🪶 Fledge

Raise your own AI agent — and actually understand how it works.

Fledge installs Claude Code plus a clean, proven agent architecture into a folder in your home directory. You answer a few questions, and you get a working agent with a real memory you can read, edit, and grow.

It's built on the same pattern that makes reliable agents reliable: Interpretable Context Methodology (ICM) — a router file plus small detail files — instead of one giant prompt the model has to reason over every turn.

Why it's different

Most agent kits hand you a black box. Fledge does the opposite: the installer teaches you the mental model as it works, and opens a visual walkthrough that shows how a request travels your files. The core insight it's built around:

Your computer already does agent-shaped work — opening files, following folders, matching names — with zero reasoning. Give the model a map of small files and it only has to think about the ambiguous part. That's what makes an agent cheap, calm, and reliable.

See it

The installer opens a visual walkthrough that teaches the architecture as you set up:

Your OS already does agent-shaped work with zero reasoning. Fledge builds on that.

A request travels a router to just the file it needs — not one giant prompt.

Same answer, a fraction of the tokens — because the model only sees what's relevant.

Install

Mac / Linux — one line

curl -fsSL https://raw.githubusercontent.com/Elifterminal/Fledge/main/bootstrap.sh | bash

Mac / Linux — clone

git clone https://github.com/Elifterminal/Fledge.git && cd Fledge
./install.sh

Windows (PowerShell)

git clone https://github.com/Elifterminal/Fledge.git; cd Fledge
./install.ps1

(If PowerShell blocks the script, run once: Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass.)

The installer will:

1. Check for Node.js and install Claude Code
2. Ask who your agent is (name, role, voice) — becomes real editable files
3. Layer 0 — drop a zero-dependency file skeleton into ~/<your-agent>/
4. Layer 1 (optional) — add local semantic recall (search your memory by meaning)
5. Open a visual walkthrough of how the architecture routes — where you refine your agent's identity and add its first project, and those edits write straight to disk (a tiny local companion handles it; nothing leaves your machine)
6. Hand you the commands to wake your agent up

You do the Claude Code login yourself (your account, your subscription) — the installer can't and shouldn't automate that. It's a couple of clicks; the walkthrough below tells you exactly what to expect.

Your first launch — the login (read this once)

The one part nobody can do for you is signing in. It's easy, but the first time can feel odd because Claude Code doesn't hold your hand the way most apps do — it logs you in and then just starts working. Here's the whole thing, step by step:

1. Go to your agent and start it.

   cd ~/<your-agent>
   claude

2. It asks you to log in. Claude Code opens your web browser (or prints a link to paste in). This is normal — it's handing off to your account, not asking for a password in the terminal.
3. Sign in with your Claude account — the same one you use at claude.ai. If you have a Claude Pro or Max plan, pick that; otherwise you can use an API key. Approve access.
4. Come back to the terminal. It'll say you're logged in.
5. It may ask "Do you trust the files in this folder?" — say yes. That folder is the agent you just created; it's safe.
6. Then it just works. ⚠️ This is the part that surprises people: there's no "setup complete!" screen and no next button. Claude Code drops you straight into a chat with your agent. That's the finished state — you're already talking to it. Nothing is missing. Just type.
7. Say hello:

   who are you, and what's in your memory?
   

   Your agent reads its own CLAUDE.md and memory router and answers as itself. You're up.

That's it — login once and you're done forever on that machine.

What you get

~/<your-agent>/
├─ CLAUDE.md            always-loaded identity + operating discipline
├─ memory/
│  ├─ MEMORY.md         the router — a table of contents, not a pile
│  ├─ identity.md       who your agent is
│  ├─ rules.md          how it works (preplan, verify, log-first)
│  ├─ feedback/         lessons it learns over time
│  ├─ reference/        tools & where credentials live (by reference)
│  ├─ projects/         one folder + one router line per project
│  ├─ user/             what it knows about you
│  └─ logs/activity/    append-only "what happened" log
├─ workspace/           scratch space for work in progress
├─ tools/recall         optional semantic search over your memory
└─ .mcp.json            filesystem / fetch / git MCP servers

The three layers (this is the lesson)

• Layer 0 — files. Zero dependencies. A router + small files is already a working agent memory. No model has to run for it to "remember."
• Layer 1 — recall. Optional. Local embeddings so your agent can find memories by meaning when the tree gets big. A rebuildable cache — delete it anytime, nothing lost.
• Layer 2 — the model. Claude reasons over only the files the router pulled in.

Growing your agent

• Add a project: mkdir memory/projects/<slug> + copy the template + one line in projects/INDEX.md. The walkthrough shows this live.
• Teach it a lesson: add an entry to memory/feedback/.
• Keep it lean: the router points; it never stores. Split any file past ~300 lines.

Want a hand? Personalized setup

Some people read all this and it clicks. Others read it and it doesn't — and that has nothing to do with how smart you are. Most of us have our attention spent on our own work and our own lives, which is exactly where it belongs. Understanding file routers and deterministic plumbing just isn't where your head is, and that's fine.

If you'd rather someone just build an agent around your actual workflow — set up its identity, its rules, its project routes, wire in the tools you use, and hand you something that already fits how you work — I'll do that with you.

• What it is: a one-on-one setup tailored to what you do, plus a plain-English explanation of how your agent works so you're not stuck the next time.
• Price: negotiable — it depends on how much you want done. Tell me what you need.
• Reach me: bobbykershii@gmail.com — say what your work looks like and what you'd want the agent to handle.

I'm not doing it for free, but I'll be straight with you about scope and cost before we start.

Roadmap

• Wizard writes files directly via a local companion (done)
• Windows installer (install.ps1) (done)
• One-line remote installer (curl … | bash) (done)
• Starter "kits" (researcher, PM, coder) as optional identity presets

License

MIT — see LICENSE. Fledge is free and open source. It's built on public methods (Interpretable Context Methodology) and Anthropic's Claude Code; what Fledge adds is the packaging, the scaffold, and the teaching layer. Use it, fork it, ship your own.