Overview

The screen the app opens on. Four tiles, a scrape bar, and the past week's failures with the fix one click away.

On this page

The tiles

Four numbers across the top, refreshed on every visit:

Tile What it counts
runs this week Every run of the past 7 days — the CLI’s and agents’ included, because everything shares one store
pages fetched this week The same window’s successfully fetched pages
LLM spend this month Metered spend since the 1st — the same number the monthly budget enforces. Flat-plan and local providers bill $0 and never move it
watches How many watches are registered

When this month’s spend has reached the budget, a Monthly LLM budget reached banner appears above the tiles: metered calls are paused until the 1st or until the budget is raised. Flat-plan and local providers keep running.

Quick scrape

Paste a URL into the bar at the top and press Enter. Magpie opens the Scrape screen with the URL filled in; nothing is fetched until you press Scrape there.

Activity

The Activity card appears once any run exists. A 14 d / 30 d switch sets the window, and the chart stacks each day’s scrape, crawl and other runs, with failed runs marked. Under it:

  • Where it went — the same window’s runs bucketed by command.
  • LLM spend by provider — one row per provider: dollars, calls, tokens.

A day with nothing in it reads as an empty No runs in the last N days card.

Failures and fixes

The Failures card lists the past 7 days’ failed runs — cancelled runs are not failures — grouped by what went wrong. Each group names the reason and shows the newest run’s verbatim error:

Group Meaning The card offers
Bot challenge The site served a bot check Retry with browser, Capture a session
Quality gate The page came back thin or walled Retry with browser
Timeout The site was too slow — often transient Retry, Open in form
Blocked by robots.txt robots.txt disallows the path Open in form (adjust the scope in Crawl)
Network Couldn’t connect Open in form, Retry
Other Read the error Open in form

Retry with browser re-runs the scrape with Render set to Browser. Retry re-runs it with the options it failed with — secrets were never kept, so a redacted key or cookie jar falls back to your saved key or captured session, and a run that used custom headers, typed text or proxy credentials refuses and points at the form instead. Capture a session jumps to Settings → Sessions. Open in form pre-fills the Scrape or Crawl form; crawls keep no per-run options, so their remedies always go through the form.

Runs started by the CLI or an agent keep no reason or URL, so they read as Other with nothing to act on.

Agent, watch and sessions

Three tiles under the cards:

  • Connect an agent — the MCP server’s state: off with a Start MCP server button, the bound port (:47219) with Copy endpoint, Copy for Claude Code, Add to Claude Desktop and Add to Cursor, or needs a plan while the trial is used. The full toggle lives in Settings → Integrations.
  • Watch pulse — watches that changed in the last 24 h, the failing count, and when the next check is due.
  • Sessions — how many logins are captured, with a warning listing expired ones and a Capture in Settings button when there are none.

Recent runs

The newest runs across every command, the same table as History — open one for its output, its cost ledger and its reproduce command. The screen remembers nothing you need to come back for: a run in flight shows its live counters on its own screen, and everything it produced lands here when it ends.

v1.0.0-rc.4 · updated