// Layer 02 · Agent · Harness
An agent that remembers. In a folder you own.
Install it as a plugin for Claude Code or Codex. It captures every session, compiles what it learned into a brain, and loads that brain back before you type a word. Everything it knows lives as plain markdown on your disk.
▸ /plugin marketplace add github:AgentLayer1/agentlayer-agent-marketplace▸ /plugin install agent-kevin@agentlayer
Inside a claude session launched from the folder. Relaunch, then /agent-kevin:init.
35+
skills
50+
tools
4
hooks
1
folder
The folder
- AGENTS.md
- SOUL.md · IDENTITY.md · USER.md
- knowledge/
- index.md
- memory/index.md
- raw/sessions/2026-09-11.md
- projects/
- TASKS.md
- checkout/tasks/co-014-retry-webhooks.md
- dashboard.html
$ cd ~/Agents/Ace && claude
▸ session start · Ace
- loadedwho it isSOUL · IDENTITY · AGENTS.md
- loadedwho you areUSER.md
- loadedthe brainknowledge/index · memory/index
- loadedthe boardprojects/TASKS.md · 14 open
- loadedthe lessons12 corrections, compiled
- loadedthe traillast session · 3 commits since
context primed · 6/6
> pick up where we left off on the webhook retries
Yesterday you settled on exponential backoff capped at five attempts
(co-014, thread 09-10). The branch is two commits ahead of dev.
Next: the idempotency key on the handler, then the test that replays
the failing batch. Starting with the handler.
2 files changed · running tests
// One launch. Six things loaded before the first prompt. Nobody pasted any context.
Every session is captured on exit. Captured sessions compile into a wiki. The wiki loads before you type a word in the next session. That loop is the whole product; the skills and tools exist to feed it and to act on what it knows.
// What it is
Not a chat wrapper. An operating system for one agent.
01
It remembers.
Every session is captured when it ends and compiled into a wiki. The wiki loads before your first prompt in the next one. No "as I mentioned last time", no pasting context.
▸ capture → compile → load, every session
02
Everything is a file.
Its character, its memory, your projects, your tasks: plain markdown in one folder on your disk. Browse it in Obsidian, commit it to git, grep it, take it with you if you ever leave.
▸ no database, no lock-in
03
Runs where you already work.
A plugin for Claude Code or Codex. One folder works from either. No new app, no new account, no metered API key: your session does the thinking on the subscription you already pay for.
▸ claude · codex · one home
04
Projects, not chats.
One markdown file per task with frontmatter, a checklist, and a thread. A board and a mission-control dashboard regenerate from them: one static HTML file, no server, opens anywhere.
▸ tasks/co-014-retry-webhooks.md
05
Wired into your systems.
56 tools out of the box: tasks, knowledge, worktrees, GitHub review, database queries, a headless browser, web search, page speed, Search Console. Packs switch on when you add a credential.
▸ SEO · Browser · Database · GitHub · API
06
Private by construction.
Local by default. Secrets sit in a deny-gated store its own tools cannot read, transcripts are redacted before they persist, and nothing leaves the machine unless you commit the folder somewhere.
▸ your keys, your disk, Apache-2.0
// The harness
The harness is the product. The model is a setting.
What makes an agent effective in a business is not which model answers. It is what the agent knows before it starts, what it is allowed to do, and what gets recorded. Those live in files around the model, and they are what ships.
Operating manual
One file says how the agent works, what it may do, and how you like to work. Every host reads it at session start.
Permission rails
Allow and deny lists, a sandboxed shell, and secrets in a store its own tools cannot read. Set to your data classification, not to a default.
Hooks
Session start loads the context. Session end captures it. A guard checks every shell command before it runs. Skills advise; hooks enforce.
Audit trail
Every session captured to a dated file, every task change threaded, the whole folder under git. Who asked, what it did, what was decided.
Observability
A dashboard of today, the board, sessions, and reports, plus a radar of what every running session is doing and where it left off.
Host-neutral
Claude Code today, Codex today, the next host when it ships. One folder works from any of them. The model is a setting.
// Open source as agent-kevin, Apache-2.0 · running in production daily · forked into every deployment
// Mission control
Many agents. One radar.
Every session is a file, so every session can be watched. The radar reads what each agent is doing across every home, marks it running, blocked, or waiting on you, and hands you the resume command. You never hunt for the stuck one.
Blocked · answer to continue · 1
⎇ main
Release checklist
Two options for the rollback window. Pick one and the tag ships.
Running · now · 2
⎇ feat/usage-charts
Usage charts on the billing page
Query written, chart wired, running the replay test against staging.
⎇ main
Morning customer-bug sweep
Reading three support channels, clustering duplicates, mapping each to a task.
Needs input · your move · 1
⎇ main
Refund dispute #4811
Timeline rebuilt from logs and the read-only replica. Draft reply ready for review.
Idle · >1h · 3
⎇ fix/webhook-retries
Webhook retries
Backoff capped at five, idempotency key on the handler, tests green. PR open.
⎇ main
Weekly compliance digest
Access reviews three of five done, evidence filed, digest written.
⎇ main
Search Console audit
Four decaying pages found, rewrite briefs drafted, waiting on the editor.
// Seven sessions, three homes, one operator. Nobody is attached to any of it right now.
// Layer 02 · a head start, then yours
open · Apache-2.0 · self-hosted
Seeded head start
The open-source agent, host-neutral, running in production daily.
private · forward-deployed engineering
Tuned to you
Forked under your keys, named, given a character, permission rails set to your data classification, wired into your workflow.
Give it a folder.
Two commands, five minutes of questions, and it wakes up with a name, a character, and your first project. Everything it learns from there is yours.
▸ /plugin marketplace add github:AgentLayer1/agentlayer-agent-marketplace▸ /plugin install agent-kevin@agentlayer
Inside a claude session launched from the folder. Relaunch, then /agent-kevin:init.