Claygent: What It Is, What It Costs, How to Prompt It

Facts checked October 16, 2026. Every Clay fact below comes from Clay's own pages: the Claygent page, the pricing page, Clay University's AI pricing, AI models, Claygent Builder, global agent and Actions and Data Credits docs, Clay's changelog and announcements, and Clay's guides and blog. The Navigator price comes from Clay's launch post of August 20, 2025, the newest Clay page that states one. Clu did not run a project through Claygent for this review, so this page has no accuracy numbers. Clu's own plans are on the pricing page.

What is Claygent?

Claygent is the AI agent inside Clay that researches the web, one row at a time. Clay's definition: "Claygents are agents in Clay that perform tasks on your behalf." You add Claygent as an AI column or a workflow step, point it at the inputs it needs (a company domain, a name, a title), write the question, and pick a model. It visits real pages for each row and writes the answer into one or more columns.

Clay lists the jobs it sees most: account research and meeting prep, finding data no provider sells, personalizing outbound from news or job posts, scoring and qualifying lists, ranking leads from 1 to 10 and finding the right contact. Clay's Claygent page counted 5,060,235,046 Claygent runs on October 16, 2026.

Claygent or Use AI?

Clay has two AI columns. Clay's FAQ says "Use AI" is for "offline tasks like content generation," while Claygent "has access to a range of tools" for last-mile research on the web. If the answer sits on a web page, use Claygent. If you only need text written from data already in the row, use Use AI. That split is why Clay lists two prices for each model, one for content generation and one for web research.

Two kinds of Claygent

From Clay's Claygent page, read October 16, 2026. Clay University lists Account Research Agents as "in open beta," and you cannot bring your own model key for them during the beta.

Claygent at a glance

Is Claygent expensive?

Claygent is cheap per row on Clay's fixed-price models and expensive in three places: heavy models, browser runs and plan minimums.

Every Claygent prompt costs 1 Action. Clay says Actions start at "less than $0.01 each," and Launch includes 15,000 a month. The bigger line is data credits, set by the model. Clay's AI pricing doc puts 80% of models, including all of Clay's own, on fixed pricing: a set number of data credits per row. The rest, mostly reasoning models, are variable, and Clay charges "the actual cost of each run, with 0% markup." Clay withholds an estimate based on the 75th percentile of past runs for that model, times the number of rows, then refunds or charges the difference once the run finishes. If your balance reaches zero mid-run, processing stops, and Clay says you are never charged beyond your balance.

Clay's published credit costs per row, from the table Clay labels "Select model pricing reference" (read October 16, 2026). Claygent is the web research column:

The table has two gaps. It prices 16 models, while Clay's model list has 36 text models, including GPT 6 Sol, Claude Opus 5 and Gemini 3.1 Pro, which the table does not price. And five models it does price (GPT-5.1, Claude 4.5 Sonnet, Claude 4.6 Opus, Gemini 2.5 Flash and Gemini 2.5 Flash Lite) are missing from Clay's "complete list of available models." For any model outside the table, run 10 or 50 rows first and read the charge per row.

On top of research, contact data has its own price. Claygent writes text answers; a verified email or phone number comes from Clay's data marketplace on more data credits. Clay's docs say each fully enriched record typically costs 6 to 20 data credits, and phone number enrichment starts on Launch. The Clay waterfall and Actions breakdown prices that side row by row.

The plan sets the floor. Free gives 100 data credits and 500 Actions a month, enough for about 100 Helium rows. Launch is $185 a month (2,500 data credits, 15,000 Actions) or $167 billed annually. CRM integrations start on Growth at $495 a month ($446 annually). Top-ups on Launch and Growth cost 30% more than plan credits. The Clay pricing guide covers the two meters in depth.

What 1,000 rows of Claygent research cost

At Clay's published entry rate of $0.05 a data credit, a 1,000-row web research job costs:

Two things move the real number. An Argon job of 3,000 credits is more than Launch's 2,500 monthly credits, so a Launch account pays for the rest at the 30% top-up premium. And variable-priced models can only be estimated: Clay's own advice for complex research prompts is to run them "on 10 or 50 rows" first and look at the charge per row.

How to build a Claygent, step by step

Claygent Builder, launched May 7, 2026, is where Claygents are built, tested and deployed. Clay says it is "free on every plan, even trials." The steps, from Clay's Builder doc:

1. Start the agent. Open Claygent Builder and start from scratch, describe the agent to Sculptor in plain English, or pick a template. You can also save the prompt of an existing table column as a Claygent.
2. Give it context. Add your business context and any documents it should read, switch on web search, and turn on the find contacts and jobs tool if the job needs people.
3. Pick a model. Start with Helium or GPT-5 Mini at 1 credit a row on web research. Move up only when test rows show the lighter model failing.
4. Add skills. Skills are reusable instructions saved for the whole workspace, for example your rules for scoring a lead. A Claygent loads a skill only when a step calls for it, and Clay's FAQ says a run with skills "is still one action plus data credits."
5. Test for free. Run up to 10 inputs at a time without spending credits, swap inputs, and compare models on the same rows.
6. Deploy. Click "Add to table" in the Builder, or add an AI column to a table and pick the saved Claygent. Turn off auto-run before you add the column, as Clay's prompt guide advises, then run 10 or 50 rows and check the charge per row before the full list.

Builder also keeps a version history with one-click rollback, so a bad prompt change can be undone.

Claygent prompt examples

Clay's April 2026 prompt guide treats a prompt as a contract with "three parts: context, named sources, and a locked output schema," with "Not found" as the floor. The three prompts below follow that shape. Paste one into a Claygent, then insert your own columns where the braces are.

1. Does this company sell to hospitals?

Context: You research B2B software companies for a sales team that sells to hospitals.
Input: the company website {Company Domain}.
Task: Decide whether this company sells to hospitals or health systems.
Sources: Read the homepage, the customers or case studies page, and the industries or solutions page on {Company Domain}. Do not use other websites.
Output as JSON:
{"sells_to_hospitals": "Yes | No | Not found",
 "evidence": "one sentence copied from the page",
 "source_url": "the page you read it on"}
If the pages do not answer the question, return "Not found" in every field. Do not guess.

2. Which ERP do they name in their job posts?

Context: You qualify manufacturing companies for an ERP integration product.
Input: the company website {Company Domain}.
Task: Find which ERP system the company uses, as named in its own job posts.
Sources: The careers or jobs page on {Company Domain}, and the job posts linked from it.
Output as JSON:
{"erp": "the product name as written, or Not found",
 "job_title": "the job post that names it",
 "source_url": "the job post URL"}
Only report an ERP that a job post names. If none does, return "Not found".

3. Did they open a new office this year?

Context: You look for expansion signals for an outbound campaign.
Input: {Company Name} and {Company Domain}.
Task: Find whether the company announced a new office or location in 2026.
Sources: The newsroom, press or blog pages on {Company Domain}.
Output as JSON:
{"new_office_2026": "Yes | No | Not found",
 "city": "the city named, or Not found",
 "announcement_date": "the date on the post",
 "source_url": "the post URL"}
Only count announcements dated 2026. If you find none, return "Not found".

Run each one on 10 rows in Claygent Builder first. If more than one or two answers come back without a source URL, tighten the sources line before you pay for the full list.

Models, Navigator and prompts

Models. Clay's AI docs list 36 text models for its AI features, with the model picked per Claygent: four of Clay's own (Helium, Neon, Argon and Clay Navigator), 15 from OpenAI (up to GPT 6 Sol and GPT 6 Luna), six from Anthropic (up to Claude Opus 5 and Claude Sonnet 5), six from Google's Gemini line, Kimi K2.6 and K3, and three GLM models from Z.ai. Clay added the open-weight Kimi K2.6 and GLM 5.2 on July 16, 2026, for research "at lower cost." You can also connect your own API keys and skip Clay's data credits, and each run still counts 1 Action, though Clay says tasks run through its APIs are "up to 2× faster" because of higher rate limits. Clay's guide sums up the trade: a heavier reasoning model "returns sharper judgment on messy pages but costs more per row," and a lighter one "needs an even tighter prompt."

Navigator. Navigator is Claygent with a browser, and it now appears in Clay's model list as Clay Navigator. Clay's August 2025 launch post says it handles "clicks, form fills, filter toggles, pagination, and scrolling," built on "GPT-5's deep research capabilities," which opens up public data that sits behind a search form, such as government registries. Each run comes with a Claygent Replay showing what Navigator clicked and filled. The same post priced Navigator at 6 credits a run, said private API keys were not supported for it, and called it "still early." Navigator does not appear in Clay's current pricing reference, so check the cost on a few test rows.

Prompts. Clay's April 2026 prompt guide puts accuracy on the prompt more than the model, and gives three rules: force a structured schema, require the source URL, and let the model answer "Not found." It also recommends turning off auto-run before you add the column, so the agent does not fire on every row before you have tested it. The three examples above follow all three rules.

Claygent timeline

Where Claygent works well

• Data points no provider sells. "Does this company sell to hospitals?", "Which ERP do they mention on their careers page?", "Did they announce a new office this year?" These are the questions Claygent was built for, written straight into a column.
• Research at list scale on cheap models. At 0.5 to 1 credit a row, a 1,000-account research pass costs $25 to $50 in credits on Clay's entry rate.
• Hard-to-reach public data. Navigator works through search forms and filters a plain scraper cannot, and the replay makes each answer auditable.
• Model choice. Thirty-six text models, plus your own keys, let a team match cost to the difficulty of each question.
• Free iteration. Claygent Builder tests up to 10 inputs at a time at no cost, and its version history rolls back a bad prompt.
• Everything else is in the same table. The research column sits next to Clay's enrichment waterfalls, scoring formulas and exports.

Where Claygent struggles

• Guessing. Clay's own guide warns that "a model with no permission to fail will fabricate to fill the field." Without a citation column and a "Not found" option, wrong answers look exactly like right ones.
• Paywalls and logins. A Clay University lesson says Claygent reaches public data "with the exception of information hidden behind paywalls or password-protected areas."
• Cost you only see afterwards. Variable-priced models bill after the run. Clay withholds an estimate first and stops at a zero balance, so a run cannot overdraw the account, but the per-row cost is unknown until the test rows finish.
• Auto-run on a new column. If auto-run is on, the column fires on every row as soon as it exists, which Clay's guide flags as paying "for a run you have not validated yet."
• Someone has to build it. Prompts, schemas, model choice and table logic are a skill. Clay's docs point teams to a directory of Clay experts and agencies for help with setup.
• Prices lag the model list. Clay's price table covers 16 of its 36 text models, so the newest models have no published per-row price until you test them.

Clay's other agents: Account Agents, the global agent and Sculptor

Clay's agent line is wider than the research column.

• Account Agents, in Clay's description, reason over what you know about an account, decide the next best action and remember what they learned. They read CRM data, warehouse tables and Gong call transcripts in Audiences, keep a persistent memory per account, and write fields back to the CRM only after a person approves. Clay University lists Account Research Agents in open beta on Launch, Growth and Enterprise, the plans with Audiences, at 1 Action per record plus variable data credits on the Claygent pricing model. You cannot bring your own model key during the beta.
• Clay's global agent, announced October 8, 2026, is a chat agent that plans and runs GTM plays across Clay: it builds and updates workflows, analyzes Audiences, finds companies and people, and drafts campaigns. It is in Alpha on Launch, Growth and Enterprise, and chatting with it is free while in Alpha; the work it sets up uses credits. It will not send a campaign, and it will not create or edit a table. It runs on the same engine as the Clay CLI.
• Sculptor is Clay's assistant inside tables and Claygent Builder, where it builds and edits tables and writes Claygent prompts, and its Analyst mode answers questions about a table in natural language. The global agent took over Sculptor's entry points in Workflows, Search and the home screen.
• Clay Sequencer sends cold email from Clay: you can buy inboxes and domains inside Clay, warming starts at setup, and Clay agents write "lead-specific copy" from your enrichments. When a lead replies, Clay drafts the follow-up and posts it in Slack for the rep to approve. The Sequencer and pricing pages mention email only. LinkedIn steps and call steps do not appear on either page.
• Clay MCP, generally available as MCP for Reps since July 14, 2026, brings Clay search, enrichment and Functions into Claude, ChatGPT, Claude Code, Codex, Microsoft 365 Copilot and Glean, with per-rep credit budgets. The Clay MCP server review covers it.

What Claygent leaves to other tools

Claygent answers questions. A full outbound motion needs a few more jobs, and Clay prices each one separately:

1. Verified contact data. Emails and phone numbers come from Clay's marketplace on data credits; phone numbers start on Launch.
2. CRM sync. Native CRM integrations start on Growth, $495 a month.
3. Sending. The Sequencer covers email; LinkedIn and call steps need another tool.
4. Upkeep. A table that runs weekly needs someone to own its prompts, models and credit budget.

Each of the four is a plan tier, a credit line or a separate tool.

Alternatives for research per row

Claude with web search

For one-off research, a chat assistant does the same job by hand. Claude Pro is $20 a month, or $17 billed annually, and reads a company's site and news on request. For research across a list, Anthropic's API charges $10 per 1,000 web searches plus model tokens (claude.com/pricing, read October 16, 2026). You write the loop and the output schema yourself, and Clay's three prompt rules apply just the same.

Clay's lighter models

If you already pay for Clay, the cheapest alternative is a cheaper model in the same column. GPT-5 Nano at 0.5 credits a row, or Helium and GPT-5 Mini at 1, handle yes-or-no questions with a cited source well enough to test first, before you reach for Argon or a variable-priced model.

Clu (our product)

Clu puts per-row research inside agents. The Research skill reads a company's website, news and careers page and writes back what they do, recent news, hiring and who to contact, with the source on every line. The Qualification skill scores each lead from 0 to 100 against your ICP, gives the reasons and checks whether the lead is already in your CRM. Research costs 3 credits a lead for the company or the contact and 5 for both, qualification costs 5, and both run on every paid plan from Starter.

Clu, our GTM agent, turns one sentence into an agent: "every Monday, find new heads of sales at 50 to 300 person SaaS companies, research and qualify them, get verified emails and phone numbers for the ones above 70, add them to a sequence and post the run in Slack." The contact data comes from a waterfall across 25+ providers, and a missed email or phone number costs 0 credits. Pro includes 5 active agents on a schedule or a CRM trigger; Enterprise runs unlimited agents. See how agents work, or the full Clu vs Clay comparison.

The verdict on Claygent

Claygent gives a GTM team web research on every row, with 36 text models to choose from, a browser agent for search forms, free test runs in Claygent Builder and a published credit price for the models most teams start with. On Clay's lighter models it costs about 1 credit a row. Buy it if you have someone who will write schemas, demand citations and tune the model per question; on those terms it answers questions no data vendor sells.

Look elsewhere if the job is the same every week and the goal is a qualified, verified, sequenced list, because each extra step in Clay (phone numbers, CRM sync, call steps) is another plan tier or another tool. More options are on the Clay alternatives page, and Clu's plans are on the pricing page.

See pricing

FAQ

What is Claygent?

Claygent is Clay's AI research agent. It takes inputs from a Clay table or workflow, follows a prompt you write, reads the web for each row and writes a structured answer into one or more columns. It is included on every Clay plan, including Free, and you build and test Claygents in Claygent Builder at no cost.

Is Claygent free?

Claygent is on every Clay plan, including Free, and building and testing in Claygent Builder costs nothing. Running it does cost: each row is 1 Action plus data credits, and Clay's Free plan includes 500 Actions and 100 data credits a month, enough for about 100 rows on Clay's Helium model.

How much does Claygent cost?

Each Claygent run costs 1 Action plus data credits set by the model. Clay's AI pricing page lists fixed web research prices of 0.5 credits a row on GPT-5 Nano, 1 on Clay's Helium, GPT-5 Mini, Gemini 2.5 Flash and Gemini 2.5 Flash Lite, and 3 on Clay's Argon. Variable-priced models are billed at their actual token cost with 0% markup. Clay says data credits start at $0.05 each.

Is Claygent expensive?

Per row, Claygent is inexpensive on light models: about $0.05 of credits a row on Helium at Clay's entry rate. It gets expensive with heavy reasoning models, Navigator runs, re-runs from auto-run, and the plan tiers needed for phone numbers (Launch, $185 a month) and CRM sync (Growth, $495 a month).

What is Claygent Builder?

Claygent Builder is Clay's hub for building, testing and deploying Claygents, launched May 7, 2026 and free on every plan, trials included. You describe the agent to Sculptor or start from a template, add skills and documents, test up to 10 inputs at a time for free, then add it to a table.

Which models does Claygent use?

Clay's AI docs list 36 text models for its AI features: Clay's own Helium, Neon, Argon and Clay Navigator, plus models from OpenAI, Anthropic, Google, Kimi and Z.ai, up to GPT 6 Sol, Claude Opus 5 and Gemini 3.1 Pro. Clay publishes per-row prices for 16 of them.

Can I use my own OpenAI API key with Claygent?

Yes. Clay says you can always use your own API keys instead of Clay data credits, and each run still counts 1 Action. Clay says its own keys run up to 2× faster. At launch, Navigator did not support private keys, and Account Research Agents do not accept your own model key during their open beta.

What is Claygent Navigator?

Navigator is Claygent with a browser, listed in Clay's models as Clay Navigator. Clay's launch post says it clicks, fills forms, toggles filters, paginates and scrolls to reach public data behind search forms, and records a replay of every step. Clay priced it at 6 credits a run in August 2025.

Can Claygent find emails and phone numbers?

Claygent researches and writes answers; verified emails and phone numbers come from Clay's data marketplace on separate data credits. Clay's docs say a fully enriched record typically costs 6 to 20 data credits, and phone number enrichment is not on the Free plan.

What are Clay agents?

Clay's agents are Claygents, which research row by row in tables and workflows, Account Agents, which reason over everything known about an account, keep a memory and recommend the next action, and Clay's global agent, a chat agent that builds and runs workflows across Clay. Account Research Agents are in open beta and the global agent is in Alpha, both on Launch, Growth and Enterprise.

Sources

• Clay, Claygent, read October 16, 2026
• Clay, Pricing, read October 16, 2026
• Clay University, AI pricing, AI in Clay, Claygent Builder, Global agent, Account research agents and Actions and Data Credits, read October 16, 2026
• Clay University, Sculptor and the lesson Claygent: AI web scraper, read October 3, 2026
• Clay, changelog entries of July 15, 2024 and July 16, August 3 and August 17, 2026, read October 16, 2026
• Clay Community, Claygent Builder is live, May 7, 2026, and Clay's new global agent, October 8, 2026, read October 16, 2026
• Clay, How to use Claygent for prospect research, published April 30, 2026, read October 16, 2026
• Clay, Introducing Claygent Navigator, published August 20, 2025, read October 16, 2026
• Clay, Account Agents, Sequencer and MCP for reps, read October 16, 2026
• Anthropic, Claude pricing, read October 16, 2026