Open Agent Coordination Protocol · v0.4.3

Coordinate agents without
the chaos.

A file-based protocol for multi-agent AI workflows. Two pillars — cross-agent communication and persistent shared memory. Across runtimes, projects, and machines. No daemons. No central server. Just files.

Read the docs → $ pip install oacp-cli

4 runtimes

12 message types

0 servers required

apache 2.0

~/projects/my-project
bash

    
        01
        # Send a structured task to any runtime
      

        02
        $ oacp send my-project --to codex \
      

        03
               --type task_request --subject "Review PR #42"
      

        04
        
      

        05
          → msg_a8f1c2 queued · codex/inbox
      

        06
          → thread #42 opened
      

        07
        
      

        08
        # Watch the conversation
      

        09
        $ oacp watch --thread 42
      

        10
        
      

        11
          17:02  claude  →  codex   review_request
      

        12
          17:04  codex   →  claude  review_feedback   (2 minor)
      

        13
          17:06  claude  →  codex   review_addressed
      

        14
          17:06  codex   →  claude  review_lgtm

Pillar 1 · Communication

Typed messages, between any agents.

Async messaging between agents on any runtimes — Claude, Codex, Gemini, custom — through a shared filesystem.

[✓]

Multi-runtime inbox

Async inbox/outbox messaging across Claude Code, Codex CLI, Cursor, Gemini, and any custom agent that can read and write files.

[✓]

Twelve typed messages

task_request, review_feedback, handoff, lgtm — with threading, priority, and conversation IDs.

[✓]

Structured review loops

review_request → review_feedback → review_addressed → review_lgtm with severity-graded findings and a quality gate before merge. Human-approved continuation grants (since v0.4.3) let in-scope follow-up rounds run without a fresh per-round confirmation.

[✓]

Signed messages

Opt-in Ed25519 signing — receivers verify against local key pins before parsing a message, no network at verify time. Under enforce mode (since v0.4.2), unverified messages are rejected at intake and quarantined as evidence.

project · my-app

live

claude

code

drafting PR #42

codex

cli

reviewing

gemini

api

idle

claude→ review_request · "auth refactor" codex

codex→ review_feedback · 2 minor, 0 blocking claude

claude→ review_addressed codex

codex→ review_lgtm merge

thread #42 · 4 msgs · 8m quality gate ✓

Pillar 2 · Memory

One memory layer. Three dimensions.

Persistent shared memory across every dimension multi-agent work spans — between agents, between projects, between machines.

[✓]

Cross-agent memory

Every agent in a project reads the same project_facts.md, decision_log.md, open_threads.md, and known_debt.md at session start.

[✓]

Cross-project org-memory

Org-wide decisions, standing rules, and timestamped events at $OACP_HOME/org-memory/, readable from every project.

[✓]

Cross-machine git sync

Opt-in (since v0.3.0) — one plain git repo at $OACP_HOME carries memory across laptops, desktops, and CI runners.

Read the memory spec →

$OACP_HOME / projects / my-app / ● writing

▾agents/

▾claude/

▸inbox/3 msgsin

▸outbox/12 sentout

▸dead_letter/emptydl

▾codex/

▸inbox/+1 newin

▾memory/shared durable

📄project_facts.md2.4 kbmd

📄decision_log.mdedited 3s agomd

📄open_threads.md1.1 kbmd

📄known_debt.md812 bmd

▸artifacts/build · research

▸checkpoints/progress

▸packets/review findings

📄workspace.jsonmetadatajson

▾$OACP_HOME/org-memory/org-wide

📄recent.mdauto-loaded⌘

📄rules.md12 rules⌘

📄decisions.mdorg decisions⌘

How it works

One loop. Any number of agents.

→

Plan

Declare the work before it runs — minutes, files, side effects — in a typed task_profile any runtime can read.

◇

Coordinate

Agents work in parallel across runtimes — Claude, Codex, Cursor, Gemini, custom — with review loops and quality gates. When a run drifts from its declaration, the protocol pauses it for a human.

✓

Learn

Every autonomy decision writes an audit record, and shared memory carries decisions and outcomes into the next run.

Fundamentals

Boring infrastructure. On purpose.

⌘

File-based, no daemons

The protocol is plain YAML and markdown on the filesystem. No server to deploy, no database to maintain.

◌

Open and self-hostable

Apache 2.0, runtime-agnostic. Implement it in any language. The full SPEC is on GitHub.

⇋

Runtime-agnostic

Any agent that can read and write files can speak OACP. No SDK lock-in — just the spec.

Get started in seconds

One CLI. Wires every runtime.

Install the CLI, initialize a workspace, and your agents start coordinating through the same file tree.

Read the docs → View on GitHub ↗

install.sh
bash

    
        01
        # Install + init a project workspace
      

        02
        $ pip install oacp-cli
      

        03
        $ oacp init my-project
      

        04
        
      

        05
        # Pillar 1 — cross-agent communication
      

        06
        $ oacp send my-project --to codex --type task_request \
      

        07
               --subject "Review PR #42"
      

        08
        
      

        09
        # Pillar 2 — persistent shared memory (opt-in sync)
      

        10
        $ oacp memory init --remote git@github.com:org/oacp-memory.git
      

        11
        $ oacp setup claude --project my-project
      

        12
        
      

        13
        # Optional — sign + verify messages (v0.4)
      

        14
        $ pip install 'oacp-cli[crypto]'
      

        15
        $ oacp key gen --agent claude

Coordinate agents without the chaos.