OPENCOVEN · PERSISTENT AI FAMILIARS

Where AI Meets
Continuity.

Named familiars with memory, tools, identity, and continuity — running locally, inside your projects, on your terms.

Remember
Memory and context that survives across sessions.
Coordinate
Multiple agents, one shared runtime.
Act
Tools, terminals, and docs under local control.
Own
Open, inspectable, yours to extend.
coven cli local · active
memory · forge 3 notes
resumed feat/runtime-attach · 4 files staged 2h ago
prefers terse PR summaries, no trailing recap persisted
Drives the agent CLIs you already use
Claude Code · Anthropic Codex · OpenAI + more adapter-ready
01

AI that can stay

Familiars keep a name, role, memory, and place in the work across every session.

02

Local runtime substrate

Coven keeps execution bounded to explicit projects and tools. Not a harness — the layer beneath.

03

Shared ecosystem

Codex, Claude Code, CastCodes, dashboards, and docs built to fit together.

04

Owned by builders

Open framework, inspectable architecture, practical extension points — yours to modify.

Architecture

One neutral layer between the harness and your code.

Coven sits between every supported agent CLI and the project you're actually working in. Commands flow down, output and memory flow up, and every boundary is explicit.

Coven architecture: harness, runtime, and project layers harness layer coven runtime your project project-scoped boundary output · memory Codex Claude Code + Future Sessions Memory Adapter Tool gateway Filesystem Git Terminals Docs

pick a node

How Coven fits into your stack

Pick any node to inspect how it fits into the runtime.

  • Commands flow down: harness → runtime → your project
  • Output, state, and familiar notes, decisions, and conventions that persist across sessions." data-href="/docs/reference/glossolalia">memory flow back up the return path
  • Every boundary is enforced at runtime, not just drawn here

How It Works

Pick a layer. See what it actually does.

Memory, sessions, harnesses, and tools are the four moving parts. Each tab is a working preview of how that piece behaves in a real Coven session.

Memory

Continuity that survives the chat window.

Project context, decisions, and preferences accumulate per familiar across sessions. Step through the timeline on the right to see how a familiar's working memory grows from one session to the next.

context project: opencoven · stack: astro + vanilla
preference prefers terse PR summaries · no trailing recap
state resumed feat/interactive-deep-dive · 4 files staged

Compare

Why Coven, not just a harness?

Same agent CLIs, different runtime around them. Pick a row to see how each option actually behaves.

Capability Raw harness Claude Code, Codex direct Other tools IDE-bound agents, suites Coven this runtime

Cross-session memory

Raw harnesses lose state when the chat ends; most other tools save chat history but not curated working memory per familiar. Coven stores explicit notes that persist and surface in the next session.

Live Demo

A real Coven session, end to end.

Watch the runtime check harnesses, launch a familiar, detach, list sessions, and resume. Use the pause control to stop at any step.

coven · live session
running

Quick Start

Run your first familiar in three commands.

Coven runs locally inside the projects you already have. Verify, start, and resume — memory persists across every session.

01 · Verify

npx @opencoven/cli doctor

Checks your environment and detects supported harnesses — Codex, Claude Code, and more.

02 · Start a session

coven run claude "fix the failing tests"

Launches a project-scoped session through Coven's managed PTY, bound to the current working directory.

03 · Resume anywhere

coven sessions

Interactive browser to rejoin, summon, archive, or view logs — work picks up where you left off.

Read the full docs View on GitHub ↗

Ecosystem

A living workspace of agents that belong to you.

OpenCoven turns AI from a blank chatbox into a durable workspace: familiars with memory and tools, local runtime control, open dashboards, desktop surfaces, and an ecosystem builders can inspect, customize, and own.