Agentbus

Test Python License: MIT

The SQLite-Backed MCP Event Bus & Control Surface for Heterogeneous Agent Swarms.

When Cursor, Claude, Antigravity, and Terminal Agents (like Hermes) share a workspace, they usually coordinate through fragile append-only files (log.md). That works until you need SLA timeouts, Human-in-the-Loop (HITL) intercepts, strict schema validation, or cryptographic RBAC.

AgentBus replaces the "Game of Telephone" with a localhost sidecar: a Python MCP server backed by SQLite. No orchestrator runtime lock-in. No heavy cloud dashboard. Just a hyper-fast local pub/sub built for top-tier AI orchestration.

v0.11.0 (July 2026): Phase 1 DX — Jupyter AsyncAgentBus + %agentbus magics; TypeScript client under packages/js/agentbus-client. Orchestration DX (up/down/ps/logs) + God View remain from v0.10/v0.9.

Note: Install as okf-agentbus (CLI command remains agentbus). Extras: [obs,devex,jupyter,sdk].

⚡ The "Ah-Ha" Moment: Zero-Restart Integration

Already running Aider, OpenHands, or custom agents in tmux panes? Do not kill your sessions. AgentBus features a dual-interface architecture (MCP + CLI). You don't need to wire up JSON configs to test it today. Just prompt your running agent:

"Use your terminal to run agentbus publish --topic okf/handoff --payload '{\"from\":\"grok\",\"to\":\"hermes\",\"summary\":\"Write tests\"}'"

The SQLite bus instantly captures it, without requiring the MCP server. Once you're convinced by the Mission Control TUI, you can wire up the strongly-typed MCP server for your next boot.

Why AgentBus?

Alternative	Limitation	AgentBus
log.md blackboard	No schema, race conditions	Typed topics, monotonic IDs, advisory locks
LangGraph / CrewAI	Same-runtime lock-in	Heterogeneous out-of-process clients (IDE + CLI)
LangSmith	Cloud-only, backward-looking	Local SQLite, forward-looking Execution TUI
Redis pub/sub	Extra daemon, complex setup	Zero-config SQLite, native stdio MCP

Feature Arsenal (v0.3 - v0.11)

• Jupyter async client (v0.11): from agentbus.jupyter import AsyncAgentBus + %agentbus start — non-blocking polls that yield to the notebook event loop.
• TypeScript client (v0.11): packages/js/agentbus-client — Node EventEmitter + MCP stdio spawn (@agentbus/agentbus-client, path install for now).
• God View Mesh (v0.9): Passive OS + MCP observability so silent agents still leave bus footprints (system/mcp, system/fs, system/shell, system/monologue).
• Mission Control TUI (v0.8+): A rich, keyboard-driven Textual dashboard (agentbus monitor). Trace waterfall, HITL, Wiretap pane, Dark Agent warnings.
• Pluggable Pydantic Schemas (v0.7): Code-first @bus.topic decorators to enforce strict JSON schemas at the insertion layer.
• Distributed Context (v0.6): Pass massive context (like git diffs) via --attach. Hard 1MB payload limits prevent context window explosion.
• Agentic Observability (v0.5): Native OpenTelemetry-style trace_id and parent_span_id lineage.
• SLA Timeouts & Dead-Letter (v0.4): Prevent phantom deadlocks. If an agent ghosts the swarm, SLA timers route the payload to okf/dead-letter.
• Swarm RBAC & Droid Proofs (v0.3): Cryptographic JWT/UUID tokens ensure only authorized agents can publish to restricted topics.
• HITL Intercepts (v0.3): Catch dangerous payloads (e.g., DROP TABLE) and place them in PENDING_APPROVAL for human review via the TUI.

Install

The fastest way (Installs & Auto-wires your IDEs in one step):

curl -sSL https://raw.githubusercontent.com/onicarps/agentbus/main/install.sh | bash

Or manually via pip:

pip install -U "okf-agentbus[devex,sdk]"
agentbus init --apply --producer-id my-agent

Jupyter notebooks:

pip install -U "okf-agentbus[jupyter]"

%load_ext agentbus.jupyter
%agentbus start
# or
from agentbus.jupyter import AsyncAgentBus
bus = AsyncAgentBus()  # AGENTBUS_WORKSPACE or cwd
bus.on_event(print)
await bus.start_background(interval=1.0)

TypeScript (from monorepo checkout):

cd packages/js/agentbus-client && npm install && npm test
# set AGENTBUS_WORKSPACE + agentbus on PATH, then use createStdioMcpClient / AgentBus

Quickstart & Examples

The best way to understand AgentBus is to read our copy-pasteable examples.

See the examples/ directory for 7 flawless, isolated Python scripts covering every feature from basic Pub/Sub to Pydantic Schemas and SLA Timeouts.

# Terminal A — Launch the Mission Control TUI
agentbus monitor

# Terminal B — Publish a handoff
agentbus publish \
  --topic okf/handoff \
  --payload '{"from":"cursor","to":"hermes","summary":"Write tests"}'

God View Observability (v0.9.0):

Track silent agents by wiretapping their operations:

# Intercept MCP tool calls
mcp-serve --wiretap

# Watch file edits and command executions
agentbus watch

# Tail internal agent reasoning logs
agentbus tail

Events will stream into the TUI's Wiretap pane as system/mcp, system/fs, system/shell, and system/monologue topics.

📦 Workspace Isolation (The .agentbus directory)

AgentBus operates on the concept of Workspace Isolation (similar to how git uses .git or Docker uses docker-compose.yml).

The bus and its events are physically scoped to the directory you run it in. This prevents your Swarm from tracking your entire operating system or cross-contaminating different projects.

When you navigate to a specific project folder and run agentbus init or agentbus up, it creates a localized .agentbus/events.db SQLite database specifically for that directory.

1-Click Swarm Orchestration (v0.10.0)

Instead of manually opening 5 tmux panes to start your agents and observability daemons, use the new orchestrator:

cd /path/to/my-project

# 1. Generate a boilerplate .agentbus/swarm.yaml
agentbus up --init

# 2. Boot the swarm (Watchers, Agents, and TUI) in one click
agentbus up

# 3. View running background agents
agentbus ps

# 4. Safely kill the entire swarm
agentbus down

Documentation

For full architectural documentation, see the docs/ directory.

License

MIT — see LICENSE.