Agent Orchestrator Starter

---

Your agent setup should outlive your choice of vendor. Memory, project state and a knowledge layer as plain files, so they survive a change of model or tool.

Status (November 2026). Extracted from a personal orchestrator in daily use since March 2026. The Claude Code part is that daily setup. The vendor-neutral core was merged in November 2026. Outside Claude Code it has run in 3 real-work sessions, all on Codex: a read-only audit (August 2026), a website audit and a text-only edit of this kit's own website (November 2026), the last two with the state, inbox and friction-log rules in use. The Copilot adapter is still untested. See VALIDATION.md and STATE.md.

This repository was called Claude Orchestrator Starter Kit until November 2026. The address stays the same and every file from that version is still where it was.

Who it is for: individuals who work with an AI agent every day and want their setup to survive a change of tool. Not tested in teams, no support. Three words used throughout: harness = the agent tool you use (Claude Code, Copilot, Codex); adapter = the notes and config for one harness; MCP = the standard protocol agents use to reach outside tools such as an Obsidian vault.

The idea in one screen

flowchart TB
    H["<b>Your harness</b><br/>Claude Code · Copilot · Codex · opencode"]
    subgraph core["Portable core: plain markdown, no product features"]
        direction LR
        AG["<b>AGENTS.md</b><br/>workflow · rules"]
        MEM["<b>memory/</b><br/>one fact per file<br/>+ pointer index"]
        ST["<b>STATE.md</b><br/>one per project<br/>ONE next action"]
    end
    subgraph vault["Obsidian vault: the one fixed integration"]
        direction LR
        IN["<b>Inbox/</b><br/>agent writes here"]
        GA["<b>Gallery/</b><br/>curated output"]
        REST["everything else<br/>agent reads only"]
    end
    H -->|"reads AGENTS.md"| core
    core -->|"via MCP"| vault
    IN -.->|"you promote<br/>what earned it"| REST

Three files carry the whole thing. Your harness reads them, your vault stores what survives, and you stay the curator.

The problem

Without this	With this
Every session starts at zero. You re-explain the same context.	The agent opens STATE.md and knows where you stopped.
Your setup lives in one vendor's config format.	It lives in markdown you can copy anywhere.
Switching tool or model means rebuilding your workflow.	Copy three files, keep working.
Decisions live in chat history you cannot search.	Decisions live next to the work, with their reasons.
You believe your rules are enforced.	You know which ones are, and which are only requested.

Quick start

Any agent (Copilot, Codex, opencode, ...)

# 1. Clone
git clone https://github.com/janrummel/claude-orchestrator-starter.git
cd claude-orchestrator-starter

# 2. Configure: put the instruction core into your project
mkdir -p ~/your-project/memory
cp AGENTS.md.example ~/your-project/AGENTS.md
cp core/memory/MEMORY.md.example ~/your-project/memory/MEMORY.md
cp core/projects/STATE.md.template ~/your-project/STATE.md

# 3. Run: open your agent in that folder and say
#    "Read memory/MEMORY.md and STATE.md, then continue."

That is the whole installation. Harness-specific notes (MCP config, what your harness can enforce) live in adapters/. Want to see it used before you install it? examples/first-session.md walks one worked, fictional task end to end.

Claude Code

Time: ~5 minutes | Prerequisites: Claude Code + a terminal

# 1. Clone
git clone https://github.com/janrummel/claude-orchestrator-starter.git
cd claude-orchestrator-starter

# 2. Copy to your Claude config
cp CLAUDE.md.example ~/.claude/CLAUDE.md   # overwrites an existing file: back yours up first
cp AGENTS.md.example ~/.claude/AGENTS.md   # the vendor-neutral core, referenced from CLAUDE.md
cp -r orchestrator/ ~/.claude/orchestrator/
mkdir -p ~/.claude/skills && for d in ~/.claude/orchestrator/skills/*/; do ln -sfn "$d" ~/.claude/skills/"$(basename "$d")"; done   # Claude Code looks for skills in ~/.claude/skills/
mkdir -p ~/.claude/memory ~/.claude/orchestrator/projects
cp core/memory/MEMORY.md.example ~/.claude/memory/MEMORY.md   # one fact per file; already using the older memory/ folders? keep them

# 3. Hooks (status line, context monitor, session chaining) need to be registered in
#    ~/.claude/settings.json: follow step 3 of docs/getting-started.md.

# 4. Start Claude Code
claude

Not yet verified: a fresh install from these steps alone on a clean machine (see STATE.md).

Claude will now:

• Read CLAUDE.md at startup and understand its role
• Route your requests to matching skills
• Remember context across sessions via memory files

On Claude Code, hooks can turn rules into blocks instead of requests. This kit ships hooks for session chaining and context monitoring only; docs/portability.md lists what else is possible and what is shipped.

Session Chaining (CLI)

When context fills up, you no longer need to open a new terminal. Once the hooks are registered (step 3 of docs/getting-started.md) and claude-loop is on your PATH:

# Use claude-loop instead of claude
claude-loop

# When context gets critical, Claude saves state automatically.
# Press Y to start a fresh session; it picks up where you left off.

See hooks/README.md for setup details.

Optional Add-ons

Connect your Obsidian vault as Claude's knowledge base. Skills like capture write to it, analyze and express read from it.

See Obsidian Setup for instructions.

For structured data storage (research items, imported datasets, skill usage stats).

This is the beginning of the README. Read the rest at the original. Read the full README ↗