Hurricane Sentinel

Plain English

What Sentinel does

Sentinel is a local platform for running teams of AI agents under real enforcement and human approval. A head agent delegates work across persistent specialist agents — each running as its own Leash-leashed OS user with model calls inspected by Muzzle, inside a per-agent filesystem sandbox — through durable runs that pause for your approval before anything risky. Drive it from a web console or the CLI, give each agent its own linked memory, add your own Python tools, package a whole team to hand to someone else, expose any agent as an HTTP endpoint, and mix local and hosted models with every call filtered on the way out.

• Enforcement is real, not advisory: every agent runs as a dedicated Leash-leashed OS user with model traffic inspected by Muzzle, confined to its own filesystem sandbox — and the single invariant is that no path bypasses that leashed worker + approval gate. The whole platform is security-audited and hardened.
• Teams, not just one agent: a head agent delegates to teams of specialists; each team's leader breaks the task down, fans it out to members, and reports results back — with a shared team workspace, bounded fan-out, and hard per-agent isolation.
• Durable, approval-gated runs: a flagged action pauses, waits for your yes/no, then resumes exactly where it left off — in the CLI or a web console where the whole delegation tree streams live and colour-coded.
• Persistent agents with their own soul, skills, tools, model, and a linked “vault” memory you can browse as a graph — plus programmable HTTP endpoints that turn any agent, team, or the head into a callable API (sync or async, with structured output).
• Bring your own tools: drop in a Python file, pick which functions become tools, and Sentinel serves them to agents through a private tool host — approval-gated, sandboxed, with their dependencies installed in one click. Never imported by the control plane; scanned, never run, until an agent calls one and you approve.
• Portable teams: export an agent or a whole team as a secret-free package — skills, tools, model choices, souls, guidelines, and your uploaded tools — and import it into another deployment. Give a teammate the same working team; they just add their own keys.
• Any model, filtered on egress: local Ollama models or hosted OpenAI / Anthropic / Gemini, added with a key from the console — keys live only in Muzzle, and every call is routed through it so secrets and PII are scrubbed before they leave. Vision models can take image and file input right in the chat.
• Mounted workspaces: give an agent files from outside its own sandbox — a local folder or an SSH target — as a named, synced workspace, attachable to an agent, a team, or a single chat. Because runs are heavily sandboxed (no kernel mounts), Sentinel materialises the source as a real directory the run reads inside its allow-list; read-only by default (opt in to writable), the SSH key kept write-only and the host key pinned, with changes pushed back to the source only when you say so.
• Built for real work: tool-use scaffolding, forgiving tool-call parsing, self-correcting validation, context compaction, cron schedules, MCP tool servers, saved multi-step workflows, mounted workspaces, and an in-console editor for your Muzzle and Leash policy — all from a mobile-friendly web console you can sign a phone into by scanning a QR from your desktop.