Give your AI agents the whole web, for a fraction of the tokens.
Vessel is a browser for AI agents and the web layer around it that reads less: it opens any page, PDF or document, keeps only what answers, remembers it, and shows you the tokens it saved on every single call. Measured on live sites: 94.6% fewer tokens per question, 22.7× cheaper whole tasks, 61 agents that cost zero model tokens.
What is in Vessel, in plain words
Pick a product and watch what your agent would do. Every run ends with a receipt: the tokens it read and the tokens it avoided. Use one product, or all six behind the same key, the same ChatGPT or Claude connector and the same dashboard.
Vessel Browser
A real browser your agent steers. It sees a tidy list of the page, not the whole page. Pay, send and delete wait for a person. Every step is recorded.
Vessel Search
Ask a question. It searches the live web, opens the best pages and returns only the sentences that answer, with the source next to each.
Vessel Read
Any page, PDF, scan, Word, Excel, slide deck or Google Doc, reduced to the passages that answer your question. Personal data hidden on request.
Vessel Agents
61 ready-made agents for jobs, prices, reviews, filings, papers and pricing pages, or a goal in plain English with the result shape you want back.
Vessel Monitor
Point any agent at a page, a search or a feed. Pick a schedule. Get what changed, a trend and a weekly report in Slack or email.
Vessel Memory
Everything your agents read, saved once and searchable later with sources. The next question costs a few hundred tokens.
Across all six: a receipt on every call (tokens kept and avoided, per team and per tool), connectors for ChatGPT, Claude, Cursor and Codex, and a signed identity websites can verify.
Access the web behind the login.
Your agent signs in with your own logins, including one-time codes, does the work on the live site, and brings back a result your software can use. It records every step, and anything that pays, sends or deletes waits for a person.
Verify it’s you
Enter the 6-digit code from your authenticator app.
{{totp:acme}} · the seed never leaves the server| Invoice | Date | Amount | Status | |
|---|---|---|---|---|
| INV-2291 | Oct 3 | $4,120.00 | Paid | Download PDF |
| INV-2297 | Oct 9 | $980.00 | Due | Download PDF |
| INV-2304 | Oct 14 | $2,375.00 | Due | Download PDF |
How do I start? Three steps, no code.
Most people are done in ten minutes. Developers get an API and SDKs on top.
Create a workspace
Free to start. Your pages, documents and results stay in your workspace; teams that need it get a private deployment.
Connect what you already use
In ChatGPT or Claude, add Vessel as a connector in Settings. In Cursor, Codex or Claude Code, paste one line. For your own agents, use the API or SDK.
Ask, in your own words
"Read this PDF and tell me the refund policy." "Watch Greenhouse for new PM roles." "Compare these three pricing pages." The answer comes back with the tokens it saved.
Where does the token bill come from, and what changes?
Hover or tap a row to see the before and after. Numbers are measured on live sites with the real tokenizer.
Tools paste the whole page into the chat. A 30-page article costs 27,611 tokens to find two sentences.
Every step re-reads the page as HTML or a full tree; retries after a bot wall burn more.
Dropped into a chat, the whole file is parsed and billed, every time you ask.
People and agents re-read the same job boards, price lists and filings daily.
How is Vessel different from TinyFish, Firecrawl, Browserbase and Apify?
All four are good at what they sell. Vessel is the one built around spending fewer tokens and proving it, with monitoring and memory included. Facts from their public pages and docs, 13 October 2026.
| Capability | TinyFish | Firecrawl | Browserbase | Apify | Vessel |
|---|---|---|---|---|---|
| Token receipt on every call (kept vs avoided) | not offered | not offered | not offered | not offered | yes, per call and per team |
| Return only the passages that answer a question | Fetch returns clean content | Question / Highlights formats (model-assisted) | Fetch returns content | not offered | yes, deterministic, with sentence mode |
| Goal in, structured JSON out; batches | yes, 100 per batch | agent endpoint | via Stagehand | via actors | yes, 200 per batch, your model key |
| Ready-made agents without a model key | agent needs a model | extract needs a model | not offered | marketplace, priced per actor | 61 included |
| Documents: PDF with OCR, Office, Google Docs, YouTube captions | pages via Fetch | PDF and Office (parse) | not offered | per actor | yes |
| Monitoring with trends and weekly reports | page and topic monitors | monitor (alerts) | not offered | schedules | any agent, trends, 7-day reports |
| Memory: ask what was already read | not offered | hosted indexes | not offered | datasets | your own collections |
| Approvals before pay / send / delete | not offered | not offered | not offered | not offered | yes, server-side |
| Connector for ChatGPT and claude.ai | plugins, MCP | plugin, MCP | MCP | MCP | OAuth connector + savings per connector |
| Signed agent identity (Web Bot Auth) | not published | not published | yes (Cloudflare cohort) | not published | yes, your own key |
| Private deployment for enterprises | VPC (enterprise) | partial self-host (AGPL) | enterprise plans | not offered | yes (enterprise plan) |
| Pricing | search/fetch free; $0.016 per agent step; $0.002 per browser minute; prepaid wallet | credits per page | per browser-hour + proxies | per actor run + platform usage | flat plans, no per-page credits, no token markup |
How much does it actually save?
Every number below is from the real tokenizer on live sites, reproducible with the scripts in the repository. Your own numbers are on the Token savings page, per surface: API, MCP, dashboard, each connector.
Sources: docs/COMPARISON.md (Oct 2026), scripts/task-tokens.mjs, scripts/tokens.mjs.
How is Vessel different from Firecrawl, Browserbase, Apify and TinyFish?
They are good products. Vessel differs in three ways: it is built to cut the token bill and proves it on every answer, it bundles browser, documents, agents and memory behind one key with no model gateway, and the agent identifies itself instead of pretending to be a person.
| Capability | Firecrawl | Browserbase | Apify | Vessel |
|---|---|---|---|---|
| Private deployment for enterprises | partial | no | no | yes (enterprise plan) |
| Question-focused reading with measured receipts | claims (model-written answers) | no | no | passages + exact token counts |
| Live browser sessions with approvals and rewind | sandbox | yes | actors | yes |
| Ready-made extractors without a model key | no | no | marketplace (paid actors) | 61 built in, free |
| Documents: PDF with OCR, Office, Google Docs, YouTube | PDF/Office | no | per actor | yes |
| Knowledge base you can inspect | hosted indexes | no | datasets | yes, on your disk |
| Signed agent identity (Web Bot Auth) | no | yes (cloud) | no | yes, your own key |
| Connectors for ChatGPT and claude.ai with per-connector savings | plugin | no | no | OAuth 2.1, savings per connector |
| Pricing | credits per page | per browser-hour + proxies | per actor run | flat plans; no per-page credits, no token markup |
Which problems does it solve, by role?
Teams running LLM apps
Every web or document call your product makes goes through one layer that returns passages, not pages. Savings are metered per surface so finance sees the line move.
Cut the bill →Agent builders
MCP, OpenAI/Anthropic/Gemini tool schemas, TypeScript and Python SDKs, a skill file the model reads first, token budgets per call.
Integrations →Sales and marketing
Lead enrichment from a domain list, competitor pricing watches, brand mentions anywhere, SEO + AIO reports on every page you ship.
Try the free report →Recruiting
Greenhouse, Lever, Ashby and job boards as watches: one CSV of new roles each morning, zero model tokens.
Recruiting use cases →Finance, legal, research
SEC filings, Federal Register, court opinions, arXiv and PubMed; 200-page PDFs answered from the right pages; PII redacted before a model sees it.
Analyst use cases →Operations and support
Supplier portals with logins become routines; approvals gate anything that pays; the help-centre crawl answers tickets from passages.
Routines and approvals →What does it cost?
Free to start. Flat monthly plans with no per-page credits, no browser-hour meter and no markup on your model's tokens. Private deployment for enterprises.
Questions people ask
What is a browser for AI agents?
A service that runs real browsers and returns what a model can act on: compact ranked snapshots instead of HTML, the passages that answer a question instead of whole pages, rows instead of browsing. Vessel adds documents, agents and memory on the same key; the model you use is your choice.
How much does it cut LLM token costs?
On live sites, question-focused reading sends 5–15% of a page to the model; whole browsing tasks measured 22.7× cheaper than raw HTML; scheduled agents cost zero model tokens. Your own numbers are on the Token savings page, per surface, counted with the real tokenizer.
Can we run it on our own infrastructure?
Yes, under the enterprise plan: the same product deployed privately in your cloud or on-premises, with SSO, audit exports and support. Most teams start on the hosted plans.
Does it bypass CAPTCHAs or bot protection?
No. It detects challenges and stops so no tokens are wasted on retries, lets a person take over, and can sign its requests with your own key (Web Bot Auth) so supporting sites can recognise your agent. Sites that need a residential proxy are marked honestly in the extractor library.
How are the token numbers measured?
With the same tokenizer the models use, on live sites, per observation and per whole task. The scripts are in the repository and the dashboard shows your own numbers per surface: API, MCP, dashboard, built-in agents and each ChatGPT or Claude connector.
Which models and frameworks work with it?
Any. ChatGPT and claude.ai through an OAuth connector; Claude Code, Cursor, Codex and Gemini CLI through MCP; OpenAI, Anthropic and Gemini SDKs through tool schemas; LangChain, Vercel AI SDK, n8n and Zapier through REST. Agents use your model key or a workspace key; there is no model gateway and no markup on tokens.
Can it read PDFs, scanned documents, Google Docs and YouTube?
Yes: PDF text layers and OCR for scans, DOCX, XLSX, PPTX, ODT, RTF, images, Google Docs, Sheets and Slides share links, and YouTube caption tracks. Audio and video files are not transcribed because there is no speech model on the server.
Spend tokens on answers, not on pages.
Create a workspace in a minute. Connect ChatGPT, Claude, Cursor or your own agents. Every answer comes with a token receipt.