Claygent: what it is, prompts that work and what it costs per row

What is Claygent?

Claygent is Clay's AI research agent: a column in a Clay table that reads a prompt, searches the web or visits pages, and writes a structured answer into the row. Clay's own FAQ puts it simply: "Claygents are agents in Clay that perform tasks on your behalf." It is the part of Clay that answers questions no database holds, such as "does this company hire SDRs?" or "what does its pricing page say?"

Clay's Claygent page describes two kinds. General-purpose Claygents run in tables and workflows, one row at a time. Account Agents (in beta, per Clay's page) run in Audiences and workflows across every account in a segment, and can draw on CRM and warehouse context. Table Claygents are on every plan including Free. Account Agents are listed for Launch, Growth and Enterprise. This guide is mostly about the table version, which is what most people mean by "Claygent" or "Clay AI agent".

The best use is last-mile discovery. Waterfall enrichment finds the contact and the firmographics; Claygent finds the unstructured fact, such as a hiring page, a press quote or a pricing model, that makes a message specific. For where that fits in a whole system, see our waterfall enrichment guide.

How does Claygent work?

Each run takes the inputs you map from the row, follows your instructions, optionally uses web access, and returns text or fields. Clay's Claygent Builder docs list the controls: business context that fills from workspace settings, attached documents, a web search toggle, a find contacts and jobs tool, and a model choice that spans frontier, open-weight and Clay-managed models. Clay's Claygent page says every decision includes a reasoning trace, which you can open to see why a cell says what it says.

The model matters. Clay's changelog introduced Claygent Neon on 15 July 2024 as "our most capable model yet" at the time, able to parse PDFs and split output across several columns instead of one text cell. Clay's current AI pricing doc lists its own models Helium and Argon for web research, plus OpenAI, Anthropic and Gemini models and, since a July 2026 changelog entry, the open-weight Kimi K2.6 and GLM 5.2 through Fireworks AI. Claygent Navigator, announced in Clay's 26 August 2025 product roundup, is the browser-enabled version: it clicks, fills forms, applies filters on search portals, parses PDF, XML and CSV files, and records each step in Claygent Replay. Clay's current AI pricing doc gives no credit price for Navigator, so test it on a few rows and check the estimate before a full run.

Since August 2026 the Claygent Builder also supports Skills: reusable instruction sets that, per Clay's changelog, "Claygent loads automatically at the right moment during a run". Put shared rules, such as your ICP definition or the answer-unknown-rather-than-guess rule, in a Skill instead of repeating them in every prompt.

Splitting output into columns is what makes a Claygent useful in a system. A single paragraph is hard to filter on. A column called "hiring_sdrs" with the values yes, no or unknown can drive a conditional run, a score or a routing rule.

How much does Claygent cost per row?

Clay's pricing page (checked 22 October 2026) says each run uses one Action plus Data Credits set by the model. An Action is a platform unit; Clay's credits doc says an Action costs a few tenths of a penny and Data Credits a few pennies. Clay's pricing page says Data Credits start at $0.05 each, and the real rate depends on your plan. Our Clay pricing guide has the plan maths.

About 80% of models are fixed price: a flat number of credits per task, known before you run. The rest, mostly heavy reasoning models used for multi-step web research, bill the actual model cost with no markup. For those, Clay withholds credits upfront, based on the 75th percentile of past runs, then refunds the surplus or deducts the shortfall. Clay's AI pricing doc gives these web research rates, as of October 2026:

Two cautions. First, the $0.05 figure is Clay's starting price, so your cost per credit may differ; treat the last column as a ceiling guide, not a quote. Second, Neon is listed by Clay as fixed price but without a figure in the current doc. The July 2024 changelog said two credits, which may be out of date.

A worked example: 2,000 target accounts, each checked for a hiring signal with Helium, is 2,000 credits (about $100 at the starting rate) and 2,000 Actions. Run the same list on Argon and it is 6,000 credits (about $300). If only the 400 accounts that pass a firmographic filter get the Claygent, the Helium bill drops to about $20. Gating the expensive column behind a cheap filter is the biggest saving, and it needs no clever prompting.

What does a plan buy? Clay's pricing page lists, as of October 2026, 500 Actions and 100 Data Credits a month on Free, 15,000 Actions and 2,500 Data Credits on Launch (one part of the page shows 3,000), and 40,000 Actions and 6,000 Data Credits on Growth. Because every row needs one Action plus its credits, Free covers about 100 Helium rows or 200 GPT-5 Nano rows a month, Launch about 2,500 to 3,000 Helium rows, and Growth about 6,000 Helium rows or 2,000 Argon rows, before any other enrichment draws on the same credits. Clay's pricing page also says you are not charged Actions or Data Credits when an enrichment returns no result. It describes rollover of unused Data Credits on Launch and Growth both as up to twice the monthly amount and as up to one month's credits, so check the current terms on your plan.

Bringing your own API key removes the Data Credit cost for that step but still uses one Action, per Clay's credits doc. You then pay the model provider directly, and Clay shows only Actions. Clay adds that its own keys run roughly twice as fast because of higher rate limits.

What are good Claygent prompts for GTM research?

Clay's prompt guidance, as summarised in its University lessons, is to give context and the goal, step-by-step instructions and an explicit output format. The six prompts below follow that shape. Each names the inputs, tells the agent to answer "unknown" rather than guess, and asks for a source. The inputs in curly brackets are columns from your table.

1. Hiring signal. Prompt: "Visit {company_domain}. Find its careers page. Does it have open roles titled SDR, BDR or Account Executive? Return JSON with open_sales_roles (number), example_title, source_url. If no careers page is found, return unknown." Output columns: open_sales_roles, example_title, source_url. Use it as a trigger for outbound to sales leaders.

2. Pricing model. Prompt: "Open the pricing page of {company_domain}. Return pricing_model as one of: free, self-serve, sales-led, usage-based, not_found. Quote the sentence that supports it in evidence." Output columns: pricing_model, evidence. Useful for segmenting product-led from sales-led accounts.

3. B2B or B2C filter. Prompt: "From the homepage of {company_domain}, say whether this company sells to businesses, consumers or both. Return business_model and one line of evidence." This is cheap, so run it on a small model such as GPT-5 Nano before spending on anything heavier.

4. Tech stack check. Prompt: "Check {company_domain} for mentions of HubSpot, Salesforce or Attio on its site, careers page or documentation. Return crm_mentioned and source_url, or not_found." Treat it as a hint, not proof. For a verified stack, a technographics provider is stronger, which our GTM tech stack guide covers.

5. Personalised first line. Prompt: "Using only the facts in {recent_news} and {hiring_signal}, write one sentence, under 25 words, that references something this company did in the last 90 days. If neither field has a fact, return skip." Output column: first_line. Feeding it earlier Claygent outputs, rather than letting it search again, is cheaper and easier to audit.

6. Lead score with reasoning. Prompt: "Score this account from 1 to 10 against our ICP: {icp_summary}. Return score and a one-sentence reason. Use only the fields provided." Output columns: score, reason. Clay's own example asks for a 1 to 10 rank from employee count, industry, tech stack and growth signals; we add the reason column so a human can sample it.

Always give the agent a permitted failure. A prompt with no "unknown" or "not_found" option pushes the model to fill the field with something plausible.

How accurate is Claygent, and how do you check it?

Clay publishes no accuracy rate for Claygent, and we have not tested one, so treat any figure you see as unverified. The structural limits are documented. Claygent cannot reach paywalled or password-protected pages, Understory's guide notes it may be blocked by anti-scraping measures, and it writes text, so verified emails and phones still come from data providers.

The failure to watch is a confident wrong answer. Check results in this order:

1. Test on the free tier first. Clay's Builder docs say you can test up to 10 inputs at a time without spending credits.
2. Add a source_url column and open ten cells. If the URL does not support the claim, the prompt needs work.
3. Add an "unknown" option and count how often it appears. Zero unknowns on a hard question is a warning sign, not a success.
4. Compare two models on the same 10 rows. The Builder supports comparing versions or models.
5. Sample 20 rows by hand each time you change the prompt, and keep the sample in a sheet so you can see drift.

Hard-to-verify answers should never go straight into an email. Use them to rank and route, and let a verified field or a human check what is actually sent.

How do you set up a Claygent column step by step?

Menu labels in Clay change, so follow the names in your own workspace. The sequence below follows Clay's Builder docs.

1. Open a table that already holds a company domain column, and filter it to a test set of 10 rows.
2. Open Agents in the left navigation to use the Claygent Builder, or add an AI column in the table. Start from scratch, from a template, or describe the agent to Sculptor, which generates the prompt, variables and output format.
3. Write the goal, the steps and the output fields. Map variables such as {company_domain} to table columns.
4. Toggle web search on only if the answer lives online, and pick the cheapest model that passes your test rows.
5. Run the free test on up to 10 inputs and check each against the source.
6. Deploy with Use a saved Claygent in the table. Updates to the saved agent then flow to every table that uses it.
7. Leave auto-run off until the first batch is checked, then add a condition so only qualifying rows run.

If you want this built into a full system rather than a single column, the Clay agency service covers Claygent research, scoring and CRM sync.

Claygent vs Clay's other AI features: what is the difference?

Clay has two AI column types. A Claygent column does web research: it goes and fetches new information. A Use AI column, which Clay's changelog names alongside Claygent whenever it adds models, transforms text you already have, for example turning a job title into a persona. Clay's AI pricing doc prices the two tasks separately, as web research and content generation, and generation is cheaper on the same model: Claude 4.5 Sonnet is 1.5 credits for generation and variable for web research, and GPT-5 Nano is 0.2 against 0.5. If the answer is already in your table, use Use AI and do not pay for browsing.

Account Agents are the other branch. Clay's page says they charge one Action per record with variable credits only, and CRM or warehouse exports add no extra charge. They suit segment-wide, ongoing research; table Claygents suit one-off enrichment of a list.

Claygent vs ChatGPT or Claude through the API: which is cheaper?

Calling OpenAI, Anthropic or Perplexity directly, from n8n, a script or Claude Code, gives you full control of the prompt, retries and storage, with no Action count. As of October 2026, Anthropic's web search tool costs $10 per 1,000 searches plus tokens, and OpenAI's web search tool for reasoning models is also $10 per 1,000 calls plus search content tokens. That is $0.01 a search before tokens, against about $0.05 a row for Helium at Clay's starting credit price. But Anthropic's docs note that comparative research can use 10 or more searches, and you build the row mapping, retries and CRM sync yourself. Claygent gives you the table: row mapping, conditional runs, waterfalls, CRM sync and a UI a non-engineer can edit.

Choose Claygent when the research is one step in a Clay workflow and you value speed of build. Choose the API when you run very large batches, need custom logic or want one pipeline outside Clay. For search rather than browsing, a dedicated tool such as Exa is a third option. You can also use your own key inside Clay, which keeps the table and moves the model bill to your provider. Our Clay alternatives comparison covers when to leave Clay entirely.

See how Tentt builds Claygent research into Clay systems

Talk to Tentt about a Clay build