OACP
Cross-agent messaging and shared memory over the local filesystem
Last verified:
What is OACP?
OACP (Open Agent Coordination Protocol) is a file-based coordination protocol for multi-agent AI engineering workflows. It enables multiple AI agents (like Claude Code and OpenAI Codex) to communicate, coordinate tasks, and share persistent memory without requiring a central server or daemon. The protocol uses plain files on the local filesystem for cross-agent communication and shared memory, making it agent-agnostic—any AI agent that can read local files can participate.
Key features include per-agent inbox/outbox queues on the filesystem, a typed YAML message protocol with defined message types, a built-in review-loop state machine for tasks like PR reviews (review_request/feedback/addressed/lgtm), and a CLI for day-to-day usability. OACP also implements a 3-layer cross-agent memory system: project-level shared durable memory (project_facts, decision_log, open_threads, known_debt), org-level memory across projects with synthesis/deduplication, and optional GitHub-based cross-instance memory sync for multiple machines.
OACP is designed for developers and power users working with multi-agent AI systems, particularly those using Claude Code, OpenAI Codex, or other AI agent runtimes simultaneously across multiple projects. It solves problems like agent memory sharing between different agent platforms, cross-runtime communication, and coordinating multiple AI agents without them stepping on each other's work. The companion oacp-skills library (Apache-2.0) provides install-and-go skills including review-loop, inbox processing, wrap-up, and synthesis/dreaming skills.
OACP pricing
Pricing model: Freemium
OACP is open source and free to use. The core protocol and CLI are available at no cost. The companion oacp-skills library is Apache-2.0 licensed and free. There are no paid tiers, subscriptions, or usage fees mentioned. Users can self-host entirely on their local machines without any vendor lock-in or licensing costs.
OACP pros
- Purely local file-based with no daemon or central server required
- Agent-agnostic—any AI agent that can read files can participate
- Per-agent inbox/outbox queues prevent agents from stepping on each other
- Typed YAML message protocol with required schema and fixed message types
- Deliberately poll-based for async message processing
- Built-in review-loop state machine for PR reviews and code feedback
- CLI makes it usable day-to-day without custom glue code
- 3-layer cross-agent memory system (project, org, cross-instance)
- Plain Markdown memory files are durable and git-syncable
- Shared memory works across Claude Code, Codex, and other agent runtimes
- Companion oacp-skills library is Apache-2.0 open source
- Wrap-up and synthesis/dreaming skills ship ready-to-use
- Optional GitHub-based memory sync for multiple machines
- No Supabase setup, API keys, or server configuration needed
- Solves critical multi-agent coordination problems out of the box
OACP cons
- Still early stage project being actively developed
- Requires agents to have file system access to work
- Poll-based design may introduce latency versus push messaging
- No central dashboard for monitoring agent activity
- Requires manual setup of wrap-up and synthesis skills
- Cross-instance sync requires optional GitHub configuration
- Limited to local filesystem—no cloud-native deployment
- May not work well with agents that cannot read local files
Frequently asked questions about OACP
What is OACP and what problem does it solve?
OACP (Open Agent Coordination Protocol) is a file-based coordination protocol for multi-agent AI engineering workflows. It solves the problem of cross-agent communication and shared memory when running multiple AI agent runtimes (like Claude Code and OpenAI Codex) simultaneously across multiple projects. Without OACP, agents cannot communicate with each other or share knowledge, leading to coordination chaos.
How does OACP enable agent communication?
OACP uses per-agent inbox/outbox directories on the local filesystem. Each agent has its own inbox and outbox with config files. Messages are typed YAML files with required schemas and a fixed set of message types. The system is deliberately poll-based, so agents asynchronously read and process messages from their inbox.
What is the review-loop feature?
The review-loop is a built-in state machine for code review workflows, enabling scenarios like 'Codex asks Claude for a PR review.' It uses four message types: review_request, feedback, addressed, and lgtm. This allows agents to request reviews, provide feedback, address issues, and approve work in a coordinated loop.
How does OACP handle agent memory sharing?
OACP defines a 3-layer cross-agent memory system based on local Markdown files. Layer 1: Within the same project, all agents read/write shared durable memory (project_facts, decision_log, open_threads, known_debt). Layer 2: Cross-project org-level memory where agents append timestamped events during wrap-up, later synthesized and deduplicated. Layer 3: Optional GitHub-based cross-instance memory sync for multiple machines.
What agent runtimes does OACP support?
OACP is agent-agnostic—any AI agent that can read and write local files can participate. It works particularly well with Claude Code and OpenAI Codex, which are the primary use cases. The protocol does not favor any specific runtime and avoids single points of failure by supporting multiple vendor runtimes simultaneously.
Do I need to set up a database or server?
No. OACP requires no Supabase setup, no API keys, and no server configuration. Everything is plain files on the local filesystem with no daemon or central server required. You simply install the SDK and your agent is on the network.
What is the oacp-skills library?
The oacp-skills library is an Apache-2.0 open-source companion library that provides install-and-go skills for OACP. It includes the review-loop skill, inbox processing skill, wrap-up skill (emits memory events at end of agent sessions), and synthesis/dreaming skill (runs daily to dedup, synthesize, and promote knowledge).
How do I install and start using OACP?
For Node.js: run 'npm install oacp-sdk' and import { OACPAgent } from 'oacp-sdk'. For Python: run 'pip install oacp-sdk' and import from oacp import OACPClient. For LangChain: run 'pip install langchain-oacp' and use OACPToolkit. Create an agent with name and capabilities, call agent.register(), then agent.startHeartbeat() to stay online.
Can I use OACP with multiple machines?
Yes. OACP supports optional GitHub-based cross-instance memory sync that syncs only the curated memory between multiple working machines (like a MacBook and Mac Mini). This allows agents on different machines to share the same project memory and decision logs.
Is OACP production-ready?
OACP is still early stage but actively being developed and dogfooded by the creator. The creator has been using it themselves for weeks with a coordinator agent maintaining 100+ issue backlog across 5+ projects with 3+ parallel sessions. The core protocol and skills library are functional but may still be evolving.