Swarmvault: A Lean LLM Agent Workflow

Adding more "skills" to an AI agent is a trap; every new capability consumes context, and eventually, the model's performance degrades. Most agent frameworks suffer from this—they are bloated with features and complex configurations that you have to memorize before you even write a line of code. I wanted something minimal enough to hold in my head, so I built Swarmvault.

git clone https://github.com/AnmarHani/swarmvault && cd swarmvault && ./install.sh

Context-First Design

The framework consists of only 13 core skills. This isn't an arbitrary number; it's what remained after stripping away overlap and redundancy. The philosophy here is that if a capability is missing, it should be a thin extension rather than a heavy new skill competing for the same context window. Every skill is modular—you can drop one into any agent's directory, and it functions independently without needing the rest of the vault.

Spec-Driven Execution

Instead of relying on "vibes" or hopeful prompting, this follows a strict technical pipeline:
1. A requirements interview that produces a real SRS.
2. A design phase where decisions are formally logged.
3. Tickets generated in dependency order with clear "definition of done" tests.

A strong model audits the work against the spec at every milestone. This eliminates the "filler" code and fluff documentation that usually plague agent outputs.

The Vault System

The "vault" is simply a local folder structure:

• 10 Projects
• 20 Memory
• 30 Plans
• 40 Sessions
• 50 Decisions

It uses a zero-dependency script to give agents query, sync, claim, and context capabilities. This allows you to start a fresh session and say "continue project X," and the agent picks up exactly where it left off. Because agents "claim" tickets, you can run multiple agents in parallel without them duplicating work.

Model Routing and Compatibility

The system includes an optional orchestrator that distributes tasks across different platforms (Claude Code, Codex, Cursor) based on model strengths and usage limits. Adding a new CLI is handled via a launch-command template rather than a complex plugin system.

Current Support:

• Verified Platforms: Claude Code, Codex
• Best-effort Platforms: Gemini CLI, OpenCode, Cursor, Copilot
• OS: Linux, macOS, WSL (Native Windows is best-effort)

There are no background processes, no telemetry, and no hosted services—just plain Markdown files on your disk under an MIT license.

https://github.com/AnmarHani/swarmvault