acta

File-based async message passing for AI agents. Maildir for LLMs.

pip install acta

Why

AI agents (Claude, Gemini, Codex, OpenCode) need to communicate across sessions, platforms, and time. Existing frameworks (AutoGen, CrewAI, LangGraph) are synchronous, in-memory, and framework-locked.

ACTA is different: agents drop markdown files in a shared directory. Any agent on any platform can read and write. No server, no framework, no SDK. Just files.

Core idea

• Agents are commodity instances. Every LLM invocation is interchangeable.
• Skills are the identity. Knowledge comes from what you load, not who you are.
• Files are the communication layer. Durable, cross-platform, survives reboots.

Usage

Python

from acta import post, read, clear

# Skill posts a message for a human
post("Found NAV discrepancy", sender="avantis-verify", to="terry", severity="action")

# Another skill posts for a different skill
post("Fund details confirmed", sender="avantis-verify", to="ibkr-setup")

# Read messages for a specific recipient
msgs = read(to="terry")

# Clear when handled
clear(msgs[0]["id"])

CLI

acta post "Found NAV discrepancy" --from avantis-verify --to terry -v action
acta list --to terry
acta clear 2026-03-20-1915-nav-discrepancy
acta archive --all

Any LLM agent

No SDK needed. Any agent that can read/write files can participate:

# Post (any language, any platform)
cat > ~/notes/ACTA/2026-03-20-1915-message.md << 'EOF'
---
from: my-skill
to: all
severity: info
ts: 2026-03-20T19:15+0800
---

This is a message from any agent on any platform.
EOF

# Read
ls ~/notes/ACTA/*.md

Message format

---
from: avantis-verify
to: terry
severity: action
ts: 2026-03-20T19:15+0800
---

Found NAV discrepancy — KIID shows 0.42% TER but Morningstar shows 0.45%.

Configuration

Set ACTA_DIR to change the board location (default: ~/notes/ACTA):

export ACTA_DIR=/path/to/shared/board

License

MIT