# Webhound — An AI research agent priced by the dollar, not a subscription. Set a budget, get a cited Report or a sourced Dataset, and trigger it from inside Claude, Codex or Cursor via a native MCP server.

> Source: The Agents Index — https://theagentsindex.com/webhound (structured, researched, re-verified)
> Facts last verified: 2026-08-23

Webhound is a web-based AI research agent that runs an investigation against a dollar budget you set, rather than a subscription. Tell it a question or a dataset to build, set a spend limit, and it keeps searching, reading and verifying until the budget is used. Founded by Moe Khalil and Theo Schmidt as part of Y Combinator's Summer 2023 batch, the two-person San Francisco team launched Webhound in mid-2025 and has rebuilt pricing and interface several times since, with a public changelog documenting each change. It returns either a cited Report (sources, claims and working materials) or a structured Dataset (a table with a source trail on every value), exposes the same engine behind an HTTP API and ships a native Model Context Protocol (MCP) server so agents inside Claude, Codex, Cursor, Manus and other coding agents can trigger a research run, recall past findings and read the output directly.

| Fact | Value |
| --- | --- |
| Website | https://www.webhound.ai |
| Pricing | Pay-as-you-go, no subscription. New accounts start with $5 in credits that pre-fill a Flash Report. Flash floors start at $1 (about 15 minutes / ~1M input tokens). Pro floors start at $10 (about 2h30m). Deep coverage is $25 (about 6h15m). Budget can be topped up mid-run. |
| API | Yes |
| Best for | Developers and technical researchers who want an agent-triggered research tool, callable via MCP from inside Claude, Codex or Cursor, that returns either a cited Report or a structured, sourced Dataset, priced transparently by the dollar rather than an opaque credit system. |
| Not for | Buyers who need a specialized corpus such as academic literature (Elicit, Undermind) or US case law (GC AI), or who want an established, multi-year track record rather than a two-person team still iterating on its own pricing and interface. |

## Pricing

| Tier | Price |
| --- | --- |
| Pay-as-you-go | From $1 / no subscription |

## Verdict

Webhound's pitch is refreshingly literal. Give it a research task and a dollar figure, and it keeps searching, reading and verifying until the budget runs out, with no subscription, no seat count and none of the opaque per-action credit pricing some rival agents use. Its two output modes are a real point of difference inside this site's research-agents category. A Report reads like a cited answer, while a Dataset behaves more like a structured deliverable, with a source trail attached to every individual value rather than just the document as a whole, genuinely useful for market-mapping or lead-list work and not only narrative Q&A. Its MCP server is the other real differentiator. Claude, Codex, Cursor, Manus and a dozen other coding agents can trigger a Webhound run, recall past findings for free and read the output as a tool call, an integration none of this category's other listings currently offer as a first-class feature. The honest catches: this is still a two-person, YC-backed team that has rebuilt its own pricing and interface more than once since a mid-2025 launch as a free tool, its own blog is candid that building trust in AI-generated research is an ongoing problem it is actively working on rather than a solved one, and it makes no claim to a specialized corpus the way Elicit, Undermind or GC AI do. It researches the open web generally, same as Perplexity or GPT Researcher, just priced and delivered differently. Pick Webhound for agent-triggered research callable from inside Claude Code or Cursor, or for a sourced dataset rather than a narrative report. Look elsewhere if a specialized corpus or an established, multi-year track record matters more than novel pricing and delivery.

## How it works

1. Pick an output: Report, Dataset, Chain or Ask. Set a budget. Hound 1.0 controls the research effort, with Flash starting at $1 and Pro at $10. New accounts start with $5 in credits and a Flash prompt pre-loaded.
2. Write the brief: target, boundaries, deliverable, proof standard. Attach files or past Webhound sessions when they carry useful context. Optional Plan mode lets you hash out scope with the planner before the run starts.
3. Webhound plans and runs the investigation: browsing public web pages and PDFs, running Python for calculations and parsing, pulling from connected services or a custom API via stored credentials, and using dedicated scrapers for 14 platforms.
4. Steer the run via chat, register a check-in moment for the agent to pause on, or set a checkpoint where it waits for your input. Top up the budget if a lead is worth chasing further.
5. Inspect the result: cited Reports with traced claims and tool-chain popovers, or structured Datasets with a source trail on every value. Pipelines link sequential runs so later research builds on earlier findings.
6. Trigger runs and read results directly from Claude, Codex, Cursor, Manus or another MCP-capable agent, or call the API v2 from your own product. Publish any session as a permanent /p/{slug} link under one of four licenses.

## Who it's for

- Developers who want a research agent callable from inside Claude, Codex or Cursor via MCP, rather than a separate standalone chat product
- Analysts who need a structured, sourced Dataset (market maps, company lists, lead lists) rather than only a narrative Report
- Buyers who want transparent, pay-as-you-go pricing tied to actual compute rather than an opaque per-action credit system

## Strengths and weaknesses

- ✓ Transparent, per-run dollar pricing (roughly $1 per 15 minutes / ~1M input tokens) instead of an opaque credit system, with the budget toppable mid-run.
- ✓ A native, two-way MCP server lets Claude, Codex, Cursor, Manus and a dozen other coding agents trigger a run, recall past findings for free and read results directly, a real integration none of this category's other listings currently offer as a first-class feature.
- ✓ Structured Datasets carry a source trail on every individual value, not just a document-level citation list, genuinely useful for market-mapping or lead-list work.
- ✓ Every claim in a Report and every cell in a Dataset is traceable back to the raw tool chain (search, page visit, code run), so a buyer can audit a citation rather than trusting it.
- ✓ One free $5 Flash Report pre-fills the composer on a new account, so a buyer can produce a real run before paying anything.
- ✓ 14 dedicated platform scrapers (LinkedIn, Crunchbase, PitchBook, Zillow, Indeed, TripAdvisor, eBay, Yelp, Telegram, Airbnb, Pinterest, Google Flights, Expedia Hotels, plus Google Maps, Amazon) cover jobs the open web route would miss.
- ✗ A two-person team, per Y Combinator's company page, with no disclosed funding beyond its YC Summer 2023 participation, a smaller and less capitalized operation than several other listings in this category.
- ✗ The product's own public changelog shows pricing and surface changes since its 2025 launch: free at launch, subscription tiers, pay-as-you-go, Flash/Pro floors, then Deep-read retirement, plus a chat-first home rolled back to a composer. Iteration, but not yet a settled feature set.
- ✗ The company's own blog states plainly that building trustworthy AI research is an active, unsolved problem for them ("time to first trust"), not a claim of solved reliability, a real caveat for research meant to be cited without independent verification.
- ✗ No specialized corpus such as academic literature or US case law the way Elicit, Undermind or GC AI offer inside this same category. It researches the general open web, so it competes on delivery and pricing rather than domain depth.
- ⚠ Pricing and product surface have changed more than once since its 2025 launch; expect continued iteration rather than a fixed feature set.
- ⚠ No independent, third-party benchmark of Webhound's research accuracy or citation reliability was found.
- ⚠ A small, two-person team with no disclosed funding beyond its YC Summer 2023 participation.

## Key features

- **Dollar-budget pricing on Hound 1.0** — No subscription. Set a budget from $1 (Flash, about 15 minutes / ~1M input tokens) to $25 (Deep, about 6h15m) and the Hound 1.0 harness keeps researching until the budget is spent, with mid-run top-ups to chase a lead further.
- **Reports and Datasets** — Two output modes. A cited narrative Report for comparisons or recommendations, and a structured Dataset with a source trail on every individual value for market maps, lists and catalogs.
- **Native MCP server for coding agents** — Claude, Codex, Cursor, Manus, OpenCode, OpenClaw, VS Code, Hermes, Pi, Antigravity, Windsurf and Cline can trigger a run, recall past findings for free, check on runs in progress and ask follow-ups, all as MCP tool calls.
- **API v2 with 58 endpoints** — REST API at https://api.webhound.ai/api/v2 with Bearer auth, structured JSON responses and full coverage of research, extraction, pipelines, files, publications, authors and account.
- **Pipelines and Stacks** — Pipelines link sequential Reports and Datasets so a later run builds on an earlier one's findings. Stacks are pre-built, expert-designed research bundles, with the Founder Stack shipping as a ten-report sequence.
- **Traced claims and tool-chain popovers** — Every claim in a Report and every cell in a Dataset carries a structured trace (evidence, method, sources, confidence). Clicking opens the full tool chain: searches, page visits, code runs.
- **Code execution and stored secrets** — The agent writes and runs Python during a run for calculations, deduplication, charts and file generation, and uses stored credentials to call Notion, Slack, Dropbox, your own API or any service with an HTTP endpoint.
- **Publications and shared workspaces** — Publish any session or folder as a permanent link at /p/{slug} under one of four licenses (View Only, Attribution, Open, Open + Commercial). Other users can copy a publication into their own workspace with attribution preserved.
- **Mid-run steering, checkpoints and check-ins** — Redirect a run via chat while it works, register a check-in moment for the agent to pause and ask, or set a checkpoint where it pauses for input before continuing.
- **Service connections and 14 dedicated scrapers** — Direct access to LinkedIn, Reddit, X, YouTube, Google Maps, Amazon and Crunchbase, plus dedicated scrapers for Zillow, Indeed, TripAdvisor, eBay, Yelp, Telegram, Airbnb, Pinterest, PitchBook, Google Flights and Expedia Hotels.

## Use cases

- **Agent-triggered competitive research** — A developer building an AI product wires Webhound's MCP server into Claude Code, Codex or Cursor so their own agent can trigger a research run, recall past findings for free and read the cited result back as a tool call, instead of switching to a separate research app.
- **Building a sourced lead list or market map** — A GTM or research team asks Webhound for a Dataset, a structured table of companies, contacts or products with a source trail on every value, rather than a narrative Report that still needs manual data extraction.
- **Budget-capped due diligence** — A buyer sets a fixed $10-25 budget for a narrow research question, gets a cited Report back, and only spends more if the initial pass surfaces something worth chasing further, instead of committing to a monthly subscription for occasional use.
- **Pre-built expert research with Stacks** — A founder pays one price for the Founder Stack, ten sequenced Reports covering market, first customer, substitutes, wedge, downside and the other questions every founder gets pressed on, with prompts designed by domain experts rather than the buyer's own brief.

## Integrations

Model Context Protocol (MCP) for Claude, Codex, Cursor, Manus, OpenCode, OpenClaw, VS Code, Hermes, Pi, Antigravity, Windsurf and Cline · REST API v2 at https://api.webhound.ai/api/v2 (Bearer auth, 58 endpoints) · LinkedIn, Reddit, X, YouTube, Google Maps, Amazon, Crunchbase, Notion, Slack, Dropbox, plus a custom API via stored credentials · 14 dedicated platform scrapers: Zillow, Indeed, TripAdvisor, eBay, Yelp, Telegram, Airbnb, Pinterest, PitchBook, Google Flights, Expedia Hotels · Public web pages and PDFs (with Unpaywall DOI lookup for open-access papers)

## Sources

- https://www.webhound.ai
- https://www.webhound.ai/guide
- https://www.webhound.ai/pricing
- https://www.webhound.ai/mcp-setup
- https://www.webhound.ai/docs/api
- https://www.webhound.ai/changelog
- https://www.webhound.ai/news
- https://www.ycombinator.com/companies/webhound
- https://www.producthunt.com/products/webhound