oassis — web scraping, crawling and browser control for AI agents

Turn any web page or document into markdown, map and crawl whole sites, and drive a real browser. Pay per call with a wallet: no API key, no signup, no monthly plan.

oassis.dev · MCP: https://api.oassis.dev/mcp · OpenAPI · llms.txt

Try it now, no account (MCP)

Over MCP every tool works without a key or an account, except web_search_exa, which is paid. Send your own client User-Agent: requests carrying the default of curl or an HTTP library are refused.

List the 11 tools:

curl -X POST https://api.oassis.dev/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -H 'user-agent: MyAgent/1.0' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Map oassis.dev (free over MCP):

curl -X POST https://api.oassis.dev/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -H 'user-agent: MyAgent/1.0' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"web_map","arguments":{"url":"https://oassis.dev"}}}'

Real response (trimmed):

{"urls":["https://oassis.dev/","https://oassis.dev/web/v1/scrape","https://oassis.dev/llms.txt",
"https://oassis.dev/openapi.json","https://oassis.dev/privacy","https://oassis.dev/terms",
"https://oassis.dev/support"],"discovered":7,"offered":14,"sources":{"sitemap":7,"page":0}}

Read a page (web_scrape) with the same client:

curl -X POST https://api.oassis.dev/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -H 'user-agent: MyAgent/1.0' \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"web_scrape","arguments":{"url":"https://oassis.dev","formats":["markdown"]}}}'

Search is the one paid tool over MCP. Calling it without a key explains the price and how to pay:

web_search_exa is paid, not free: each search costs $0.007, which is the price Exa charges, passed through.
Two ways to carry on: pay per call with a wallet (x402, USDC on Base), or use an API key.

Related MCP server: pagewatch

Pay per call (x402)

Without payment, the request answers 402 with the exact price of that call. Nothing is charged without a payment.

Step 1: ask for the page. Real response for oassis.dev with markdown and controls:

curl -i -X POST https://api.oassis.dev/web/v1/scrape \
  -H 'content-type: application/json' \
  -d '{"url":"https://oassis.dev","formats":["markdown","controls"]}'

HTTP/2 402
payment-required: eyJ4NDAyVmVyc2lvbiI6MiwiZXJyb3IiOiJQYXltZW50IHJlcXVpcmVkIi…

The payment-required header (base64) states: 2000 micro-USDC = $0.002, network eip155:8453 (Base mainnet), asset USDC 0x8335…2913, recipient payTo 0x23Fc…Dfc75d.

Step 2: sign the payment with your wallet. Use an x402 SDK for your language: the client signs an EIP-3009 USDC authorization and the facilitator settles it. The wallet only needs USDC on Base; the facilitator pays the gas.

Step 3: repeat the same request with the payment proof in the PAYMENT-SIGNATURE header. The response comes back with PAYMENT-RESPONSE.

MCP includes a one-time 100-credit trial for scrape and map. Other tools and further use require a Bearer API key with account credits. HTTP accepts the same API key or x402.

The 11 operations

Over HTTP each one is paid. Each one is served at api.oassis.dev/web/v1/… and at web.oassis.dev/v1/…. Both hosts count and charge the same.

MCP and HTTP share credit prices. One credit = $0.002; verified accounts receive 1,000 welcome credits once. Any GET on a route returns, for free, the fields it expects and what it costs.

Pricing

Minimum charge per call: $0.002. The price is quoted before the work and charged after; if the work does not happen, you are not charged. Balances do not expire and there is no monthly fee.

Crawls and batches are charged up front for the pages or urls they may read. A refund for work that never ran is a credit to an API-key account. A wallet payment (x402) is settled on-chain and cannot be reversed from here, so with x402 you pay for every url or the limit you ask for: map the site first.

Example: markdown + controls of one page = $0.002. A 10-page crawl with markdown = $0.02, because each page charges its outputs plus its links.

Free over MCP

Paying per call over HTTP (x402) or with an API key reads the same content on your own responsibility, with no MCP restrictions.

MCP

Streamable HTTP at https://api.oassis.dev/mcp. Published in the official MCP registry as dev.oassis/scraper-crawler, and under the organization github.com/oassisdev.

initialize and tools/list are always free.

Agent tasks

Ask a question in plain words and get a structured answer with its sources. The agent plans its own steps: it decides what to search, which pages to read, and when it has enough.

Tasks require an API key with credits over MCP/HTTP, reserving a maximum and charging actual execution cost. x402 can independently pay a fixed task package.

curl -X POST https://api.oassis.dev/tasks/v1/research \
  -H 'authorization: Bearer oas_…' \
  -H 'content-type: application/json' \
  -H 'idempotency-key: compare-hosts-1' \
  -d '{"task":"Compare pricing and limits of three hosting providers for AI agents",
       "mode":"research","urls":[],
       "limits":{"max_cost_usd":0.10,"max_duration_seconds":180,"max_steps":10}}'

The answer is 202 Accepted with a Location header. Read it until it is final:

curl https://api.oassis.dev/tasks/v1/research/TASK_ID -H 'authorization: Bearer oas_…'

Statuses: queued, running, completed, partial, failed, cancelled. A task is completed only when it has findings that cite pages it really read, nothing is left open and no limit was hit. Otherwise it is partial, with the reason and what is missing. A task with no page read at all is failed.

Cancel with POST /tasks/v1/research/TASK_ID/cancel. Cancelling a finished task changes nothing.

Budget. max_cost_usd is reserved from the balance before the task starts. What the task spends is charged at the end, and the rest of the reserve is returned in the same transaction. A request that would exceed the balance is refused with 402 and nothing is reserved.

Prices. Each client price is at least twice what one call costs us in the worst case (a 100% markup, which is a 50% gross margin). The start fee is charged once when a task is accepted:

Every attempted model call and page read is charged, including the ones that fail: the model or the browser ran. A call that is refused before it runs (budget exhausted) is not attempted and not charged. Each model call is capped (planning at 256 output tokens, the answer at 1,500) and must return within 30 seconds for planning and 60 seconds for the answer; a call that runs past its limit counts as attempted, and is charged.

The minimum budget is $0.03: it pays the start fee, one planning call, one page read and one answer.

Idempotency. Idempotency-Key is per account. The same key with the same request returns the same task. The same key with a different request is refused with 409. Keys are kept with their task. MCP has no idempotency key: retry tasks over HTTP.

Uncertain operations. A payment to Exa that may have gone through, but whose outcome is unknown (for example after a crash), is charged at the most it could have cost. It is never retried, and the task ends as partial, or failed when no page had been read, and says so.

Sources and evidence. Each source has an id, the URL, the time it was read, and the fragment of the page the answer may use. A page is cut at 4,000 characters and the cut is marked. The model sees at most 12,000 characters of evidence in one call. Page content is treated as data: it cannot change the task, call tools or authorise spending. A finding that cites a source the task did not read is removed. The task reports how many were removed as findings_removed. The server checks the citations, not the wording: a finding can cite a real page and still add a conclusion the page does not state. Read the cited evidence before you rely on a finding.

MCP. web_task_start, web_task_status and web_task_cancel run the same code as the HTTP endpoints.

Limits of version 1. Tasks research public pages only. They do not submit forms, buy, book or send anything. Crawls, browser sessions and actions are not part of tasks yet.

Network safety. Urls are checked before they are read: only http and https, no credentials, no obvious local or private hosts or addresses, and no IPv6 literals. This is a filter, not a full protection against server-side request forgery: a public name can resolve to a private address, and redirects are followed. Do not rely on it to keep an agent away from your internal network.

Model. Planning and answers run on Workers AI. The default model is @cf/meta/llama-3.3-70b-instruct-fp8-fast; set AGENT_MODEL to change it. The rates live in src/agent/rates.ts.

Run your own copy

npm install
npm run check      # typecheck, dead-export scan, tests
npm run deploy     # Cloudflare Workers

Needs a D1 database, a KV namespace and the Browser Run binding; see wrangler.jsonc. Payments need X402_PAY_TO and, to settle on Base mainnet, CDP keys.

License

Elastic License 2.0. You may read, run, modify and host it for yourself; you may not offer it to third parties as a hosted or managed service.

Account credits

1 basic single-format page = 1 credit = $0.002. Buy 2,500 / 10,000 / 25,000 credits for $5 / $20 / $50, once or optionally monthly. Purchased credits accumulate without scheduled expiry. MCP and HTTP use the same Authorization: Bearer oas_… key and balance. Other operations consume their quoted dollar price divided by $0.002; fractions are supported.

Generate or revoke keys at https://oassis.dev/dashboard/billing. Verified accounts receive 1,000 welcome credits once. MCP without a key has a lifetime 100-credit scrape/map trial, associated with a salted Cloudflare IP hash (shared networks share the trial). When exhausted, the tool requests signup and an API key. Status/cancellation and discovery are free.

See billing operations for migrations, Stripe activation and alert delivery.

Definitive API families

The same paths are used on every API host. There are no old-path aliases or redirects.

• Web Tools: /web/v1/scrape, /map, /search, /session, /act, /feedback. Session close: DELETE /web/v1/session/:id.

• Extraction Tasks: POST /tasks/v1/scrape/batch and POST /tasks/v1/crawl; status: GET on the creation path plus /:id; cancellation: POST on that status path plus /cancel.

• Research Tasks: POST /tasks/v1/research; GET /tasks/v1/research/:id; POST /tasks/v1/research/:id/cancel.

• Agent catalog: GET /agents/v1 and GET /agents/v1/:agent. Create a service: POST /agents/v1/:agent.

• Owned services: GET /agents/v1/services (latest 100), GET /agents/v1/services/:id, POST /agents/v1/services/:id/cancel. Cancellation is idempotent, retains the report and stops future runs. A read already in progress may finish.

Clients must replace /v1/*, /web/v1/crawl*, /web/v1/scrape/batch* and /agent/v1/tasks*. These old URLs are no longer mounted and return 404. Existing job/service identifiers remain valid on the definitive paths. API keys, MCP tool names and billing rules do not change.