Your personal AI assistant, as a Claude Code or Codex plugin. One markdown folder, one plugin, a brain that learns who you are session after session.
Kevin is a portable, file-based personal AI assistant that plugs into the agent CLI you already use, Claude Code or OpenAI Codex. Everything that makes Kevin Kevin (personality, memory, knowledge, projects, tasks) lives in your own directory as plain markdown. Any AI can read it. You can browse it in Obsidian or Finder. If you ever want to leave, you take the folder and go.
It is not a chat wrapper. It is an operating system for personal AI:
- A 59-tool MCP server for tasks, knowledge compilation, reports, home history, worktrees, database queries, GitHub review, search, page speed, a bundled browser, and Google Search Console.
- A 31-skill library, related work grouped into one skill with playbooks (briefing, goals, SEO, focus, engineering), covering onboarding, version history, project lifecycle, a daily focus view with standup and a session radar, goals from the day to the year, morning and evening briefings, tax deadlines and bookkeeping, trip planning, worktree setup, API-request drafting, read-only SEO auditing, and audio and video transcripts and summaries.
- A knowledge pipeline that turns every conversation into structured, queryable memory.
- Opt-in packs (SEO, Browser, Database, GitHub, API, Xcode) and a bridge to community skill libraries via skills.sh.
- You drive. Ask in plain words and Kevin picks the skill and playbook; nothing runs unasked, and you never have to remember a command.
graph LR
A[Sessions] -->|capture| B[Knowledge]
B -->|informs| C[Projects]
C -->|generate| D[Results]
D -->|feed back into| A
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.
Kevin is named after the loyal minion. Helpful, enthusiastic, a little nerdy.
You need Bun ≥ 1.1, Git, and a host: Claude Code or Codex. Full prerequisites, the local clone install, and Windows notes are in Install.
Pick a home for the brain, then install the plugin from it:
mkdir -p ~/Documents/Agents/Kevin && cd ~/Documents/Agents/Kevin| Claude Code | Codex |
|---|---|
claude, then /plugin install agent-kevin --marketplace AgentLayer1/agentlayer-agent-marketplace |
codex plugin marketplace add AgentLayer1/agentlayer-agent-marketplace then codex plugin add agent-kevin@agentlayer |
Relaunch and run /agent-kevin:init |
Create the home from Claude Code, then open it with codex and run $upgrade once |
Skills: /agent-kevin:<skill> |
Skills: $<skill> |
Five minutes of questions later you have a home. See Onboarding and Hosts.
Want a head start? The wizard turns eleven prompts about your company into a seed bundle; hand the zip to init and the agent wakes up named, characterised, and briefed. A teammate's seed export does the same from an existing agent. → Seed bundles
Always launch from the agent home. The plugin loads only for sessions started there, and that is also what keeps several agents on one machine apart. Reach your code through permissions.additionalDirectories in the home's .claude/settings.local.json (init writes it there, since the path is this machine's), not by launching from a repo.
This README is the short version. Everything lives at agentlayer.one/docs, also served as llms-full.txt for models.
| Section | Start with |
|---|---|
| Getting started | Install · Onboarding · Your first session · Updating |
| Dashboard | Today · Tasks and projects · Sessions · Brain · Reports and scheduler · Capabilities · Persona and system |
| Platform | The agent home · The brain · Capture · Sync · History · Self-evolution · Seed bundles · Multiple agents |
| Agent | Hosts · Claude Code · Codex · Hooks · Configuration · Tasks · Daily rhythm · Architecture |
| Modules | Plan and run · Build and ship · Reach and see · Brain and memory · Skills · MCP tools · Second-model review · Browser · SEO · Accounts |
| Engineering | The engineer skill · Principles · Design and review · Pull requests · Worktrees · Specs and plans · Coding rules · Verification · API collections · Releases |
| Reference | CLI · Upgrades · Naming · Changelog |
| Workstation | The rig: Ghostty · cmux · editor and tools |
| About | Privacy · Platforms · FAQ · History · Contributing |
- Memory that compounds. Hooks capture every session; the
knowledge-compileskill distils them into user facets, concept articles, and active memory that load next launch. → The brain - Projects, not just chats. One markdown file per task with frontmatter, threads, and a generated dashboard. → Tasks
- A focus page for the day. The
focusskill keeps today to three, shows what carried over and what the roadmap has in flight, and thegoalsskill sets the day, week, month and year with you; one page for all your work and one per project. → Daily rhythm - One pass to bring everything current. The
syncskill runs compile → lint → flywheel → dashboards and ends with a next move. → Sync - Every change can be undone. The
historyskill turns on local version history for the home in one question, no git knowledge needed; sync saves a snapshot each run. A home in iCloud, Dropbox or OneDrive keeps its history in~/.local/state(%LOCALAPPDATA%on Windows), where syncing can't damage it. → History - A mission-control page regenerated on every sync, self-contained, zero external requests. → Dashboard
- Engineering by playbook. The
engineerskill routes code work to a playbook (bug fix, feature, refactor, performance, forensics, prototype, and more) backed by 23 named principles, proves the result on the running artifact, and strips comments before you see the diff. → Engineering - Pull requests, three ways. Review a teammate's PR with verified findings and paste-ready comments, prep to present your own, or brief a second model on your branch and have its findings verified, fixed, and committed on one dossier. → Pull requests
- Multiple homes, multiple personas. One plugin install, separate brains, told apart by the launch folder. → Multiple agents
- Hand an agent to a teammate. A seed bundle carries persona, curated knowledge, and setup with fork semantics; credentials travel as names only. → Seed bundles
- Subscription-billed. The MCP server returns prompts; your session does the thinking on your host's plan. → Hosts
- Private by construction. Secrets in a deny-gated store Kevin cannot read, transcripts redacted before they persist. → Privacy
Pull the new plugin version on your host (/plugin update agent-kevin@agentlayer in Claude Code; remove and re-add in Codex), then run the upgrade skill. The plugin update refreshes code; upgrade reconciles your home from the CHANGELOG's Upgrade blocks, backing up first. → Updating · Upgrades and releases
Pull requests welcome: new opt-in packs, read-mostly MCP tools, platform hardening, docs. Open an issue first for architectural changes; Kevin's contract with the markdown home is intentional. See CONTRIBUTING.md.
Licensed under the Apache License, Version 2.0. agent-kevin is © AgentLayer · agentlayer.one. See NOTICE for the attribution stanza. Third-party skill libraries installed via the configure-skills skill are not bundled and carry their own licenses.
Built by AgentLayer · agentic infrastructure for AI-native operations