klura-ai/klura/klura

Klura

Klura runs maintained web-data tools locally. Start with a managed package; its signed contract, local package, browser work, and scrape output stay on the user's machine.

Default: use a managed package

1. Call search_packages for the site or capability.
2. Call show_package to inspect the selected package or capability.
3. Call install_package before use. list_installed_packages shows active immutable artifacts.
4. Use call_package_capability for one verified read, or start_scrape_run for a durable bounded collection. For a declared local browser realm, use open_package_login, wait for the user, then complete_package_login before selecting its session on a call or run (a consent realm completes by itself and its capabilities use it without a session name); clear_package_session removes only that local session. A run pins that session generation until terminal state. Use wait_scrape_run, get_scrape_run, list_scrape_runs, and list_scrape_run_items to follow it; resume_scrape_run, cancel_scrape_run, and discard_scrape_run are explicit lifecycle operations. run_consumer_doctor shows current structural scheduler contention. remove_package changes only the active local pointer.

Consumer results are typed data. Read kind, operation, code, and declared outcome fields; never infer success, absence, retryability, or a next action from response wording, a status alone, or a text-matching rule. A consumer failure does not automatically start discovery. Explain the result or ask the user whether they want an explicit authoring attempt. See klura://reference#consumer-tools.

Explicit authoring only

Use discovery tools only when the user asks to map, build, or maintain a tool, or explicitly chooses authoring after no managed package covers the job. The compact path is start_session → observed browser work → end_drive; follow the phase, checkpoint, audit, and schema feedback returned by tools. A successful session must save a complete reusable strategy, or be honestly closed with abort_session when the user stopped or the site cannot be completed. See klura://reference#reverse-engineer-playbook and klura://reference#triage.

For an explicit authoring task, preserve user literals exactly, use only observed traffic and structural verification, and keep incomplete progress in the discovery artifact rather than inventing runtime heuristics. save_strategy and end_drive return the next required action when an audit or checkpoint applies.

Boundaries

• Managed public packages are read-only. Their sole source-bearing profile is an exact signed, maintainer-reviewed browser page program; do not ask klura to execute unreviewed package code, choose arbitrary egress, or bypass browser/security controls.
• Use the host's normal confirmation policy for installation, removal, and target-site reads. A scrape file path belongs to the user; MCP paths must already be absolute.
• Do not imply that a package is hosted by klura or that a scrape is cloud-run. The V1 registry is signed static metadata; execution is local.