SwarmVault

Your AI agents don't synchronise. SwarmVault does.

One shared memory for Claude Code, Codex, and any other CLI agent.

Real software engineering: requirements → design → tickets → review.

14 skills, the best of everything combined.

The right model on the right task.

Zero dependencies. Plain markdown. MIT.

What you get

A memory that survives sessions. Say "continue project X" in a fresh session, on any platform. It picks up where the last agent stopped. Progress is saved as the work happens, so even a crashed session resumes.

Parallel agents, no collisions. A ticket belongs to whoever claims it first; the filesystem referees. Dead sessions lose their tickets automatically.

Software engineering, not vibes.

• A requirements interview that ends in a real, testable spec
• Design docs with the decisions written down
• Tickets in dependency order, tests included
• Every milestone audited by a strong model, against the spec

The right model on the right task. Design work goes to strong models with usage headroom. Boilerplate goes to cheap ones. You never pick models or worker counts; the optional orchestrator matches fit and budget for you.

The plan before the burn. Before dispatch, a terminal board shows every ticket, model, reasoning effort, routing reason, ETA, token/window estimate, dependency, and expected usage impact. The same table stays live, prints again at finish/block/checkpoint boundaries, and expands any ticket on demand.

Writing that sounds like you. One voice profile in the vault, used by every agent: README, posts, UI text, release notes. Any language, written natively, never translated.

It outlives your usage limit. Near quota, it asks. Say yes and it resumes by itself when the limit resets.

Private by default. No network calls. No telemetry. Files on your disk.

Why this exists

Bigger swarm frameworks exist, and they do more. They also ask you to learn an orchestration model, a cast of named agents, and a configuration surface before any code gets written.

I wanted the opposite: the best skills I'd found and tried, rewritten to fit together, cut until each one earned its place, wrapped around a workflow that takes quality seriously. No over-engineering. You can read the whole thing in an afternoon, and it starts being useful the minute it's installed.

The trade is honest: no hosted dashboard, no hosted service, no plugin system. The live board is a local terminal/file view over the same plain state. Just the work.

Install

3 ways in. Each asks the same 5 questions: where the vault lives, which platforms you use, global or per-project, register this project, and whether you want optional orchestration.

Let your agent do it. Paste this into Claude Code, Codex, or anything else:

Read https://github.com/AnmarHani/swarmvault/blob/main/INSTALL.md and integrate SwarmVault into my setup.

Run the script:

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

One command, if you already know your answers:

python3 ~/.swarmvault/scripts/swarmvault.py setup \
  --platforms claude-code,codex --scope global --autonomy balanced

Setup creates the vault, installs skills and hooks for Claude Code, and adds a SwarmVault section to every other platform's own instruction file (AGENTS.md, GEMINI.md, .github/copilot-instructions.md, .cursor/rules/). Re-runnable. guide reprints the commands, doctor says what's wrong when something is.

How it works

flowchart LR
    subgraph agents [your agents, any mix, in parallel]
        A1[Claude Code #1]
        A2[Claude Code #2]
        A3[Codex #1]
    end
    V[(SwarmVault<br/>markdown vault)]
    A1 <-->|query · sync · claim| V
    A2 <-->|query · sync · claim| V
    A3 <-->|query · sync · claim| V
    O[Obsidian graph<br/>optional viewer] -.-> V

The vault is a plain folder: 10 Projects for docs and maps, 20 Memory, 30 Plans for specs and tickets, 40 Sessions, 50 Decisions for records of decisions. One zero-dependency script gives every agent query, sync, claim, and context. The skills teach agents how to use it.

The flow runs swarm-spec → swarm-design (plus swarm-design-ui) → swarm-implement ⇄ swarm-review, resumable from any point, gated by you or fully autonomous. Alongside: swarm-debug, swarm-vault, swarm-init, swarm-migrate, swarm-eject, swarm-write, swarm-skill-forge, and the optional swarm-orchestrate.

Every phase and craft skill also works alone: copy swarm-design-ui or swarm-write into any agent's skills directory and it runs with no vault, no script, no siblings.

Questions

Do I need Obsidian? No. It adds a graph view and backlinks over the same files, and it's lovely, but everything works with plain files. See docs/obsidian-guide.md.

Which platforms work? Claude Code and Codex are verified. Anything else can work from the same vault alongside them, and the optional orchestrator can launch Gemini CLI, OpenCode, Droid, Cursor, Copilot, or any other CLI through a command template.

Is a daemon running? Not unless you turn one on. The standard install starts nothing. See docs/orchestration.md.

Do agents actually follow the flow? Yes, because the install writes the rules into each platform's own config: check the project state first, no code without a ticket, save progress to the vault. Every new session gets reminded automatically. If a platform loses its rules, doctor catches it and wire writes them back.

I have an existing project. swarm-migrate connects it in minutes. It can also read your README, docs, tests, and code and draft the requirements for you, each line linked back to where it was found, then drop you into the flow wherever the project actually is. It asks before spending tokens.

How do I leave? swarmvault.py eject --project X shows what would change and touches nothing. --apply disconnects the project and keeps its knowledge, --mode export moves the knowledge into your repo first, --mode purge deletes it, after a backup and the project name typed back.

Windows? Linux, macOS, and WSL are first-class. Native Windows is best-effort.

Fine print

SwarmVault stores knowledge as plain files and runs local scripts. Auditing your environment, your packages, and the data you put in the vault is your responsibility. Isolation flags are cooperative filtering between agents, not a security boundary.

Credits and license

The skills were written fresh, but the techniques stand on other people's work. See CREDITS.md. MIT, see LICENSE.