cli-anything-safari
A command-line interface for Safari browser automation on macOS. Wraps the
safari-mcp Node.js MCP
server in a Python Click CLI.
Feature parity is guaranteed. Every Click command is generated
automatically from safari-mcp's tool schema (bundled as
resources/tools.json). All 84 tools are reachable with the exact
argument names and types the MCP server expects.
When to use this CLI
Each CLI invocation spawns a fresh subprocess, so there is per-call
overhead. If your agent speaks MCP natively (Claude Code, Cursor, Cline,
etc.), using safari-mcp directly over MCP stdio will be faster.
Use this CLI when:
- Your agent framework does not speak MCP (Codex CLI, GitHub Copilot CLI, custom scripts, older agent frameworks).
- You need to script browser automation from bash —
cli-anything-safari --json tool snapshot | jq '...'. - You run in CI/CD and want cron-able, subprocess-friendly output.
- You're debugging interactively from Terminal.
Installation
Prerequisites
- macOS — Safari MCP is macOS-only.
- Safari — already installed on macOS.
- Node.js 18+ —
brew install nodeor from https://nodejs.org/ - Python 3.10+
- Enable Apple Events for Safari: Safari → Develop → Allow JavaScript from Apple Events
Install the CLI
The first tool call will download the safari-mcp npm package (one-time, a few MB).
Command Structure
The CLI has 5 top-level commands:
Usage Examples
Discover the tool surface
Call a tool (schema-driven)
Navigate and read in one round-trip
Form fill (bulk)
safari_fill_form takes an array of {selector, value} objects.
Pass it as a JSON string:
Run cli-anything-safari tools describe safari_fill_form to see the
exact schema, including any new fields safari-mcp adds upstream.
Network monitoring
Storage
Raw JSON escape hatch
When you need to pass a complex nested object or want to drive the CLI from a pre-built JSON blob:
Interactive REPL
The REPL banner prints the absolute path to this SKILL.md so agents can self-discover capabilities.
JSON Output
All commands support --json as a global flag:
State Management
The CLI maintains a small amount of in-memory state for REPL display only:
last_url— last URL the CLI navigated to (updated after every successfultool navigate,tool navigate-and-read, ortool new-tab)current_tab_index— last known active tab index
There is no persistent session, no undo/redo, no document model.
Every CLI invocation starts with fresh state. Safari MCP itself is
stateless per-call: each tool command spawns a fresh
npx safari-mcp subprocess, performs the action, and exits. This is a
deliberate design choice; see HARNESS.md and TEST.md for the
reasoning behind the deviation from the standard undo/redo pattern.
Output Formats
All commands support dual output modes:
- Human-readable (default): indented key-value text for
dictresults, bullet lists for arrays, plain text otherwise - Machine-readable (
--jsonflag): structured JSON for agent consumption
For AI Agents
When using this CLI programmatically:
- Always use
--jsonflag for parseable output. - Check return codes — 0 for success, non-zero for errors (URL validation failures, MCP call failures, invalid JSON args).
- Parse stderr for error messages; use stdout for data.
- File-handling tools have inconsistent path arg names — always
check
tools describe <name>first:tool save-pdf --path /tmp/x.pdftool upload-file --selector ... --file-path /tmp/x.txt(note:--file-path, not--path)tool export-storage— no path arg; pipe JSON output to a filetool import-storage --path /tmp/x.jsontool screenshot/screenshot-element— return base64 in the JSON response, no path arg (decode it yourself)
- Snapshot before click — refs from
tool snapshotexpire on the next snapshot. Always snapshot → find ref → click in close succession. - Discover tools via
tools list— the bundled registry is the source of truth for what's available. Do not hard-code tool names that may change upstream. - Use
tools describe <name>to learn the exact schema (required args, enum choices, JSON-typed args) before constructing a call. Never assume parameter names from the description — for example,safari_evaluatetakes--script(not--code) even though the description says "JavaScript code to execute".
Agent-Specific Guidance
Finding the right tool
Use the introspection commands. The CLI is guaranteed to reflect the MCP server 1:1:
Tool selection strategy
tool snapshotovertool screenshot— structured text with ref IDs is orders of magnitude cheaper and carries the refs needed for clicks.tool click --refovertool click --selector— refs are stable within a single snapshot, selectors may be brittle.tool navigate-and-readovernavigate+read-page— saves one round-trip.tool click-and-readoverclick+read-page— saves one round-trip.tool native-clickonly when regular click fails with 405/403 (WAF blocks, G2, Cloudflare) — it physically moves the cursor.
Refs Expire
Refs from tool snapshot expire when you take a new snapshot:
- First snapshot: refs
0_1,0_2,0_3... - Second snapshot: refs
1_1,1_2,1_3...
Always snapshot → click in close succession. If in doubt, snapshot again.
Tab Ownership Safety
Safari MCP tracks tab ownership per session. Tools that modify a tab
(navigate, click, fill) are blocked on tabs the session did not open.
To operate on a specific page, always start with tool new-tab --url ....
Error Handling
Common errors:
npx not found→ install Node.js 18+safari-mcp package not found on npm registry→ check networkNot macOS→ harness is macOS-onlyAppleScript denied→ enable "Allow JavaScript from Apple Events" in Safari → DevelopBlocked URL scheme: file→ URL validation rejected the input (by design)
URL Validation
The CLI validates URLs before passing them to safari_navigate,
safari_navigate_and_read, and safari_new_tab. Blocked schemes:
file, javascript, data, vbscript, about, chrome, safari,
webkit, x-apple, and other browser-internal schemes. The raw
command also enforces this for navigation tools.
Multi-Session Warning
Safari MCP enforces a single active session by killing stale Node.js
processes older than 10 seconds. If you run two CLI instances at once,
one will kill the other's backend. There is currently no daemon
mode — for latency-sensitive workflows, drive the CLI from a
long-lived Python script that imports
cli_anything.safari.utils.safari_backend.call() directly to avoid
re-spawning the subprocess on every invocation.
Links
Security Considerations
URL Validation
All navigation tools (tool navigate, tool navigate-and-read, tool new-tab, and raw safari_navigate*) pass the url argument through
utils/security.py which blocks dangerous schemes and optionally blocks
private networks (set CLI_ANYTHING_SAFARI_BLOCK_PRIVATE=1).
Tab Isolation
Safari MCP enforces per-session tab ownership upstream — tools cannot operate on tabs the session did not open.
Profile Isolation
Set SAFARI_PROFILE env var to use a separate Safari profile for
automation:
This keeps cookies/logins/history separate from the user's main browsing.
JavaScript Execution
tool evaluate and tool run-script run arbitrary JavaScript in the page
context. Treat untrusted input with the same care as any dynamic code
execution path.
Clipboard
tool clipboard-read and tool clipboard-write touch the system
clipboard. Be careful when running inside a user's active session —
overwriting the clipboard mid-task is disruptive.
Regenerating the tool registry
If you upgrade safari-mcp, regenerate the bundled schema:
The parity test (test_parity.py) pins the expected tool count; update
it when the upstream tool list changes.
More Information
- Full documentation:
cli_anything/safari/README.mdin the package - Test coverage:
cli_anything/safari/tests/TEST.mdin the package - Architecture analysis:
safari/agent-harness/SAFARI.md - Methodology:
cli-anything-plugin/HARNESS.md - MCP backend pattern:
cli-anything-plugin/guides/mcp-backend.md
Version
1.0.0 — targets safari-mcp 2.7.8 (84 tools). Bundled tool registry is
regenerated via scripts/extract_tools.py when safari-mcp upgrades.
