You're in. Two MCP servers, nothing held back.

Four read tools: list_campaigns, get_campaign_metrics (impressions, clicks, cost, conversions over the last 7 days), get_search_terms (sorted by spend, so wasteful keywords surface fast), and list_negative_keywords. Ask "how did my ads do this week" and Poke just answers.

Four mutating tools: pause or enable a campaign, set a daily budget, add or remove a negative keyword. Each refuses to act unless called with confirm: true, so Poke always asks before it spends or changes anything. No budget blowout from a misread text.

Hit connect, authorize Google Ads once, and the server prints a personal /sse URL. Paste it into Poke and you're live. Each URL is its own identity, tokens stored per user, zero cross-user leakage.

A /disconnect link deletes your token and kills any live session on the spot, then hands you a fresh URL. You hold the keys to your own ad account.

Core (run_command, healthcheck), files & git (read, write, patch, list_dir, git_status), system & app control (clipboard, open/quit/focus apps, desktop notify, say, screen capture, tail logs), and background jobs (start, status, output, kill).

run_command executes in a fresh zsh -f subshell with optional cwd and timeout. A non-zero exit comes back as normal output, not a tool error, so Poke can read a failing build and reason about it instead of giving up.

Builds, installs, and dev servers run detached and return a jobId immediately. Poll status, tail output, or kill them later. No more "tool timed out" on a three-minute install.

The bearer token is stripped from every child process env, so a prompt-injected echo $TOKEN can't exfiltrate it. Every state-changing call is appended to an audit log. It listens on localhost only and reaches Poke through a cloudflared tunnel you control.

./setup.sh compiles the project and installs a LaunchAgent so the bridge auto-starts at login. Point a cloudflared tunnel at localhost:3000, paste the URL and token into Poke, done.