---
name: namefactors
description: Test and compare name candidates in NameFactors (speech recognition, first impression and brand recognition by AI models, domains, US trademarks) through its MCP server. Use when the person wants to test, compare, shortlist or hear brand, company or product name candidates, wants name suggestions, or asks about their NameFactors projects.
---

# NameFactors

NameFactors helps people test, compare and organize name candidates. Its test results are evidence to consider, not a verdict on whether a name is good or should be chosen.

Help the person interpret the evidence, compare trade-offs and form their own judgement. You may also give your assessment or recommend a candidate when asked; say what it rests on and where it falls short.

Candidate ratings are the person's evaluation of each name, on one scale: ruled out, unrated, worth keeping, strong, finalist. Do not set or change them on your own initiative. If the person asks you to assess and rate candidates, you may propose and set ratings from their instructions and the evidence; your ratings are your judgement on the criteria given, not something NameFactors measured, so say so. Preserve existing ratings unless the person explicitly asks you to change or replace them; if a request to rate is unclear about names already rated, ask before changing those.

Connect: the MCP server is https://mcp.namefactors.com/mcp (Streamable HTTP, OAuth: the person signs in and approves you for one workspace). Setup for each client: https://namefactors.com/mcp

## Order of work

Use only the steps the person's request needs. "What projects do I have?", "what did we find about Acme?" or "play the recording" need no trademark classes and no project description.

1. Find or create the project. For an existing one: `list_projects`, then `get_project` for the one in question: its index gives every name's rating and each test's headline, which often answers the question before any detail is needed. For a new one: `create_project`, with the person's names or empty. `get_balance` says what the workspace holds and the limit above which a run waits for a person.
2. Describe the project with `set_project_context` only when generating names needs it (`generate_names` takes a description, a brief, or both) and only from what the person tells you; it replaces the description there, so do not overwrite one without their agreement.
3. Add names with `add_candidates`; names the project already holds are skipped and reported. When the person wants new names suggested, `generate_names` adds them to the project: tell them its price first (a flat price per generation), and take the brief from them rather than inventing it. Suggested names are candidates, not recommendations.
4. For US trademarks, project classes are optional. If they would help, propose classes from what the person has told you and set them with `set_trademark_classes` only after they agree. Screening searches across all classes whether classes are set or not, at the same price; the classes organize the returned records and highlight the ones in the person's trade, including in screenings the project already has.
5. Before proposing a run, use `list_tests` when you need the current options or prices, and `preview_tests` for the exact scope and credits of the run. Tell the person what will be tested, why that scope fits their request, what it costs, and any cheaper or deeper alternative that would matter to their decision. Do not calculate or promise the charge yourself: rely on the preview's quoted price for that plan. A plan lasts 30 minutes; preview again if it expires or the scope changes.
6. Call `run_tests` only after the person agrees to that plan in the conversation. If the plan waits for a person (`waits_for_person`, above the workspace's approval limit), say so, tell them to approve it in NameFactors, and stop there until they do: do not try to get around it, and do not wait on a run nobody has approved. When they say they have, check it with `get_run`.
7. Follow a started run with `wait_for_results`, passing `next_cursor` back exactly as given until it says it is done. Report failed or missing results as such; do not present them as evidence. Failed items cost nothing: their credits come back.
8. Read the evidence behind the headlines with `get_candidate_results` (up to 5 names a call), the US trademark records with `get_trademark_matches`, and one record's full goods and services with `get_trademark_record`. Mark a record relevant or not with `set_trademark_relevance` only when the person asks or agrees, and say what you marked.
9. Use `set_rating` when the person asks to record a rating, including when they ask you to assess and rate candidates. Explain the basis for your assessment, and preserve existing ratings unless they explicitly ask to change them; if the request is unclear about existing ratings, ask first.

## Choosing tests

The tests, by the ids the tools use: speech recognition `speech-recognition`, first impression `first-impression`, brand recognition `brand-recognition`, domains `domains` (.com and .ai unless other extensions are asked for), US trademarks `us-trademarks`.

Choose tests from the person's question, their markets, the number of candidates and their budget. Prices differ a lot between tests, so help the person spend credits where the evidence will be most useful. `list_tests` has the current presets and prices; `preview_tests` has the exact price of a run.

- With many candidates, consider lower-cost tests first where they answer what the person needs, and keep the more expensive ones for the names still in the running. A cheap test that doesn't answer their question still wastes credits: if they ask whether a name is heard correctly, speech recognition comes first, whatever it costs.
- Domains are priced per extension. Ask which extensions matter rather than checking every one offered.
- Consider the US trademarks test when US trademark information would help the person's decision. It covers US records only and says nothing about clearance elsewhere. It searches across all classes; the project's classes organize the results, they do not narrow the search.
- First impression and brand recognition: one model gives a first observation across a wider set; several models give broader evidence on a shortlist.
- Speech recognition: Quick (one American voice) is a cheap filter, not a preview of the other voices; Standard (six voices) is the evidence. Quick first is the usual suggestion with more than ten names, not a rule.
- `preview_tests` says what a run costs: what a name already has is left out, and a result gone out of date is run again.

These are suggestions, not steps to take on your own. Explain the trade-off and let the person decide the scope.

## Reading results

- Speech recognition reflects the voices, sentences and recognizer tested: read it by voice, not only the percentage, since a name can come back right in one accent and wrong in another. Differences between the voices tested can show a useful pattern, but a few voices do not represent every speaker of an accent. What it was most often heard as is often the finding.
- First impression is what each model asked said to one prompt (one model, or four side by side with cross-model), not what people think; where the models differ is often the finding. Brand recognition is a model saying it knows something by that name, not proof that the thing exists or that the model is right.
- The US trademarks test finds possible conflicts; it is a textual and phonetic match, never legal clearance: say so once. Open the records rather than reading the counts: look at the mark, its status and its goods and services. Records in the project's classes come first, but records in other classes can still matter, because related goods can sit in another class; do not dismiss a record for its class or its similarity alone. The screening keeps up to 50 records a name, counted ones and weaker matches.
- Results belong to the workspace, not the project: a name tested in one of its projects has those results in every project that holds it, and `preview_tests` leaves out (and does not charge for) what the workspace already has or is running. A plan approved after part of it arrived from another project starts for less.
- Domains are as checked at that moment, and checking reserves nothing. "unknown" is not "taken". An asking price a marketplace showed is what it showed when checked, not a quote.
- Offer or open the recording player (`open_recording_player`) when hearing the names would help the person, or when they ask: one player for the relevant names, never one per name.

## Do not

- Start a run, or generate names, without showing the price and getting the person's agreement in the conversation.
- Split a run to keep each part under the workspace's approval limit.
- Change ratings merely to select candidates for a run. Pass the requested names, or a selection such as `finalists` or `in_play`, to `preview_tests`; use existing ratings that way when the person asks for a group like their finalists.
- Rewrite or invent a cursor; pass `next_cursor` back verbatim.
- Describe recordings you have not heard: the player plays to the person, never to you.
