCP?
Yes, with a guard. The update_context tool lets Claude write to your role, projects, challenges, goals, and freeform notes. Appends save straight away. Anything that would remove existing context returns a preview instead and writes nothing until you approve it on a second call, and removed items are archived rather than deleted by default.
Does it work with Cursor or Claude Code?
Yes. Both connect over HTTP with the same endpoint and bearer token. Claude Code can add it in one command: claude mcp add --transport http adviserry. Cursor takes a JSON block under Settings then MCP. Gemini CLI and OpenClaw work the same way. The docs page has a copy-ready config for each client.
How much of my context window does an Adviserry search use?
Less than you would think. A search returns the passages that matched your query, not whole issues, and each excerpt is capped at roughly 1,200 characters. The default is 5 results and the maximum is 10, so a typical search adds a few thousand characters of body text rather than five full newsletters.
Is my newsletter data sent to Anthropic or OpenAI?
Some of it, and here is exactly where. Content is stored in Supabase under row-level security. OpenAI generates the embeddings and summaries. Anthropic receives subjects and summaries when it drafts your actions. MCP results go to whichever client you connected, so a Claude Desktop query puts those passages into your Claude conversation. Neither provider trains on API data.
Put your sources inside your assistant
Connect Gmail, add the channels you watch, paste one config block. $49/month after a 7-day free trial, and you can cancel before it ends and pay nothing.