For AI agents · MCP server
Give your agents a memory — and a phone line to you.
Every agent you run is trapped in its window: it forgets between sessions, and it can’t reach you once you’ve walked away from the screen. Squirrel Brain is the missing half — one persistent brain your whole fleet shares, plus a real line to your phone: a push, a spoken message, or an actual call it places itself. Wire it in over MCP in about three lines.
{
"mcpServers": {
"squirrel-brain": {
"url": "https://geczbtsjfbvfukdzdemr.supabase.co/functions/v1/mcp-server",
"headers": {
"X-API-Key": "sb_your_personal_key",
"Authorization": "Bearer <supabase-anon>",
"X-Agent-Name": "your-agent"
}
}
}
}That’s the whole integration. Exact values + every tool’s parameters live in the developer docs.
Works with Claude Code, Claude Desktop, Cursor & self-hosted agents (OpenClaw, LangChain, your own scripts).
ChatGPT can’t connect custom MCP servers yet. And if a human never touches any of this — the app does everything on its own.
The two things agents can't do alone
Remember everything. Reach you anywhere.
One brain your whole fleet shares
Not a vector store you bolt on and babysit. It's the same brain the human uses — so what your coding agent files, your research agent (and the person) can find later. Memory survives the session, the tool, and the model. Semantic search, a daily brief, boards, and a forever-note that never gets wiped.
A line to the human's actual phone
Every other agent is trapped behind a screen the person has to be looking at. Squirrel Brain is the last mile: when something can't wait, your agent places a real phone call and speaks — or fires an escalating alarm, or a push. It reaches the human in the physical world, on its own, unattended.
What it unlocks
What your agent does with hands and a voice.
A coding agent finishes a 40-minute migration at 2 a.m., runs the tests green, and — instead of waiting in a terminal nobody's watching — calls your phone: “Migration's done, tests passed, staged for review.”
A research agent digs up that the visa window you asked about closes Friday. It doesn't bury it in a chat log — it schedules a call for 8 a.m. with the summary, so the deadline actually reaches you.
An assistant agent remembers, across every session and every tool you run, that your daughter's recital is Thursday — because it lives in one shared brain, not that agent's ephemeral context window.
Any agent can schedule its own follow-up: “re-check the deploy in four hours.” It sets the alarm on itself, then acts when it fires — a real feedback loop, not a one-shot request.
The API
32 tools, live today.
The complete set your agent can call — names match the server’s tools/list exactly.
Every tool, with parameters, is in the developer docs.
Built to be trusted
Handing an agent your phone, safely.
Scoped, revocable keys
Each agent gets a personal API key, scoped to one brain and killable in a tap. No shared god-token.
Every write is attributed
Nothing anonymous ever lands in your brain — each item is stamped with the exact agent that created it, and you can wipe one agent's writes in one call.
The server enforces honesty
An agent can't report a success it didn't achieve — the portal verifies the real effect before it confirms. No hallucinated “done.”
Spend is capped
Per-agent and fleet-wide cost ceilings are enforced server-side, so a runaway loop can't run up your bill or your phone.
How it works
Three steps. No glue code.
Mint a key in the app
Open Squirrel Brain and generate a personal API key. It's yours — scoped to your brain, revocable anytime.
Point your agent at the portal
Add the MCP server to Claude Code, Claude Desktop, or Cursor — or wire it into your own agent. The 32 tools appear automatically. No glue code.
It acts on your brain
Now your agent can remember, set reminders, read your day, file things, and reach you on your terms — every action traceable to the agent that made it.
Agent FAQ
Agent questions, answered.
What does Squirrel Brain give an agent that a vector DB doesn't?
Two things. First, it's not just recall — it acts: set a real iPhone alarm, add a calendar event, file a photo, append to a forever-note. Second, it's the delivery layer: an agent can place an actual phone call to the human and speak, so time-sensitive results don't die in a chat log. And the memory is shared with the person and every other agent, not siloed to one tool.
Can my agent schedule its own callbacks?
Yes. An agent can create an alarm or a cell_alert for a future time and act when it fires — a real feedback loop. It can re-check a deploy in four hours, follow up on a lead next Tuesday, or ring you the morning a deadline lands.
Is it safe to give an agent a line to my phone?
It's built for exactly that. Every agent uses its own scoped, revocable key; every write is attributed to the agent that made it (and wipeable per-agent); the server enforces honesty so an agent can't claim an action it didn't perform; and per-agent + fleet spend caps stop a runaway loop cold.
Which AI tools can connect?
Anything that speaks MCP and can send an API-key header: Claude Code, Claude Desktop, Cursor, and custom agents you host yourself — OpenClaw, LangChain, or plain scripts. Consumer chat apps like ChatGPT can't connect custom MCP tool servers yet.
Can Claude or ChatGPT set an alarm on my phone on their own?
Not by themselves — chat assistants can't fire a real, time-based alarm on your iPhone. Squirrel Brain closes that gap: the app turns anything you say, snap, or forward into real alarms, and MCP-capable agents can drive it directly through the portal.
Give your AI a memory that lives on your phone — and can ring it.
Pre-beta — join the launch list and developer invites go out first.
Join the launch list