◇ The operating system for AI agents
Jack your agent in. Instantly expert. Never forgets.
Plug in a pack and it's a senior specialist from turn one. Plug into memory and it knows every decision, lesson, and session — across tools and devices. Like Neo learning kung fu, for AI.
in plain terms
One install. It bolts onto the
coding agent you already use.
-
1
Install it
One command adds OpenSquid to Claude Code, Cursor, or Codex as an MCP server. Nothing to migrate — your agent just gains new tools.
$ npm i -g opensquid -
2
Load a pack
A pack is a folder of plain-text files — rules, skills, and a step-by-step workflow. You read it, edit it, version it in git. Not a prompt. Not a black-box plugin.
packs/fullstack-flow/… -
3
It just runs
Your agent now remembers your decisions, follows the workflow, and can't skip the checks. The discipline lives in the files — not a prompt it forgets.
● gates armed · memory on
001 — the operating system
Swap the pack. Swap the agent.
The engine never changes. The pack you load decides what the agent becomes — a full-stack
engineer, an SEO strategist, a contracts lawyer, your team's exact way of working. It's not a coding
tool. It's the machine any discipline runs on. experience = orchestrator(loaded packs).
Plug in a pack.
A pack is an app for behavior — a discipline, a domain expert, a persona, a lens. Load it and the agent works like a senior specialist from turn one. Packs are living: yours adapts to how you work, and never overwrites what it has learned.
Plug into memory.
Every decision, lesson, and session — captured, gated, durable. Your words are never evicted; nothing rots. A brand-new context wakes up already knowing everything, across tools and devices.
a pack can be anything that enhances an agent
- Codingfullstack-flow
- SEO & contentseo-aeo-expert
- Legalcontracts · IRAC
- Designbrand · visual system
- Researchsource-grounded
- Financeanalysis · modeling
- A personayour standing rules
- A lenssecurity · a11y · perf
the moat — why it can't drift
The model is a stochastic part in a deterministic system.
Determinism where correctness is binary — state, order, gates. Probability where it pays — the reasoning. A state machine makes the model's unreliability harmless, not absent: an illegal step simply doesn't exist to take.
- a total FSM — no undefined step
- zero-LLM gates + three always-on floors
- guess-free — evidence, or it's flagged
- the model reasons and writes the code
- drift is contained, never trusted
- swap the model — it's just an actor
002 — packs are gated state machines
The runtime is ours.
The flow is a pack.
Here's a real request running the coding pack. Every step has to pass a check before the next one starts — and when the agent drifts, the gate catches it and makes it fix it, instead of shipping the mistake. Scroll to watch a lap run.
- 01SCOPE
- 02PLAN
- 03AUTHOR
- 04CODE
- 05DEPLOY
SCOPE
Plan it with you.
The gate is a deterministic, zero-LLM check — it can't be talked past. A drift is blocked with a fix, the stage reruns as a fresh Ralph lap, and the commit stays blocked until the whole flow clears — drift caught at the boundary, not at the end.
anti-self-grading
The model doesn't get to grade the model.
A lesson is promoted only on external evidence — you confirm it, or it's applied enough times and the measured pass-rate improves. The promotion path makes zero LLM calls, enforced by an audit that must find none.
The model may propose a lesson. It never promotes its own.
- Mem0
- Letta
- Zep
- Auto-Memory
- Dreaming
The line no one else holds.
continuity
Start anywhere.
Continue anywhere.
One agent workspace across sessions, tools, devices, models, and compute. Local state is the trust base; cloud sync is the fabric. The agent never owns the source of truth — so nothing rots when you move.
local substrate → portability → cloud sync → device registry → capability fabric
003 — surface
Wire it to the agent
you already use.
It's just MCP at the boundary. One install, register the server, add the gates.
recall · memorize · workgraph_ready · log_phase · store_lesson