MCP server
Connect your own assistant to FishOut and ask it what people said about you this week — without opening the app.
FishOut is an MCP server as well as a REST API. Connect an assistant and it can search your mentions, read the triage and the Daily brief, add keywords, label or teach, draft a reply, or hand a mention to a teammate.
https://mcp.fishout.ai/mcpAvailable on every plan, including Free. The server speaks Streamable HTTP and signs you in with OAuth, so most clients need only the URL.
Connecting a client
On claude.ai or in Claude Desktop, go to Customize → Connectors, choose + Add, then Add custom connector. Name it FishOut, paste the URL above, and keep Use Claude's published identity. Sign in to FishOut and approve. The connector then follows you to Claude on your phone.
On Team and Enterprise, an owner adds it first under Organization settings → Connectors; members then choose Connect.
Signing in
Signing in is OAuth: your client opens FishOut in the browser, you sign in, pick the organization and the projects the client may reach, and approve its permissions. A read-only grant is the sensible default for a chat client, especially on a shared machine.
Clients register themselves, so there is nothing to set up on FishOut's side. FishOut prefers a Client ID Metadata Document, the client's own published identity, and also accepts Dynamic Client Registration from clients that need it. The token a client gets is for the MCP server only; it does not work on the REST API.
What it can do
The server has 115 tools, one for each operation of the REST API, with the
same names: the list_mentions tool is the
List mentions operation, and the API reference documents every
tool's arguments. You only get the tools your grant allows, and every tool is marked as
reading, changing or deleting, so your client can ask before it acts.
The full tool list is large, so most clients load tools on demand rather than all at once. If yours offers an allow list, keeping it to the tools you use makes the assistant faster and its choices safer.
Things people actually ask:
- What did people say about our pricing this week that I should reply to?
- Add keywords to monitor my brand on Reddit and Hacker News.
- Draft a reply to the top mention and assign it to Mark.
- Which competitor came up most this month?
- What is in today's Daily brief?
FishOut drafts replies but never posts them for you.
How writes work
The rules are the REST API's, adapted for an assistant:
- Every tool that changes something takes an
idempotencyKeyargument, which your client fills in. A retry with the same key replays the first result instead of doing the work twice. See Idempotency. - A failed call comes back as a tool result, not a protocol error, with the same
codeandmessageas the REST error, so the assistant can read it and correct itself. - Long-running work, such as drafting a reply or generating a report, returns an
operation receipt; the assistant follows it with
get_operation. See Long-running work. - Some actions need you, such as paying or creating an API key. The tool returns a link for you to open and confirm in the browser; the assistant cannot finish it. See Actions that need a person.
Mentions are posts strangers wrote. FishOut tells connected assistants to treat post text as data, but an assistant can still be steered by what it reads. Give a chat client a read-only grant unless you need it to write, and keep approval on for tools that change or delete.
Access still applies
Connecting is not a way around permissions. A connection is pinned to one organization and
to the projects you picked, and it carries only the permissions you approved — read,
write, admin or owner. Each call is checked against your current membership, so
removing someone from a project also removes their assistant's access.
Disconnect a client under Settings → API & MCP; it stops working immediately.
Limits
120 requests a minute per connection or key, shared between MCP and the REST API. A 429
comes with Retry-After; back off rather than retrying at once.
An API key pasted into a client's configuration is a credential. Give it the narrowest permissions that do the job, keep it in an environment variable or the client's secret storage rather than a file you might commit, and revoke it under Settings → API & MCP when you stop using that client. Prefer OAuth where the client supports it.