🐕 leash

See what your AI agents did last night.

One command. No signup. Costs, runaway loops, dead crons, failing workflows: your whole agent fleet, across every machine and cloud platform it runs on.

npx -y getleash

Live demo fleet → · Quick start · Dashboard · Watchdog · Privacy · FAQ

Why

You run agents now. Claude Code sessions that work for hours, launchd and cron jobs, GitHub Actions on a schedule, Vercel crons. They ship while you sleep. They also die silently, loop on the same tool call at 3am, and burn API budget while you sleep. Nothing watches them.

leash is the missing pane of glass: what your agents did, what it was worth, what broke, and the exact command to fix it. Local-first, metrics-only, one npx away.

Quick start

npx -y getleash

Ten seconds later you have:

• The value of the work: pay-as-you-go API value of your last 30 days, per project, with the multiple vs your flat subscription ("9.8× a $200/mo Max")
• Every scheduled agent on the machine (launchd, cron, systemd): zombies pointing at deleted scripts, jobs that stopped writing logs, last runs that crashed
• Every cloud agent, checked live: GitHub Actions and Vercel crons verified with the logins already on your machine (see platforms)
• Runaway loops: the same tool call repeated 10+ times with identical input, priced
• Fixes: each problem comes with a copy-paste command, not a shrug
• Press Enter at the end and the whole thing becomes a live dashboard (optional, free)

Nothing leaves the machine unless you say so. --json for scripts, --days N for the window, --offline to skip every network call.

The fleet dashboard

→ Open the live demo fleet (three machines, real layout, curated data)

npx getleash connect gives you a private URL. No account, no email: the URL is the key. Every machine you connect with the same fleet token lands on the same page, and a SessionEnd hook refreshes it after each Claude Code session.

• Hero: what your agents did, in dollars, with the subscription ROI multiple
• 🌙 While you slept: a midnight-to-7am timeline per night, each session placed at the hour it actually ran
• The pulse: 30 days of cost, day by day
• Your fleet: one health dot per agent, problems first
• Needs attention: severity, machine, cause, and where to get the fix

Live platform checks

Agents don't just live on your laptop. leash checks the platforms where they run, with credentials that never leave your machine and never go to leash:

| Platform | Setup | What gets checked | |---|---|---| | GitHub Actions | none: uses your gh login or GITHUB_TOKEN | last run status, disabled workflows, schedules GitHub silently stopped firing | | Vercel | none: uses your vercel login or VERCEL_TOKEN | every cron across all projects and teams, enabled/disabled, failed deployments | | Render | getleash link render <api-key> | cron jobs: suspended, stale, last successful run | | Railway | getleash link railway <token> | scheduled services, latest deployment status | | Cloudflare Workers | getleash link cloudflare <api-token> | workers with cron triggers |

npx getleash link shows what's connected. Tokens are validated before saving, stored in ~/.leash/providers.json (chmod 600), and only ever sent to their own platform's API, read-only.

The budget guard

Claude Code has no native spend limit. leash gives you a hard one, in one command:

npx getleash guard --daily 25 --hourly 5

A PreToolUse hook estimates your spend from local transcripts (cached 2 min) and blocks tool calls past the cap, with a clear message. The hourly cap is the loop killer: it stops a runaway session in about two minutes, long before the daily cap would. Fail-open by design: if anything breaks, Claude Code works normally. --status to check, --off to remove, backup of your settings kept at settings.json.pre-leash.

The live monitor

npx getleash watch

top for your agents: sessions running right now, cost ticking per session, current tool call, burn rate in $/h. Refreshes every 2 seconds.

🐕 Watchdog

A dashboard has to be looked at. The watchdog barks.

Every push is diffed against the machine's previous one. The moment something new breaks (a cron goes zombie, a workflow starts failing, a loop fires at 3am) it pings your Discord channel with what broke and a link to the fleet. Known problems stay quiet: no alert fatigue.

npx getleash watchdog --discord <your-webhook-url>

(Discord → channel settings → Integrations → Webhooks → New Webhook. It proves the webhook with a test ping before saving anything.)

Pricing

| | | |---|---| | Everything you can see: report, live monitor, budget guard, fleet dashboard, night replay, platform checks | Free, forever | | 🐕 Watchdog: the alerts that come to you | Free during the beta, then $15/mo. Never charged without explicitly opting in. |

The philosophy: the visibility is free, the vigilance is paid.

How it works

flowchart LR
    subgraph machine["Your machine (local, default)"]
        T["~/.claude transcripts<br/>(costs, sessions, loops)"] --> CLI
        S["launchd · cron · systemd"] --> CLI
        R["local repos<br/>(workflow & cron definitions)"] --> CLI
        CLI["getleash CLI"]
    end
    subgraph platforms["Your platforms (your tokens, read-only)"]
        GH["GitHub API"]
        V["Vercel API"]
        O["Render · Railway · Cloudflare"]
    end
    CLI <-->|"health checks"| GH
    CLI <-->|"health checks"| V
    CLI <-->|"health checks"| O
    CLI -->|"opt-in: metrics-only snapshot"| CLOUD["leash cloud<br/>dashboard + watchdog"]
    CLOUD -->|"barks on NEW problems"| D["Discord"]

• Cost accuracy: usage is deduplicated by message.id + requestId (retries and resumed sessions would otherwise double-count), priced per model generation, within ~5% of ccusage.
• Loop detection: same tool + identical input hash, 10+ times overall or 6+ within ten minutes.
• Cloud health: workflow state plus last run from the platform's own API; "stale" is inferred from the cron expression (GitHub silently stops scheduling inactive repos).
• The gate: the budget guard is a PreToolUse hook; exit 2 blocks the call. It cannot be bypassed by permission modes.

Privacy

The local scan makes zero network calls (--offline guarantees it even with platforms linked). Everything else is opt-in and metrics-only:

| Leaves the machine (only after you connect) | Never leaves, ever | |---|---| | Costs, session counts, tool-call counts | Prompts and transcript content | | Project display names, agent labels | File paths and file contents | | Health statuses and schedules | Platform tokens (they go to their own platform only) |

An end-to-end test asserts the snapshot contains no home paths and no absolute paths. The dashboard URL is a capability: treat it like a secret, connect --off kills it locally and removes the hook.

FAQ

Yes. Open the Terminal app (macOS: Cmd+Space, type "Terminal"), paste npx -y getleash, press Enter. The report explains itself, and every problem comes with the exact command to copy and paste. Nothing is installed permanently.

The cost and session analysis reads Claude Code's local transcripts today. The scheduled-agent scan (launchd, cron, systemd) and the platform checks (GitHub, Vercel, Render, Railway, Cloudflare) watch any agent, whatever wrote it. Multi-LLM cost parsing is on the roadmap.

Within about 5% of ccusage. Usage is deduplicated by (message.id, requestId) so streaming, retries and resumed sessions are counted once, and priced per model generation, including cache-write tiers. On a subscription, the number is what your usage would cost pay-as-you-go: your ROI, not a bill.

They see metrics: project names, costs, agent health. Never prompts, paths or contents. Rotate by disconnecting (connect --off) and reconnecting for a fresh token.

leash on npm is an abandoned MongoDB package from 2012. We asked; the squatters were faster. npx getleash it is.

Development

cd cli && npm install && npm test     # build + 21 tests, all offline

TypeScript, esbuild single-file bundle, zero runtime dependencies. The cloud is three Vercel functions and a static page over Vercel Blob; its alert engine has its own integration harness (cloud/test/watchdog-harness.mjs). See CONTRIBUTING.md.

License

MIT. Built in the open, fast: see the CHANGELOG for the day-by-day story.