Squirrel Brain agent skills
Download the complete loop: a useful phone briefing, your spoken answer, delivery back to the right task, and proof that the decision moved the work forward.
Free for Squirrel Brain users
Ready to install
The downloads are free. Phone delivery runs through your own Squirrel Brain account and private MCP key. Start with Contact Your Human; the Supervisor bundle includes it too.
Core skill
A secure, reusable human-in-the-loop skill for any agent. It calls with useful context and options, captures the spoken reply, routes it, and verifies the result.
Codex
unzip -o contact-human.zip -d "${CODEX_HOME:-$HOME/.codex}/skills"Claude Code
unzip -o contact-human.zip -d "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills"Fleet skill
A dedicated watchdog for multiple work sessions. It detects stalls, verifies real artifacts, nudges the right task, and calls only after a verified question has waited at least five minutes.
Codex
unzip -o supervisor-loop.zip -d "${CODEX_HOME:-$HOME/.codex}/skills"Claude Code
unzip -o supervisor-loop.zip -d "${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills"One package, honest adapters
The safety rules, state engine, and Squirrel Brain phone loop are shared. Each agent host supplies its own session list, messaging, and recurring wake-up.
Turnkey task discovery, bounded task reads, messaging, and heartbeat automation.
The same skill installs directly. Claude must prove and wire its own team/session adapter and persistent timer before claiming unattended supervision is active.
Supported when the host can list work, read a bounded tail, message the owner, keep private state, and run a recurring coordinator.
No dead ends
The agent recognizes a real decision, repeated failure, or unsafe impasse — not just a keyword.
It names the task, describes verified state, gives numbered options, and recommends one.
Squirrel Brain rings the human's iPhone and speaks the complete decision brief.
The skill polls the private MCP Portal for the new spoken answer.
The answer goes back to the owning task, and artifact progress proves it was applied.
Built to share safely
No embedded secrets. API keys live in the user’s MCP configuration, environment, or operating-system Keychain—never in the downloaded files.
No personal defaults. The package contains no person’s name, account id, Portal thread, device id, or private filesystem path.
No false delivery claims. A call counts only when Squirrel Brain returns a real delivery receipt. Replies are matched by an invisible call identifier, so the human can simply say “option one” and unrelated Portal messages are rejected.
Revocable access. Each Squirrel Brain key is scoped to its owner and can be rotated from the app. Read the MCP setup and security docs.