Native Rust browser automation CLI for Chrome/Chromium via CDP. gsd-browser keeps a persistent background daemon, auto-starts on first use, and exposes 90+ top-level commands for navigation, interaction, authenticated live viewing, annotations, recording bundles, snapshots with versioned refs, assertions, structured extraction, network control, visual diffing, tracing, and stateful auth flows.

Built for AI agents, CI pipelines, and developers who want deterministic browser control without adopting a full browser test framework.

gsd-browser mcp is now a first-class, extremely powerful browser automation platform for agents. It exposes 50+ tools, live resources (real snapshot/refs/state/timeline data), and executable prompts over local stdio or hosted Streamable HTTP (Model Context Protocol).

Completed advancements include:

• Full coverage of the rich surface: versioned refs + multiple snapshot modes, semantic act + find_best, advanced forms, robust assertions + waits, browser_batch for atomic flows, live viewer + full human collaboration (takeover, annotations, goal banners, step/abort/pause/resume, sensitive mode), first-class recording & evidence bundles, visual regression, HAR/trace/PDF export, network mocking & blocking, device emulation, encrypted auth vault + state save/restore, structured extraction, prompt injection scanning, action cache for long-term self-healing, multi-tab/frame management, rich diagnostics (debug_bundle etc.), and more.
• Resources that actually query the daemon for live context (gsd-browser://latest-snapshot, current-state, active-recordings, timeline, etc.).
• Executable prompts encoding best-practice multi-step workflows (robust_login_flow, full_page_audit, autonomous_research_task, evidence_creation_workflow, debug_stuck_agent_flow, etc.).
• Standardized high-value response envelopes on every tool call: summary, structured_data, suggested_next_actions, evidence_refs.
• Seamless reuse of the proven daemon client (auto-start, named sessions for isolation + persistent cache/state, robust error handling).

This is designed to be the high-end browser backend for serious agent platforms.

Get started in seconds:

gsd-browser mcp

Point Cursor, Claude Desktop, VS Code + Copilot, or any MCP client at it.

Host it for remote/cloud clients:

export GSD_BROWSER_MCP_AUTH_TOKEN="$(openssl rand -hex 32)"
gsd-browser mcp --http --host 0.0.0.0 --port 8788

Expose /mcp over HTTPS and send Authorization: Bearer <token> from the remote MCP client.

To use OpenGSD console tokens and usage tallying:

export GSD_BROWSER_MCP_AUTH_VERIFY_URL="https://mcp.opengsd.dev/api/mcp/tokens/verify"
gsd-browser mcp --http --host 0.0.0.0 --port 8788

Each tools/call request is validated against the console, counted against the user's free quota, and rejected when the console returns a throttle response.

Tailored setup + config snippets:

./scripts/mcp-quickstart.sh cursor # claude | vscode | generic

Key documentation:

• docs/mcp.md — Full capabilities, architecture, client configs, quickstart script.
• docs/AGENT-BEST-PRACTICES.md — Golden rules, workflow patterns, "When to Use What" table, self-healing, response envelopes, prompt/resource usage (essential reading for agents).
• docs/examples/mcp-client-config.json — Ready-to-paste example.
• Root SKILL.md and the gsd-browser-skill/ pack — Complete underlying command semantics and curated workflows (the MCP tools are a direct mapping).

Run gsd-browser mcp and unleash one of the most powerful browser surfaces available for agentic work.

npm install -g @opengsd/gsd-browser

Download from GitHub Releases:

git clone https://github.com/open-gsd/gsd-browser.git
cd gsd-browser
cargo install --path cli

Install the CLI and register the Codex Plugin in one pass:

curl -fsSL https://raw.githubusercontent.com/open-gsd/gsd-browser/main/install.sh | bash -s -- --codex-plugin

The installer writes the plugin to ~/plugins/gsd-browser, updates the personal Codex marketplace at ~/.agents/plugins/marketplace.json, and runs codex plugin add gsd-browser@<marketplace> when the Codex CLI is available. Without --codex-plugin, the interactive installer offers OpenAI Codex Plugin alongside the agent skill options.

The crates.io package (gsd-browser) is not published yet. Use GitHub release assets or a source build.

The one-line installer (curl -fsSL https://raw.githubusercontent.com/open-gsd/gsd-browser/main/install.sh | bash) also sets up the gsd-browser-skill/ pack for coding agents and documents the MCP path in its header.

The daemon starts automatically on first use.

# Navigate to a page
gsd-browser navigate https://example.com
# Snapshot interactive elements and assign refs like @v1:e1
gsd-browser snapshot
# On example.com the only interactive element is the "More information..." link
gsd-browser click-ref @v1:e1
# Wait for navigation and assert the result
gsd-browser wait-for --condition network_idle
gsd-browser assert --checks '[{"kind":"url_contains","text":"iana.org"}]'
# Capture a PNG
gsd-browser screenshot --output page.png --format png

For the modern agent experience, prefer the MCP server (see top of this README).

gsd-browser view starts an authenticated localhost workbench for the active session. The URL is bound to the session, viewer id, loopback origin, expiry, and viewer capabilities. Use view --print-only when another tool needs the URL.

gsd-browser view
gsd-browser view --print-only
gsd-browser control-state
gsd-browser takeover
gsd-browser release-control
gsd-browser sensitive-on
gsd-browser sensitive-off

The viewer streams the real Chrome page, forwards pointer, wheel, keyboard, text, and paste input while in Control mode, creates annotations in Annotate mode, and starts/stops local recording bundles in Record mode. Sensitive mode keeps local human control available while cloud frame capture and evidence surfaces use redaction policy.

Annotations and recordings stay local to the daemon state directory:

gsd-browser annotations
gsd-browser annotation-get <id>
gsd-browser annotation-clear <id>
gsd-browser annotation-resolve <id>
gsd-browser annotation-export --output annotations.json
gsd-browser record-start --name checkout-bug
gsd-browser record-stop
gsd-browser recordings
gsd-browser recording-get <id>
gsd-browser recording-export <id> --output <path>
gsd-browser recording-discard <id>
gsd-browser recording-validate <id-or-path> --json

(MCP equivalents: browser_view, browser_annotation_request, browser_record_start, etc. — among the highest-leverage features for collaborative agent work.)

gsd-browser currently exposes 90+ top-level commands (the MCP server exposes the most valuable subset as 50+ discoverable tools with agent-optimized descriptions and envelopes):

• Persistent daemon with automatic startup for fast repeated commands
• Durable named sessions with explicit health reporting and no silent session replacement
• Versioned refs from snapshot for deterministic interaction (@v1:e1, @v2:e3)
• Explicit assertions with assert and multi-step automation with batch
• Shared inspection semantics across snapshot, find, wait-for, assert, and ref-driven actions
• Semantic find-best and act flows covering 15 built-in intents
• Named sessions via --session for isolated parallel browser workers
• Authenticated local viewer with human takeover, pause/step/abort, annotations, sensitive mode, and bounded recording bundles
• Structured JSON output on every command via --json
• Visual diffing, HAR export, PDF generation, and CDP tracing in the same tool
• Saved browser state plus encrypted credential replay through the auth vault
• Prompt injection scanning for agent-facing browsing workflows
• Action cache for self-healing intent mappings across sessions (especially powerful with named MCP sessions)
• Full MCP server with resources, prompts, and agent-optimized envelopes

gsd-browser merges configuration in this order:

1. Built-in defaults
2. User config: ~/.gsd-browser/config.toml
3. Project config: ./gsd-browser.toml
4. Environment variables: GSD_BROWSER_*
5. CLI flags

Example gsd-browser.toml:

[browser]
path = "/usr/bin/chromium"
cdp_url = "http://localhost:9222" # attach to existing Chrome instead of launching
headless = true
[daemon]
port = 9222
host = "127.0.0.1"
[screenshot]
quality = 90
format = "png"
full_page = false
[settle]
timeout_ms = 500
poll_ms = 40
quiet_window_ms = 100
[logs]
max_buffer_size = 1000
[artifacts]
dir = "./browser-artifacts"
[timeline]
enabled = true
max_entries = 500

Supported environment variable overrides use GSD_BROWSER_<SECTION>_<FIELD> naming:

export GSD_BROWSER_BROWSER_PATH=/usr/bin/chromium
export GSD_BROWSER_BROWSER_CDP_URL=http://localhost:9222
export GSD_BROWSER_BROWSER_HEADLESS=true
export GSD_BROWSER_DAEMON_PORT=9333
export GSD_BROWSER_SCREENSHOT_QUALITY=90
export GSD_BROWSER_SETTLE_TIMEOUT_MS=1000
export GSD_BROWSER_ARTIFACTS_DIR=./browser-artifacts
export GSD_BROWSER_VAULT_KEY=your-encryption-key

For MCP usage, place the relevant GSD_BROWSER_* variables in your MCP client's server env configuration.

gsd-browser defaults to the stable chromiumoxide CDP client for maximum compatibility.

gsd-browser --stealth navigate https://bot.sannysoft.com
# or
gsd-browser --backend stealth navigate ...
# config.toml
[browser]
stealth = true
backend = "chaser-oxide" # or "stealth", "chromey"

Effects when enabled:

• Anti-detection Chrome flags (--disable-blink-features=AutomationControlled, IsolateOrigins, etc.)
• Realistic UA + hardware (cores, memory, platform) spoofing
• CDP signal patches (webdriver, cdc_ markers, chrome object, permissions, WebGL)
• Client Hints and locale/language consistency
• (Future) human-like mouse curves via input_dispatch when chaser-oxide backend active

The following require explicit cargo features (the published binary always ships the stable default):

• chromiumoxide-backend (default) — current stable
• chromey-backend — fresher CDP definitions, adblock, fingerprint crate (drop-in, same use chromiumoxide)
• chaser-backend / stealth feature — protocol-level stealth, ChaserPage human input (compile with --features stealth)
• ferrous-backend — ergonomic locator/wait API (launch path experimental)

To build with an alternative:

cargo install --path cli --no-default-features --features chromey-backend
# or for full stealth
cargo install --path cli --no-default-features --features stealth

See also the audit and superpowers plans for the "dependency/stealth refresh" item.

Trade-off: stealth backends may lag the main chromiumoxide feature surface or have different perf characteristics. The daemon handlers/refs/viewer remain unchanged regardless of backend.

• The CLI parses commands and sends them to a local daemon over a loopback HTTP channel.
• The daemon maintains the browser lifecycle, page/frame routing, network hooks, action timeline, and session manifest state.
• --session <name> creates isolated daemon and browser instances for parallel workflows.
• The MCP stdio server (gsd-browser mcp) is a thin, high-fidelity adapter over the exact same daemon client used by the CLI.

Recommended 2026+ path: Connect via the MCP server (gsd-browser mcp). It gives you automatic discovery of 50+ tools, resources, and prompts with rich envelopes and best-practice guidance. See the dedicated sections at the top of this README, plus docs/mcp.md and especially docs/AGENT-BEST-PRACTICES.md.

When using the CLI directly (or for reference):

• The daemon auto-starts. You almost never need gsd-browser daemon start.
• gsd-browser daemon health reports the current session state and does not auto-start the daemon.
• Use --json when you need structured output.
• Prefer snapshot then click-ref or fill-ref for stable interaction, and re-snapshot after page changes. (MCP: read the latest-snapshot resource.)
• Use assert and batch when you need deterministic pass/fail automation.
• find-best and act cover 15 built-in semantic intents for common navigation, form, dialog, auth, and pagination actions.
• The live viewer + annotations + recordings + human takeover are first-class superpowers for collaborative or auditable work.
• Read SKILL.md for the full command reference and workflow patterns (this is the source of truth for MCP tool semantics).
• Install the curated gsd-browser-skill/ pack (via the main installer) for coding agents.

Licensed under either of:

• MIT License
• Apache License, Version 2.0

at your option.