Self-hosted autonomous AI operations platform. A primary agent — you name it; it works like an autonomous chief of staff — monitors your channels, drives a real authenticated browser, delegates to a swarm of specialist peers, and runs the back office 24/7 across Telegram, WhatsApp, Discord, Slack, Email and its own SwarmAI Messenger app — all under hardware-key auth + MFA.
PS>powershell -c "irm https://northpeak.app/install.ps1 | iex"
Installs Node 22+, pnpm 10+, git, then clones the repo and links the global swarmai command.
Idempotent · honours SWARMAI_HOME / SWARMAI_REPO env vars ·
full install guide
A 28-module desktop dashboard, grouped into thematic areas. Every feature ships with tests, audit logs, and master-gated approvals where mutation is involved.
Watch the whole organisation think in real time. The orchestrator agent sits at the centre of a constellation of department peers — security, finance, strategy, legal, tech, ops — with ask / reply / assign edges animating as work flows between them. Ring, Shells, Sectors and Bubbles views.
A dedicated desktop messenger where your team and the AI share one contact list. Human↔human chats cost zero tokens; the AI contact is clearly badged “uses tokens”. Markdown, files & images, reactions, reply / edit / delete, voice notes, link unfurl and full-text search — and the agent can register & invite teammates itself.
The swarm learns from its own outcomes. A Thompson-sampling bandit ranks models on a 7-signal reward, a model registry promotes shadow→active, and built-in heads cover sentiment, triage, tool-failure prediction and anomaly detection — with a fine-tune studio for custom scorers and LoRA datasets.
In idle time the agent runs supervised “playtime” sweeps — proposing new triggers, flows and skills from what it observed — while a sealed, sandboxed self-modification engine can draft, test and (on approval) promote brand-new tools. Inert by default; you decide what ships.
Every token and dollar is captured per category, agent, provider, model and tier — charted over time with drill-down and burn-down, and capped by budgets with no-cap or hard-limit modes. Autonomy you can actually afford and audit.
The agent produces real deliverables — Word, PowerPoint, Excel and PDF with letterheads, charts and brand templates — rendered inline as artifact cards. It also emits .ipg Interactive Portable Graphics: live, self-contained HTML dashboards you can open in the IPG Viewer.
Specialise the swarm for a vertical — agriculture, security ops, sales — with installable domain packs, then extend it from the Hub marketplace of tools, channels, skills and MCP servers. Author your own tools and MCP servers right from chat.
Configure the platform by chatting. Schedule cron jobs, enroll channels, edit personas, define triggers, manage masters — no YAML hand-editing required.
Fire-and-forget peer dispatch via peer_ask_background. SQLite-backed registry, multi-channel notification fan-out (Telegram/WhatsApp/Discord/Slack DM on completion), and main-session wake-up when results land.
Native Chrome / Edge extension drives your already-authenticated browser. No Playwright, no headless puppeteer — just DOM access plus opt-in screenshots from the tab you're actually logged into.
Talk to your agent from anywhere. Telegram, WhatsApp Cloud API, WhatsApp Personal (free QR scan via Baileys), Discord, Slack, plus the dashboard. Every channel uses DM-pairing — the agent only responds to authorised contacts.
OpenRouter, Anthropic, OpenAI, Gemini, Bedrock, GitHub Copilot, Ollama (local), MiniMax, Qwen, Kimi, the Claude Code / Codex / Gemini CLI agents, and any OpenAI-compatible endpoint. The Model Tree routes per task class (heavy / average / simple) with per-provider quotas, circuit breakers and configurable failover.
Scrub the timeline of any past conversation, branch from any turn into a new session, and export to JSON or NDJSON. Diff between branches to see exactly what changed.
ed25519 hardware key (CLI sign or WebAuthn / Touch ID / Windows Hello / YubiKey). TOTP MFA with 10 single-use recovery codes. Granular scopes. Bearer tokens with silent rotation at 50% TTL.
Pause work fast: soft (drain), cancel-all (running tasks), freeze (no new work), kill (immediate halt). Triggered by slash commands or natural language — "stop everything now" works.
LLM-call retries (rate limit, context overflow, malformed JSON), process supervisor (heartbeat, circuit breaker), workspace file integrity (LEDGER seal chain), capability-gap detection — plus eight Owner-selectable advanced strategies.
Native Windows / macOS / Linux shell built on Tauri 2. System tray, auto-updater scaffold, three themes (Midnight Forge / Twilight Command / Daylight Brief). Falls back to a PWA in the browser.
One command (swarmai setup --install-daemon) installs a native service unit. Windows uses NSSM, macOS launchd (user agent), Linux systemd (user unit — no root required).
Every operation appended to LEDGER.md with a chained HMAC seal. Workspace integrity check verifies the chain on boot — any tampering is detected immediately.
Per-IP, per-token, per-endpoint limits. In-memory by default (zero-config), Redis-backed when REDIS_URL is set — for multi-instance deploys behind a load balancer.
swarmai logs tails server output like docker logs -f. Colour-coded levels, JSON-aware structure, filterable by category, follows rotation.
A primary orchestrator agent at the centre, peer agents spawned on demand, channels feeding inbound traffic, the browser extension on the side, vault and master-auth keeping it all locked down.
The installer takes care of Node, pnpm, git, and the workspace. You only do three things.
One command on your platform. Idempotent — re-run anytime to upgrade.
# Windows (PowerShell)
powershell -c "irm https://northpeak.app/install.ps1 | iex"
# macOS / Linux
curl -fsSL https://northpeak.app/install.sh | bash
Interactive walkthrough — pick LLM provider, paste API key, set master passphrase, optionally enroll channels.
swarmai setup
Boots the agent server (port 7910) and dashboard (port 18789), and opens your browser.
swarmai start
Install as a native daemon — starts at boot, no Docker required.
swarmai setup --install-daemon
swarmai daemon status
A small daemon turns any machine — phone, laptop, desktop, server — into an endpoint your primary agent can drive. Outbound WebSocket only, so devices behind NAT (carrier networks, corporate firewalls, hotel WiFi) work without port-forwarding, static IPs, or SSH tunnels.
27+ native tools: clipboard read/write, screenshot, window list/focus/close/move, app launch, volume, brightness, caffeinate, system info, battery, network interfaces, process management.
50 bundled manifests cover git, gh, docker, kubectl, helm, npm, pnpm, cargo, aws, gcloud, az, psql, redis-cli, terraform, ffmpeg, jq, curl, rg, fd, and 30+ more.
Full mobile device control with Termux:API + Shizuku — battery, wifi toggle, bluetooth, brightness, volume, screen lock, screenshot, app launch, calls, SMS compose, WhatsApp compose, GPS location, vibration.
JSON-RPC server for VS Code, Zed, JetBrains, Cursor. Agent reads open buffers, fetches diagnostics, proposes inline edits (you accept/reject), drives the integrated terminal.
Single-use 6-char pair codes (5-min TTL, ambiguous chars excluded), sha256-hashed token storage, per-device tool-policy floor, master-revocable from the gateway. Token lives at ~/.swarmai/device-agent.yaml chmod 0600.
Device opens a WebSocket OUT to your gateway URL — works behind any NAT/firewall a normal browser works behind. Exponential reconnect backoff, in-flight invocation draining on shutdown.
# Windows (PowerShell)
powershell -c "irm https://northpeak.app/device-agent.ps1 | iex"
# macOS / Linux / Termux (Android)
curl -fsSL https://northpeak.app/device-agent.sh | bash
On Android: install Termux + Termux:API from F-Droid first — the installer walks you through Shizuku pairing.
Already have SwarmAI installed locally and just want the desktop tools added in-process? Use
agent.sh /
agent.ps1 instead.
TypeScript strict, ESM-only, plugin-boundary lint enforcement — the codebase is built to be readable and modifiable.
No marketing fluff — here's the real state of the project.
Your masters file, vault, audit ledger, and session DB all live on your machine. No data leaves the box except for the LLM provider you wire in.
The platform itself is free. You only pay for the LLM calls — with your own provider key. Bring OpenRouter, Anthropic, OpenAI, Gemini, or run Ollama locally for zero LLM cost.
A 28-module dashboard, 200+ tools and 1,200+ tests passing, lint-clean across a ~95-package monorepo — but APIs may still evolve before 1.0. Run it, file issues, contribute fixes.