ov

OverView — point it at a URL and it drives the web app like a user, records what the app does (pages, DOM, screenshots, network, console, journey) and what it's made of (HTML, JS, frameworks, API surface), then analyzes both for UX and software architecture and writes Markdown reports plus a single synopsis.

import ov

run = ov.observe("https://example.com")  # capture (zero config)
print(run.fingerprint)  # detected technologies
print(len(run.artifacts), "artifacts captured")

ov observe https://example.com                  # same, from the shell
ov check                                         # what system deps are present

ov is a tool + skill library with a deterministic, model-free core. The near-term "agent" that orchestrates it is a host agent (Claude Code) following the skills in .claude/skills/ — ov supplies the hands (drive the target) and eyes (observe + analyze); the host supplies the control loop. Everything in the core runs with no model and no host, so it's testable, cheap, and repeatable.

Install

pip install -e .            # the package
playwright install chromium # the browser it drives (required for capture)

The arch/API reconstruction and evidence layers have optional extras (pip install -e ".[arch,evidence]") and an optional Node sidecar; run ov check (or ov.check_requirements()) to see what's present and how to install what's missing.

What it captures

Capture is organized into probes — independent, registerable units that each write typed artifacts into a content-addressed store:

• Behavioral: navigation, network (request/response + size-capped bodies), dom (serialized DOM + ARIA snapshot + full AX tree), screenshot, console, perf, storage, websocket, sse.
• Static: fingerprint (framework/library detection), assets (inventory).

The five operate primitives — observe / act / journal / progress / snapshot_state — are the deterministic, LLM-free "hands" the host (or a scripted driver) uses to drive a journey; the journey trace is the UX evidence.

The facade

The CLI mirrors these: ov observe|analyze|diff|report|synopsis|overview|check|runs|mcp.

In mode="review" (your own target), ov.overview(url, mode="review") inserts a diff step: it compares the run against the latest prior run of the same target and reports what is new, changed, or resolved — drift detection for a downstream creation/modification agent. Artifacts are content-addressed, so this own-target diffing is cheap by design. The first review run has no baseline and is a no-op; pass baseline=<run_id> to pin a specific prior run.

In-package agents (optional — ov[agents])

The host-agent path above is the default and the pit of success. When you need a self-contained runtime — a non-Claude-Code host, programmatic use, or process isolation — ov.agents re-hosts the same skills + deterministic core as runnable agents. It is the spec's "mechanical lift" (Phase 4): nothing is rewritten, the lift is done by coact (COMPLETE a skill into an agent definition, REALIZE it onto a backend), and ov.agents adds only the bits coact deliberately leaves out — the operator's live-browser loop, the analyst's multimodal evidence-bundle call, and the orchestration topology (via aw).

pip install -e ".[agents]"          # coact[sdk,mcp] + aw + py2mcp + fastmcp

from ov import agents

# 1. Cheapest: materialize ov's agents as .claude/agents/ so Claude Code runs them
#    (host backend — zero LLM, no fan-out). The host stays the manager.
agents.materialize()

# 2. The cost gate before standing up an in-process fleet (~15× the tokens):
print(agents.estimate(["ux-analyst", "arch-analyst"]).render())

# 3. In-package study (LLM dependency-injected; omit it for a deterministic run):
result = agents.study("https://example.com", llm=my_model)  # -> synopsis + run

Each agent reuses Phase 1–3 machinery unchanged and routes by the §10 cost gate (operator → Haiku, UX-analyst → Sonnet, Arch-analyst / orchestrator → Opus). The analyst applies the cite-or-abstain contract — every LLM claim cites a mark/fact id or is downgraded to undetermined. For a foreign MCP host, ov mcp (or agents.mcp_server()) exposes study_url / capture_url via py2mcp.

Authorization & privacy

ov analyzes publicly served frontend material and observable network behavior of targets you are authorized to inspect. Foreign-target runs require an explicit authorized=True acknowledgement; secret/PII capture is off by default and storage values are redacted. It does not defeat access controls.

Status

Built in phases (see the GitHub issues). Phases 1–3 are shipped: the capture spine + operate primitives, the deterministic UX + architecture analysis + reports + synopsis, and the host-agent skill layer with the grounded evidence bundle. Phase 4 adds the optional in-package agent layer (ov[agents], above) and review mode — own-target diff / regression detection (ov.diff, ov.overview(url, mode="review")). Its remaining depth items (source-map / GraphQL / Lighthouse probes) are deferred to follow-up issues.

Author: Thor Whalen.