Yash Agrawal

Agent framework

ClosedClaw

A local, capsule-based multi-agent orchestrator for Python agents: discover or create the right capsule, route through LangGraph, gate risky work, execute allowlisted tools, and preserve evidence.

Open GitHub Back to systems
ClosedClaw artifact board with a capsule file tree, run terminal, approval gate, and evidence panel.
Artifact-style board from repository surfaces, not a product screenshot.
Status
Live repository
Proof type
Agent framework
Stack
LangGraph LangChain sqlite-vec Python JSONL audit
Source
github.com/rogue-socket/closed_claw

What it is

Agents as reusable capsules

ClosedClaw packages agents with a manifest, skills, memory, entrypoint, and logs. A LangGraph coordinator discovers matching capsules through embeddings and reranking, then runs the chosen agent as an isolated subprocess.

Why it exists

Agent prototypes need an operating surface

The repo is aimed at the point where one-off agents become hard to trust: reuse decisions, paid API calls, tool permissions, and logs need to be visible instead of buried inside a prompt script.

How it works

Discover, approve, execute, audit

01Task
02Registry
03Capsule
04ToolExec
  1. Embed the task and search the capsule registry with sqlite-vec.
  2. Use the coordinator graph to choose discovery or execution.
  3. Pause at human approval gates for creation, reuse, and paid calls.
  4. Run allowlisted tools and leave JSONL logs plus SQLite audit events.