omniscout 0.2.6

OmniScout CLI: local-first browser automation, semantic search, and research for AI agents

OmniScout

Local-first browser automation, semantic search, and research for AI agents.

No cloud APIs. No hosted browser sessions. No MCP yet. No SDK.

The CLI is the interface. Install the omniscout command and drive everything from the terminal or via JSON (--json / OMNISCOUT_JSON=1).

scout is a short alias. harness is a legacy dev alias kept for compatibility.

---

Install

Requires Python 3.11+ and Google Chrome (macOS default: /Applications/Google Chrome.app).

pip install omniscout
omniscout install                    # verify Chrome + prefetch embedding model (~80MB)
omniscout install --skill            # optional: drop agent skill into Claude/Cursor/Codex/Antigravity

If Chrome is not installed, add --bundled to download Playwright Chromium (~190MB).

Search commands auto-start the local daemon and keep the embedding model loaded in RAM across invocations — no manual warm-up step required.

---

Features

Browser automation (daemon-backed)

Long-lived daemon at 127.0.0.1:7720 for sub-second per-action latency.

• Playwright backend (default) — local Chrome with persistent profiles
• Chrome extension backend (opt-in) — drives your real running Chrome via chrome.debugger; same JSON vocabulary, real cookies and logins
• Atomic actions: navigate, snapshot, click, fill, scroll, key, hover, screenshot, pdf, eval, wait, tabs, network, upload, login, captcha, close
• Stable @eN refs from the accessibility tree (preferred over CSS selectors)
• Persistent profiles — log in once, stay logged in
• CAPTCHA: local-first manual handoff; optional 2captcha / capsolver solvers
• Network capture (CDP) with URL filters and per-request inspection
• Session restore across daemon restarts

Semantic search

• DuckDuckGo HTML search with optional local embedding rerank
• Sources: ddg, index (local crawl corpus), memory (remembered visits), hybrid (memory + DDG)
• omniscout answer — grounded one-sentence answers: direct DDG answers first (snippets, Search Assist), then extractive parsing, local LLM, and limited crawl (auto, fast, balanced, deep; extractive fallback)

Warm embedding model

Search, research, and memory commands route embeddings through the daemon. The sentence-transformers model (all-MiniLM-L6-v2) loads once (~2s) and stays hot. omniscout daemon status reports embed_model_loaded.

Content extraction

Fetch URLs to clean Markdown, plain text, or structured JSON via trafilatura + markdownify. On-disk HTML cache.

Research pipeline

Multi-step: search → crawl → extract → embed → rerank → summarize.

Browser memory

Remember visits and notes; semantic search over your browsing history.

• omniscout remember <url> — visit, extract, index
• omniscout memory list|show|note|delete|stats|clear

Workflow shortcuts

Top-level commands for agent ergonomics:

• omniscout open <url|index> — open URL or latest search result
• omniscout snapshot, omniscout context, omniscout reset
• omniscout workflow export — JSON steps from workflow state + action history

Replay & observability

Every daemon action is logged to $OMNISCOUT_DATA_DIR/daemon/actions.jsonl:

• omniscout daemon trace — recent activity table or JSON
• omniscout daemon replay <action_id> — re-run a single action
• omniscout daemon watch — live SSE event stream
• Top-level omniscout replay action-<id> and omniscout replay session-<name>

Benchmarks

• omniscout benchmark answers — latency + correctness matrix over answer modes
• omniscout benchmark startup — CLI process launch overhead

---

Quickstart

# Search
omniscout search "local-first browser agents"
omniscout answer "who is the president" --depth balanced

# Extract
omniscout extract https://example.com

# Browser (daemon auto-starts)
omniscout browser navigate https://example.com
omniscout browser snapshot --refs-only
omniscout browser click '@e1'
omniscout browser screenshot --out /tmp/state.png
omniscout browser close --all

# Research
omniscout research "state of local AI agents in 2026"

# Profiles & sessions
omniscout profile create work
omniscout browser open https://news.ycombinator.com --profile work --headful
omniscout session start --headful

Optional warm-up before a batch of searches:

omniscout warmup

---

JSON output (for agents)

Every command supports --json. Set OMNISCOUT_JSON=1 to make JSON the default for an entire shell session. Logs go to stderr; stdout is the structured result.

export OMNISCOUT_JSON=1
omniscout search "robotics simulators" --limit 5
omniscout browser navigate https://example.com --session demo

Direct HTTP (no CLI wrapper):

curl -s -X POST http://127.0.0.1:7720/command \
  -H 'Content-Type: application/json' \
  -d '{"action":"navigate","args":{"url":"https://example.com"},"session":"demo"}'

---

Architecture

omniscout CLI ──HTTP POST /command──▶ omniscout daemon (127.0.0.1:7720)
     │                                      ├─ Playwright backend
     │                                      ├─ Extension backend (opt-in)
     │                                      └─ Embed service (warm model)
     └── Search / Extract / Research engines (local Qdrant + DDG)

Python package layout (for contributors):

cli/omniscout/
  app.py              # Typer root (binary: omniscout)
  commands/           # CLI sub-commands
  daemon/             # HTTP server, backends, replay, events
  engines/            # browser, search, research, extractor, crawler
  store/              # SQLite cache, sessions, workflow, memory
  models.py           # pydantic JSON contract

---

On-disk state

Path	Purpose
profiles/	Persistent Chrome user-data-dirs
qdrant/	Embedded vector index
models/sentence-transformers/	Prefetched embedding model
memory.sqlite	Browser memory (visits + notes)
sessions.sqlite	Long-lived browser session registry
cache/pages/	Content-hashed HTML cache
daemon/	PID, port, logs, action history, session restore

Default locations:

• macOS — ~/Library/Application Support/omniscout/
• Linux — ~/.local/share/omniscout/

Override with OMNISCOUT_DATA_DIR, OMNISCOUT_CONFIG_DIR, OMNISCOUT_CACHE_DIR. Legacy HARNESS_* names are still accepted.

---

Configuration

config.toml (in config dir):

default_source = "ddg"
search_limit = 10
research_results = 8
request_throttle_seconds = 1.0
embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
embedding_local_only = true
browser_channel = "chrome"
summary_sentences = 6

Environment variables

Variable	Purpose
OMNISCOUT_JSON=1	Force JSON output on every command
OMNISCOUT_EMBED_DAEMON=1	Route embeds through daemon (default on)
OMNISCOUT_DAEMON_AUTO_START=0	Don't auto-start daemon
OMNISCOUT_DAEMON_PORT	Daemon port (default 7720)
OMNISCOUT_DATA_DIR	Override data directory
OMNISCOUT_EMBED_LOCAL_ONLY=0	Allow runtime Hugging Face fetches
TWOCAPTCHA_API_KEY	CAPTCHA solver API key

Legacy HARNESS_* equivalents work for all of the above.

---

Why local Chrome?

Using system Chrome gives you real cookies, login state, extensions, and the same fingerprint as daily browsing — without a separate ~190MB Chromium download. Falls back to Playwright Chromium automatically when Chrome is unavailable.

---

License

Modified MIT — see LICENSE. Products built on OmniScout must prominently display Powered by OmniScout on the user interface.