# classifier.dev — MCP server card for AI agents

Sort up to 1,000 texts into your own labels with a calibrated confidence per answer. No API key.

Registry name: `dev.classifier/classifier` · category: Developer tools · version 1.0.0 · updated 2026-09-19

## Connect
- Remote (Claude Code): `claude mcp add --transport http classifier-dev https://classifier.dev/mcp`

## Facts
- Repository: https://github.com/mrmps/classifier-dev (441 stars)
- Website: https://classifier.dev

Before you connect a server: read its repository, and give it only the permissions and keys it needs.

## Links
- AgentGid page: https://agentgid.com/mcp/servers/dev-classifier-classifier/


## classifier.dev: sort many texts into your own labels

classifier.dev lets your AI assistant sort a pile of texts, like support tickets, reviews or search results, into labels you name, such as bug, feature or praise. Each answer comes with a confidence number, so you read only the unsure ones. It works without an account or key, but by default the texts you send are kept to train its models unless you switch that off.

### Where to use it
- Sort a backlog of support tickets: Your assistant puts each ticket into billing, bug or question and gives a confidence for each, so the sure ones are routed and you read the rest.
- Tag articles with every topic that fits: An article can be about pricing and security at once. The assistant gets every label that applies to each text, with a score per label.
- See what share of feedback is bugs: Get only the totals, such as 412 bugs, 230 praise and 58 unsure, without pulling a thousand answers into the chat.
- Find the items worth a human look: After a big sort, ask for only the texts the model was unsure about, each with its second guess, and read just those.
- Route each ticket by team and urgency: One request labels every ticket by several things at once, for example which team, how urgent and what kind, each with its own confidence.
- Filter search results before reading them: Your assistant found forty links; it labels them relevant or not and opens only the six that matter, which saves time and tokens.

### How to start
- Add the server to Claude Code (or Codex, Cursor and others; the setup page lists them all). No key is needed.

```
claude mcp add --transport http classifier https://classifier.dev/mcp
```

- In the Claude app, add it as a custom connector with this address and leave the sign-in fields empty.
- Ask the assistant to pass share_data false in every call if your texts should not be kept for training.

### License
The code is open source under the MIT license. Using the hosted service is free within its limits and offered as is. The answers are yours, but texts you send are kept for training unless you opt out.

### Good to know
- Texts, labels and results are kept for training and staff review by default. Turn it off on every call with share_data false.
- Free use has limits per internet address, 3,000 texts a minute and 20,000 a day on the fast tier, and fewer on the smart tier.
- Reading a web page by its address, and long documents, need a paid workspace balance.
- The terms forbid using it to make important decisions about a person without a human checking.
- Every classifying tool uses up your quota or paid balance, so your assistant should not repeat calls blindly.

Checked 2026-11-10 against: https://github.com/mrmps/classifier-dev, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/README.md, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/server.json, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/src/mcp.ts, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/src/http/mcp.ts, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/src/index.ts, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/src/training-data.ts, https://github.com/mrmps/classifier-dev/blob/b0e3b449a9c438a1fc31de2d50a8f1546b7d7c49/plugins/classifier/.mcp.json, https://classifier.dev/, https://classifier.dev/mcp-setup, https://classifier.dev/pricing, https://classifier.dev/privacy, https://classifier.dev/terms, https://classifier.dev/agents.md, https://classifier.dev/skill.md, https://classifier.dev/llms.txt

### Mistakes to avoid
- Don't: Calling classify_texts with only inputs and labels on private or customer texts → Do: Add "share_data": false to the arguments of every call (An omitted setting counts as sharing, so the texts are kept for training.)
- Don't: if (r.confidence < 0.7) review(r) when tier is smart → Do: Check confidence === null first and send nulls to review (Answers re-asked by the smart tier come back with null confidence and scores.)
- Don't: Labels like "p0", "p1" or no way to say none of these → Do: Descriptive labels ("urgent bug") plus a "none of these" label when nothing may fit (Every call returns one of your labels; the label names are all the model has to go on.)
- Don't: tier "smart" on 1,000 texts without a key → Do: Split public smart batches into 200 texts or fewer, or use the fast tier (Public smart batches above 200 texts are refused.)
- Don't: Calling the server for three items already in the chat → Do: Decide yourself; use it when there are many texts to read (The server's own instructions say it only pays off when reading the inputs is the expensive part.)

### Questions
- Do I need an account or a key? No. The public server works without a key within per-address limits; a workspace key only adds paid capacity and usage records.
- Are my texts kept? Yes by default, for training and operator review. Pass share_data false on each call to opt out; signing in or paying does not opt you out.
- Where do my texts go? To classifier.dev, which forwards them to its model providers, TypeSafe and, for the smart tier, a reasoning model through OpenRouter.
- How many texts can one call take? Up to 1,000 texts and 2 to 100 labels per call.
- Is there a separate server for its documentation? Yes, at classifier.dev/mcp/docs; it reads and searches the docs and classifies nothing.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
