Skip to main content
Ask your agent what people said about you, and let it read the answer out of Nephia.
That is the whole install. Your browser opens once, you sign in and choose what the agent may do, and the connection is live — no API key to copy, nothing to paste. The official MCP (Model Context Protocol) server exposes the product API — your mentions, your queries, their buckets — as tools any model can call. It comes two ways: remote, at https://api.nephia.cc/mcp under OAuth, and local, as an npx command under an API key.

​
Try it

Once it is added, ask in plain language:
What did Reddit say about us this week? Anything negative?
The agent calls list_queries to find your monitor, then list_mentions with source=reddit and a week’s since, and answers from the scored stream:
Nothing was computed for that answer that the dashboard does not already show. What is new is that the agent could read it. More prompts that map onto one or two calls:
  • “Any leads this week?” → list_mentions with intent=["purchase_intent", "comparison"]
  • “Show me the pricing complaints on my Nephia monitor” → list_query_buckets, then get_query_results with bucket=
  • “Find people who can’t get it to deploy” → list_mentions with mode=semantic (charged — see below)
  • “Pause my brand monitor while we’re on holiday” → query_manage with action=pause

​
Clients

One URL, no key. The first call opens your browser: you sign in, a screen lists what the agent is asking for, and you decide. Works in any client that can add a remote MCP server — including the hosted ones that cannot run a local command. Claude Code
Cursor ~/.cursor/mcp.json for every project, or .cursor/mcp.json for one. A remote server is a url and nothing else — no command, no args, no type:
Cursor registers itself and opens the browser on first use. Nothing to whitelist: its callbacks (http://localhost:8787/callback on desktop, https://www.cursor.com/agents/mcp/oauth/callback on web) are accepted as they are. Claude.ai, ChatGPT, and other hosted clients Add a custom connector and give it the server URL:
What the consent screen asks for The third one arrives unticked. An agent you leave it off for still reads everything — it just cannot spend, and it will say so when a call needs it, naming the permission you did not grant. Disconnecting Settings → API keys → Connected agents → Disconnect. The agent has to sign in again to reach your account. A token it is already holding keeps working until it expires, which is at most an hour.

​
Local (stdio)

An API key and a local command. Right when you want no browser in the loop — a script, a CI job, a machine that is not yours to sign in on. Claude Code
Claude Desktop, Cursor, and anything else reading an mcpServers config
Needs Node ≥ 20. In Cursor this goes in the same mcp.json as the remote form above — a server entry is either a url or a command, never both.
Both paths serve the same tools from the same registry, and both bill the same way. Every call appears in Spend either way — under the key’s name for a local server, under the agent’s for a remote one.

​
Tools

Every tool description carries what it costs, so a model can budget before it spends. Check the balance any time with account_credits, which is free. The Source is an argument, not a tool: list_mentions takes source=reddit rather than there being a reddit_mentions beside an x_mentions. One contract, one place to look.

​
Reading, and what it costs

Reading your own mentions is free, in any volume, because they are rows you already own. One thing charges: mode="semantic". The charge is per question, not per page or per result: paging through a semantic answer is free, and asking the same question twice inside ten minutes is one credit. list_mentions and get_query_results are annotated non-read-only and non-idempotent even though they are free by default, because an MCP annotation describes a tool rather than a call: a client’s auto-approve list should have to say out loud that a tool which can debit the account is allowed to run unattended.

​
Creating and changing monitors

Queries are configured in the dashboard, not from an agent: a Query carries rules, buckets, delivery channels, an activation schedule and per-source search overrides, and a monitor composed blind is a meter nobody reads. The MCP server can read every Query, and can pause, resume or retire one — which is also how you stop it spending. Watches — the single-source, webhook-first primitive underneath — can be created from an agent with watch_create. It needs a publicly reachable webhookUrl you control; ask for it, never invent one. Without a receiver, poll matches with watch_feed (kind=events) and delivery outcomes with kind=activity, both free. See Watches and Webhooks.

​
Errors

Tool results carry actionable error text instead of raw failures: 401 tells you to reconnect on the remote server and points to NEPHIA_API_KEY on the local one, 402 to credits (check with account_credits), 404 means not found or not yours — never a 403, which would confirm someone else’s Query exists — and 429/503 include the retry delay. The underlying client already retries 429/502/503/504 with backoff, never timeouts or other 4xx. On the remote server one more refusal exists: a call the connection was not granted. It comes back as a tool result naming the missing permission, not as a broken connection — reconnect and allow it on the consent screen.

​
Large responses

Mention lists are shaped for a context window before they reach the model: capped at 25 items per call by default, bodies truncated, screen-only fields (thumbnails, captions, the label/value grid) dropped, and sentiment, intent, bucket, author and URL kept. Ask for a bigger limit to count, not to read. Beyond that, any response over 50 000 characters is truncated with an explicit notice — narrow the request or page with the returned cursor. Override with the NEPHIA_MCP_MAX_OUTPUT_CHARS env var.

​
Next steps