portable agent layer · v0

Your AI,
with a memory.

PAL is a persistent, cross-agent layer that sits inside Claude Code, opencode, Cursor and Codex — giving them identity, modes, a 5-phase algorithm, hooks, skills, and a memory that compounds across every session.

$bun add -g portable-agent-layer
01
5-phase
algorithm engine
02
16+ skills
bundled
03
4 agents
claude · opencode · …
section · layers

Watch your AI grow up.

Each scroll adds one capability to the same session. Start with bare vanilla on the left, finish with a fully-PAL'd assistant on the right.

session · interactive
PAL · offclaude code
══════ PAL ══════
🗣️ <your-agent>: Ready when you are.
user » help me ship the auth flow today
assistant » Sure — let me start with a sign-in form and routes…
1 · Observe2 · Plan3 · Execute4 · Verify5 · Learn
📋 criteria:
[x] C-1 better-auth provider wired
[x] C-2 protected route group renders user
[ ] C-3 stripe session for paid tier
🧠 memory · loaded from ~/.pal
opinion · concise (99%)opinion · prefer bun (90%)principle · read before edit
↳ hooks
SessionStart → LoadContext injected 4 frames
PreToolUse → SecurityValidator: rm -rf /tmp · blocked
Stop → captured 1 reflection, 1 opinion
🗣️ <your-agent>: On it. Pulling up better-auth + the protected route group from your stack.
00 / 05
baseline
00baseline

Vanilla AI

Drop a prompt in. The model replies. Each new session starts blank — no identity, no method, no recollection of what worked yesterday.

01+ identity

A name. A voice. A start.

PAL gives your assistant a stable identity — the name is yours to pick — and a session-start ritual. Every conversation begins where the last one left off, not from zero.

02+ modes

Right effort for the task.

Every reply is classified — minimal for greetings, native for quick work, algorithm for everything that demands structure. No more 600-word answers to a typo fix.

03+ the algorithm

Five phases. Verifiable criteria.

Complex work runs through Observe → Plan → Execute → Verify → Learn. Atomic criteria, anti-criteria, evidence required. No more rubber-stamping a half-finished task as done.

04+ memory

It remembers you.

Tracked opinions, crystallized principles, daily relationship notes, failure digests, signal trends. The assistant arrives already knowing how you like to work — and gets sharper every session.

05+ hooks

Wired into the lifecycle.

SessionStart loads context. PreToolUse blocks dangerous commands before they run. Stop captures learnings and writes the next reflection. The harness watches every event so you don't have to.

section · portability

One engine.
Any agent.

PAL lives in ~/.pal — your home directory, not the tool's. Memory, skills, hooks, telos, all portable. Switch agents on Tuesday and your assistant still knows you on Wednesday.

  1. 01

    Claude Code

    CLI · IDE

  2. 02

    opencode

    TUI

  3. 03

    Cursor

    IDE

  4. 04

    Codex

    CLI

  5. 05

    your own

    if it has hooks

[ runs anywhere bun runs ]

macOS/Linux/Windows
section · install

Give your AI a self.

One install. Your assistant picks up an identity, a method, a memory — in every agent, on every machine, for every session that follows.

$bun add -g portable-agent-layer && pal cli init
01
identity
your assistant, named
02
algorithm
5 phases, verifiable
03
memory
compounds across sessions
04
portable
~/.pal, any agent