User guide
Pick an output, set a budget, hit send. This is how the four modes — Reports, Datasets, Chains, Ask — actually work.
30-second tour
Webhound runs autonomous research for you. The home page is one composer. You pick what you want to make, type a brief, set a budget, and it goes. There are four things you can make:
| Mode | What you get | When to pick it |
|---|---|---|
| Report | A cited document with a full source trail | A question whose answer is prose |
| Dataset | A spreadsheet of rows you defined | A question whose answer is a list |
| Chain | A pipeline of reports and datasets that run in sequence | The output of step 1 should drive step 2 |
| Ask | Q&A over research you've already done | You don't need new research, just an answer |
Reports and Datasets are sessions that run autonomously until your budget runs out. Chains are pipelines of those sessions. Ask is a chat that reads your existing work without spending on new research.
Your first run
Pick the output
Click the Output pill on the composer to choose Report, Dataset, Chain, or Ask. The placeholder updates so you know what to type.
Type the brief
Be specific. Tell it what to find and how the answer should look. The first paragraph of your message is the brief — everything below it is constraints.
Set a budget
Click the Budget pill. Webhound enforces a minimum per output type and model (see Pricing & budgets). For your first Report try $10; for a Dataset try $5.
Send
The agent confirms the plan in chat, then runs. You can leave the page — your session keeps running. You'll get a notification when it finishes.
Things you might not know
A handful of capabilities that exist but are easy to miss. Each of these can save a full new run, a copy-paste, or a manual workflow.
Use a past session as context for a new one. Drag any session from the workspace right onto the chat box, type @ in chat to pick one by name, or just describe what you want — “look at my last 5 sessions on competitor pricing”. Folders work the same way.
Send guidance to a session while it's running. A running session isn't locked. Type into the chat below the status bar — “focus on Europe”, “skip the historical context”, “add a funding round column” — and the planner picks it up on the next cycle.
Add budget mid-run with a note. When you top up a running session, you can attach a short message like “spend the extra $10 on European-specific sources”. The planner reads it as a directive for the additional work.
Save API keys as Secrets so the agent can call any service. Open Account → Secrets and save any key you want the agent to use (notion_token, slack_webhook, etc). With code execution, the agent can then run real integrations for you — “post these findings to our Slack”, “upload this report to my Dropbox”, “add a row to my Notion DB for each company”. You can also paste a key into chat and tell the agent to save it.
Switch to Ask mode for one-off follow-ups. On any finished report or dataset, switch to Ask in the mode dropdown to ask things like “summarize the top 3 findings” or “how many rows have no email”. Webhound answers from the existing content and only goes back to the web if it actually has to — much faster and cheaper than starting a new run.
Browse and roll back to past versions of a report. Every time the assembler builds the final output, that version is saved. Scroll to the bottom of any output document for the version timeline; click any past version to view it or roll back.
Click any claim or cell to see the work behind it. In a report, click any colored claim pill. In a dataset, click the small icon next to a cell. The popover shows the evidence, sources, and confidence — and a “View tool chain →” link that opens every search, page visit, and code run that produced that single fact.
Copy any public publication into your workspace. Find someone's public publication in Explore that's relevant, click Copy to workspace, and the whole thing — sessions, documents, datasets, sources — lands in a new folder in your workspace with attribution back to the original. Pick up where they left off instead of starting over.
Reports
A Report is a cited document. The agent searches the web, reads pages and PDFs, builds working notes on each angle of your question, then assembles a final document with inline citations. Every claim points to the page it came from.
A finished Report has four tabs:
- Document — the final write-up. Click any sentence to see how it was found.
- Claims — every factual claim with its sources. Search across them semantically.
- Activity — the live agent log. What it searched, what it read, what it wrote.
- Audit — an optional fact-check pass that re-verifies claims against their sources.
Higher budgets buy more depth: more cycles, more sources, longer documents. You can add more budget mid-run and the agent picks up where it left off.
Plan vs One-Shot
The Mode pill chooses how the run starts.
- One-Shot — sends the brief straight to the agent. It starts researching immediately.
- Plan — the agent asks you 2–4 clarifying questions first, then runs.
Examples of briefs that work well:
"Compare the actual clinical outcomes of GLP-1 drugs for weight loss in people under 30, based on published trials and patient-reported data. Lead with what's settled vs. still contested."
"Pricing, features, and user sentiment for the top 5 PM tools used by remote startups in 2026. Format as a decision memo for a CEO, not an academic paper."
"What does the public record say about Alex Karp's strategic playbook over the last decade? Cite primary sources only — interviews, letters, transcripts. Skip secondary commentary."
Auditing claims
After a Report finishes, click Audit Claims in the toolbar. The auditor opens every cited source independently and checks whether each claim is actually supported, contradicted, or unverified. Anything it can't confirm gets flagged in the Audit tab; minor inaccuracies are repaired in place.
Audits get their own budget. You can also message the auditor mid-run to focus on specific sections or claim types — same composer, same right-side chat.
Datasets
A Dataset is a spreadsheet you describe. You say what kind of entity you want and which fields matter, the agent designs a schema (or you can edit it), then it scours the web filling rows. Every cell links to the source it came from.
A finished Dataset has these tabs:
- Data — the table. Click any cell to see how it was found.
- Sources — every page the agent visited, with what it pulled from each.
- Schema — column names, types, and descriptions. Editable.
- Activity — the live extraction log.
Mode behaves the same way: One-Shot lets the agent design the schema and run. Plan shows you the schema first so you can adjust columns before extraction starts.
Briefs that work well for Datasets:
"Every Y Combinator S25 company. Columns: name, one-line description, founder names, funding amount, primary market."
"Michelin-starred restaurants in Tokyo. Columns: name, stars, cuisine, neighborhood, price tier, reservation difficulty."
"Open-source AI inference servers with more than 1k stars. Columns: project, license, primary language, notable backers, last commit date."
Add Column & refresh
A finished Dataset isn't frozen. Two operations let you keep building on it:
- Add Column — describe a new field; the agent revisits each row's sources and fills only that column. Good for "I should also know X about each of these."
- Refresh — re-runs extraction against the same sources to catch new info. Good for time-sensitive datasets.
Both can be paused mid-run. They use the same budget pill as a fresh extraction.
Chains
A Chain is a pipeline of Reports and Datasets that run in sequence. Each step's output is fed to the next as context, so step 2 sees what step 1 produced.
Picking Chain from the Output pill lands you on a builder page. A co-pilot drafts the steps from your brief, you review, edit budgets and models, and click Start chain. Nothing runs until you confirm.
"Step 1: build a dataset of every YC AI healthcare company in 2025-2026 with funding stage and primary product. Step 2: write a market map report ranking them by traction and addressable market, citing the dataset."
Each step has its own model and budget. The Total budget pill on the home composer is optional — if you set one, the co-pilot distributes it across the steps it proposes.
Ask
Ask is a chat that reads what you've already researched and answers questions about it. It doesn't start new sessions, doesn't hit the open web — it works only with the material you point it at.
You can scope an Ask conversation to:
- Specific past sessions (attach with @ in the composer)
- Files you've uploaded (PDFs, docs, spreadsheets, transcripts)
- A whole folder (the Working in pill on the composer)
Two ways to start an Ask:
- From the home composer — pick Ask, attach what you want to query, type the question. Lives at
/ask/...and shows up under Chats in Recents. - From inside a session — flip the right-side composer to Ask mode to ask questions about that specific Report or Dataset without starting a new run.
The composer
The composer is the single text box on the home page (and the one on the right side of every session and chain). The pill row underneath is the entire control surface.
- Output — Report, Dataset, Chain, or Ask. Determines what the brief produces.
- Mode — Plan or One-Shot. Plan asks you a few questions first; One-Shot starts immediately. Hidden for Chain and Ask.
- Model — Auto, Pro, or Flash. See below for what Auto picks. Includes a Deep read toggle that lets the agent hold much more of each page in context per pass, so it can reason over more material at once (~3× cost and runtime).
- Budget — dollar cap for the run. The pill turns amber if you set it below the floor for that output + model.
- Save to — workspace folder. New sessions land here. For Ask, this becomes Working in — the folder Ask reads from.
- + — attach files, attach past sessions, or start from a CSV (Datasets only).
- Advanced — minimum spend, audit settings, schema overrides. You almost never need this on first runs.
Hover Send before clicking — the tooltip tells you exactly what's about to start (e.g. "Start research · $10 budget") so there are no surprises.
Auto resolution
The default Auto setting picks the right model for the job:
- Reports → Pro (smarter, better reasoning over long documents)
- Datasets → Flash (faster, cheaper per row, the verifier still checks every row)
The pill shows what Auto resolved to: Auto · Pro for Reports, Auto · Flash for Datasets. Override only if you have a reason — Pro on Datasets is great for niche domains, Flash on Reports is good for first drafts.
Workspace & Recents
The left panel has two tabs.
Recents is everything you've recently touched, newest first. Filter chips up top: All · Reports · Datasets · Chains · Chats. Chains appear as a parent group with their step sessions indented underneath, so you can jump from a chain into any of its child Reports or Datasets.
Workspace is your file system: nested folders, sessions, uploaded files. Drag sessions between folders. Right-click anything for quick actions including publishing.
Files & @-mentions
Two ways to give a session more context.
Files. Click + on the composer or drag onto a folder in Workspace. Supported: PDF, Word, Excel, CSV, text, markdown, VTT (subtitles). The agent reads the contents and uses them as primary context.
Sessions. Type @ in the composer to bring up a picker of past sessions. Selecting one attaches its document or table to the current run. Useful when you're starting research that builds on something you already did.
Mid-run guidance
A running session has the same composer on the right side of the page. Type into it during the run; the agent picks your message up on the next planning cycle and adjusts.
Useful guidance looks like:
"Focus on European markets, drop Asia."
"Skip the historical background — only 2025-2026 matters."
"Add a column for funding round."
Works during research, extraction, and auditing. The right-side chat is where you steer.
Code execution
The agent can run Python in a sandbox. Pandas, numpy, matplotlib, scikit-learn, and the usual suspects are pre-installed. Use it for analysis, charts, and pulling data from APIs you have access to.
"Pull this CSV from GitHub and cross-reference it against the dataset I just made. Flag mismatches."
"Hit the Airtable API at this endpoint and chart Q1 vs Q2."
Custom instructions
Tell the agent how you like to work — once. Custom instructions are applied to every session you start, so you don't have to re-explain your preferences on every brief.
Useful things to put there:
- Default report style ("decision memo, lead with the recommendation")
- Citation conventions ("primary sources only, no secondary commentary")
- Domain context ("I work in healthcare, assume FDA-aware framing")
Lives at Account → Custom Instructions. Edit anytime; new sessions pick up the change immediately.
Publishing
Publish a session or a folder to make your research public. You get a permanent URL at /p/<slug> with a versioned history; updates ship as new versions while older versions stay readable.
Each publication has:
- Title and summary — what shows up on the public page and link previews.
- License — Open, Open + Commercial, Attribution, or View Only.
- Visibility — Public (discoverable on Explore) or Unlisted (link only).
- Attribution chain — tracks who contributed and what was forked.
Right-click a session or folder in Workspace and pick Publish. To update a published work, pick Publish new version.
On any public page, click Copy to workspace to fork it into your own account. Forked sessions are marked so you can always tell which content is original vs. derivative.
API
Everything in the UI is also a REST endpoint. Start runs from your own systems, poll status, read documents, send mid-run guidance.
POST /research— start a ReportPOST /extractions— start a DatasetPOST /sessions/:id/messages— send mid-run guidanceGET /sessions/:id/status— poll progress and costGET /sessions/:id/document— read the finished ReportGET /sessions/:id/dataset— read finished Dataset rows
Full reference at /docs/api.
Pricing & budgets
Pay-as-you-go. No subscriptions. The numbers below are the minimum budget a session needs up front — not the price. You can set the cap higher; the agent works inside it, and anything it doesn't spend stays in your balance.
| Type | Model | Standard min | Deep read min |
|---|---|---|---|
| Report | Flash | $2 | $8 |
| Report | Pro | $10 | $25 |
| Dataset | Flash | $1 | $3 |
| Dataset | Pro | $5 | $15 |
A session's cost is LLM tokens plus the price of every search and scrape it ran. Pricing has the full per-call breakdown. Example reports shows what each tier actually produces.
Auto-recharge tops up your balance when it gets low. Volume bonuses kick in at $50 (10%), $100 (15%), and $250 (20%).
5 things to know
- Click any fact, see how it was found. Every claim in a Report and every cell in a Dataset opens a popover with the full tool-call trail. No more "where did this come from?".
- Message a running session. Don't wait for it to finish. The right-side chat picks up your guidance on the next cycle.
- Drag a past session into the composer. Easier than typing @. Useful for "build on what I did last week."
- Auto · Pro / Auto · Flash is showing you what it'll use. The pill text reveals the resolved model. If it's wrong for your case, override.
- Custom Instructions exists. Set them once in Account and stop re-typing your preferences.