Author
Your agent inspects the page with its own browser and writes a flat parser config — CSS selectors, field types, transforms. ~20 lines of JSON, not a codebase.
"price": { "css": "span.price", "type": "number" }I use an AI agentI'm a developer
Author parsers with the agent you already use — Claude Code, Cursor, any MCP client. Test them against the live web, then run batches, crons, and multi-step crawls that land clean rows in a SQL database your agent queries directly.
No LLM per page HTTP + MCP, one key Rows land in SQL 1,000 free credits + 1 database
npx skills add Parseium/skillsSession 04 / parser: price-monitor cron 06:00
You
“Scrape these three stores into a products table. Keep it fresh nightly.”
Your agent
→ test_parser · title 20/20 · price 20/20 · sku 19/20
→ save_parser · price-monitor v2 · green
→ run_batch · 4,980 urls queued · fair queue
→ create_cron · 0 6 * * * · nightly
SELECT * FROM products ORDER BY _scraped_at DESC
4,980 rows · private database · updated today at 06:00
| sku | price | in_stock | _scraped_at |
|---|---|---|---|
| alpine-jacket-m | $118.00 | true | 06:00:12 |
| trail-runner-42 | $89.50 | true | 06:00:12 |
| summit-pack-28l | $134.00 | false | 06:00:13 |
| ridge-fleece-l | $72.25 | true | 06:00:13 |
/01
Claude Code · Cursor · Codex
/02
Full surface over MCP — 16 tools
/03
20+ prebuilt APIs
/04
A private SQL database per dataset
01 / How it works
The intelligence happens once, when your agent authors the parser. Every run after that is deterministic — fast, cheap, repeatable, debuggable.
Test
One call runs the config through the production fetch stack and returns per-field match counts and samples. The agent iterates until green — no deploy loop.
POST /v1/parsers/test → title 20/20 · price 20/20Run
One call submits 5,000 URLs. Fair queueing, retries, and rate control are our problem, not yours.
POST /v1/batches → 5,000 urls acceptedQuery
Rows upsert into your private SQL database, keyed so re-runs keep tables clean. Your agent reads results the way it already knows how.
SELECT * FROM products WHERE price < 20Then add a cron — it stays fresh every night without another prompt.
02 / The runtime
Your agent authors the intelligence once. These are the parts that run it forever — deterministic, fair, and debuggable.
01
Agent-authored parsers
HTML, JSON, API, or regex — one compact config format your agent writes and owns. Per-field diagnostics show what matched, what nulled, and what to fix.
POST /v1/parsers/test
02
Batches with fair queueing
Submit thousands of URLs in one call. Weighted fair admission means a whale next door never starves your jobs; retries and idempotency come standard.
POST /v1/batches · 5,000 urls
03
Datasets: SQL, not blobs
Every dataset is a real, private SQL database. Keyed upserts keep re-runs clean. Query with SQL, dump as CSV or NDJSON, wire it straight into your app.
SELECT * FROM products
04
Crons
Schedule any parser or workflow. Overlap protection, durable fire records, and automatic disable after repeated failures — no babysitting.
0 6 * * * · price-monitor
05
Workflows
List page → detail pages in one linear pipeline, with URL dedupe across runs so you only ever fetch what's new.
list → detail → rows
06
Prebuilt APIs
Instagram, TikTok, YouTube and 20+ more endpoints, ready on day one. Batch them like anything else. No parser needed.
GET /v1/instagram-profile
03 / Built for agents
The MCP server isn't a demo with three tools — it's the whole product: parsers, batches, databases, SQL, crons, workflows, status. Sixteen tools, the same API key as HTTP, the same shapes back. Whatever your agent can do over the API, it can do over MCP.
01
Via MCP
Add one server entry and your agent gets the full toolset — create parsers, run batches, query results — inside Claude Code, Cursor, or any MCP client.
02
Via HTTP
A clean /v1 REST surface with one error envelope. Every error carries a hint field — written for a model to act on, not a human to google.
03
Via the docs
A step-by-step agent guide teaches the whole authoring loop: discover selectors, dry-run, save, scale. Paste it into your agent's context and go.
MCP session / test_parser complete
→
test_parser price-monitor — live urls, sample of 20
00:00.4 Fetched via production stack20 urls
00:01.1 title · price matched20/20
00:01.2 sku matched17/20
00:01.3 Diagnostics returned2.1 KB
hint
3 pages nulled `sku` — selector `.sku-code` absent on bundle pages. Samples attached; add a fallback selector and re-test.
Designed for context windows
Responses are counts and samples, never dumps. Results over the cap come back truncated with a hint for paging the rest. Your agent's tokens are the scarcest resource in the loop — we treat them that way.
04 / Reliability
The runtime doesn't try to be smart. It tries to be fair, isolated, and accountable — every run, every night.
Physically isolated
Each of your databases is a separate database — isolation by architecture, not a WHERE clause we hope holds.
Nothing lost quietly
Every billable execution is logged and attributable to its job, batch, cron, or workflow run.
Strict ownership
No API key can see, run, or even confirm the existence of another user's parsers, jobs, or data.
Prebuilt APIs
Need data before you need a parser? Instagram, TikTok, YouTube and 20+ more endpoints, one credit away. Discovery is free and unauthenticated — your agent can browse the catalog before you even sign up.
Browse the catalog05 / Pricing
Credits meter live collection work—a scrape or a prebuilt call. Creating and dry-running a parser against stored HTML is free, and SQL queries are unmetered.
Start with 1,000 free credits
Every new account can create one private database and start collecting immediately. No credit card required.
Starter
For new projects and focused recurring collections.
$49/ month
Credits replenish monthly
- 150,000 monthly credits
- 3 custom parsers
- 10 fetch / 3 browser concurrency
- 3 databases · 1 GB storage
Pro
For growing data operations with multiple live jobs.
$99/ month
Credits replenish monthly
- 700,000 monthly credits
- 10 custom parsers
- 30 fetch / 10 browser concurrency
- 10 databases · 2 GB storage
Business
For high-volume collection across many sources.
$299/ month
Credits replenish monthly
- 3,000,000 monthly credits
- 30 custom parsers
- 150 fetch / 30 browser concurrency
- 30 databases · 10 GB storage
06 / FAQ
01What does “agent-native” mean here?+
Your AI agent is the operator, not a gimmick. It authors parsers, tests them, submits batches, schedules crons, and queries results — over MCP or HTTP with one API key. The dashboard is optional.
02Do you use an LLM to extract data?+
No — and that's the point. Your agent writes a deterministic parser once; every run after that is plain fetch-and-extract. Re-running 10,000 pages doesn't cost 10,000 LLM calls.
03I don't use an agent. Can I still use Parseium?+
Yes. The REST API is plain and documented, prebuilt APIs need zero setup, and parser configs are simple enough to write by hand.
04What happens when a website changes?+
Your parser's dry-run returns per-field match counts and samples, so the agent sees exactly which fields broke and repairs the config in the same loop it used to author it.
05What can I parse?+
Rendered HTML with CSS selectors, JSON embedded in pages, JSON APIs, or regex over raw text — plus the prebuilt platform endpoints.
06Where does my data go?+
Your choice per run: returned inline as JSON, or upserted into a private SQL database you query with plain SQL and dump as CSV or NDJSON.
07Will heavy users slow my jobs down?+
No. Admission to the execution pool is weighted fair round-robin across users, with per-user and global caps. Big batches queue; they don't starve neighbors.
08Do you support scheduled scraping?+
Yes — cron-scheduled parsers and workflows with overlap protection and automatic disable after repeated failures. Upserts mean nightly re-runs keep tables clean instead of duplicating rows.
09How does pricing work?+
Every new account starts with 1,000 free credits and one private database. Credits pay for live executions; dry-running a parser against a stored body is free. Paid plans start at $49 per month.
- Author and test in one agent session
- Batches, crons, and workflows on a fair queue
- Clean rows in your own SQL database