One skill, every agent

Brain has always been model-agnostic — one ~/.brain/ shared across Claude Code, Codex, OpenCode, Antigravity. But "supported" meant we'd written an installer for that specific agent: a table of paths, a prompt file, a command format, per tool. Every new agent was a new row to maintain, and an agent we hadn't gotten to was an agent you couldn't use Brain with.

That framing is now backwards. The question isn't "did Brain's team add support for my agent?" It's "does my agent read SKILL.md?" — and increasingly, they all do.

The standard that made this possible

Over the last few months the open Agent Skills format — a folder with a SKILL.md (YAML frontmatter plus instructions) and optional scripts/ — went from a Claude convention to something read identically by Claude Code, Copilot, Cursor, Codex, Gemini CLI, and a growing list of others. Two of them (Codex, Copilot) even read the same cross-tool ~/.agents/skills/ location, so one folder installed there serves multiple agents at once.

A memory system is a near-perfect fit for that shape. Brain's intelligence lives in a deterministic CLI; the skill just has to tell the host when to recall, how to classify a write, and how to read the results. That's instructions plus a binary — exactly what a portable skill carries.

What ships

A single committed folder, included in the npm package, so npm i -g brain-memory drops it where cross-agent registries can find it:

brain-memory/
  SKILL.md              # frontmatter + the full recall / memorize / verify loop
  scripts/
    session-start.mjs   # dependency-free wrapper over `brain session-start`

The frontmatter declares what it needs, in the open format's own vocabulary:

metadata: {"requires": {"bins": ["brain"]}, "install": [{"kind": "node", "package": "brain-memory", "bins": ["brain"]}]}

Any host that reads the standard sees "this skill needs the brain binary; here's how to install it" and offers the same deterministic recall, provenance-aware memorize, and verification workflow every other integration uses. Because the skill only ever depends on the global CLI — never on a specific host's APIs — it stays byte-identical everywhere it runs.

The trust surface comes with it

This isn't a stripped-down port. The portable skill documents the whole thing: honor low_trust and quarantine_pending flags, copy receipt lines verbatim, route untrusted writes through brain verify. Wherever the skill lands, the poisoning defenses travel with it — a memory absorbed from a web page on one agent is still capped, flagged, and quarantined when a different agent recalls it.

The full mechanics are in Portable Skill Distribution. If you're on an agent Brain doesn't have a dedicated installer for, this is your path — and it probably already works.