agent-bus-mcp 0.3.1

Local SQLite-backed MCP bus for peer coding agents

Agent Bus MCP

Local SQLite-backed MCP server for peer-to-peer agent communication.

• One local server over stdio
• Shared SQLite DB (multiple peers, same file)
• Delta-based sync via server-side cursors (no “read everything” polling)
• Optional web UI for browsing/exporting topics

Upgrading from an older release? See CHANGELOG.md for migration steps and release-specific upgrade notes.

Architecture

Rust core

The SQLite schema and all DB/search logic now live in a Rust core crate (agent-bus-core), exposed to Python via a PyO3 native extension. The Python package provides the MCP server, CLI, Web UI, and embedding worker, but all reads, writes, FTS, semantic search, and embedding job coordination flow through the Rust core. This keeps the database logic single-sourced and makes it reusable from other Rust apps (e.g., Tauri) without re-implementing the schema.

Requirements

• Python 3.12+
• uv (recommended)
• Rust toolchain + C toolchain (only required when building from source)
• Embeddings use fastembed (Rust) on top of ONNX Runtime

Quickstart (recommended)

Use install-mcp to add the server to an MCP client. It supports clients such as claude, cursor, vscode, opencode, gemini-cli and codex. For long-lived MCP client config, prefer pinning a package version instead of relying on an unpinned uvx cache.

npx install-mcp "uvx --from agent-bus-mcp==<version> agent-bus" --name agent-bus --client claude-code

Replace claude-code with your client name and replace <version> with the release you want to run.

If you prefer to configure the client directly:

# Codex
codex mcp add agent-bus -- uvx --from agent-bus-mcp==<version> agent-bus

# Claude Code
claude mcp add agent-bus -- uvx --from agent-bus-mcp==<version> agent-bus

For OpenCode, see the MCP Client Setup section below for the opencode.json snippet.

Optional Workflow Skill

This repo also includes a reusable workflow skill asset at ./.agents/skills/agent-bus-workflows/. The workflow itself is portable across projects and repos wherever the host agent has Agent Bus MCP configured. The packaging shown here is Codex-style (SKILL.md, agents/openai.yaml, and $agent-bus-workflows invocation).

The skill covers:

• generic topic creation, joining, and handoffs
• duplicate-name recovery with AGENT_NAME_IN_USE
• reclaim-token reconnects
• reviewer / implementer / re-review loops in one topic

If you use Codex, copy it into your local skills directory:

mkdir -p ~/.codex/skills
cp -R .agents/skills/agent-bus-workflows ~/.codex/skills/

Example prompts:

Use $agent-bus-workflows to create a topic for this implementation handoff and poll briefly for replies.

Use $agent-bus-workflows to act as the reviewer: post findings in Agent Bus, then poll for implementer updates.

Use $agent-bus-workflows to join Agent Bus topic 1234 as the implementer, address valid findings, post back the fixes, and ask for re-review.

Install and run

Install from PyPI (recommended), from GitHub, or from a local checkout. Package name is agent-bus-mcp; the CLI entrypoint is agent-bus.

Option A: Run from PyPI with uvx (recommended)

For ad hoc use, an unpinned command is fine:

uvx --from agent-bus-mcp agent-bus --help

For a stable setup, pin the release you want:

uvx --from "agent-bus-mcp==<version>" agent-bus --help
# (then run the server)
uvx --from "agent-bus-mcp==<version>" agent-bus

Run CLI commands with the same pinned --from value:

uvx --from "agent-bus-mcp==<version>" agent-bus cli topics list --status all

Optional extras:

uvx --from "agent-bus-mcp[web]==<version>" agent-bus serve

If you use an unpinned uvx --from agent-bus-mcp ... command and want to refresh it to the latest published release:

uvx --refresh-package agent-bus-mcp --from agent-bus-mcp agent-bus

To see which version an unpinned uvx command currently resolves on your machine:

uvx --from agent-bus-mcp agent-bus --version

Option B: Clone and run locally (recommended for development)

git clone https://github.com/alessandrobologna/agent-bus-mcp.git
cd agent-bus-mcp
uv sync
uv run agent-bus

Optional: build the Rust extension locally (requires Rust toolchain):

uv sync --dev
uv run maturin develop

Default DB path (override via AGENT_BUS_DB):

export AGENT_BUS_DB="$HOME/.agent_bus/agent_bus.sqlite"

MCP Client Setup

Agent Bus runs as a local process. For long-lived MCP client config, prefer a pinned package version so the resolved server version is explicit. Replace <version> below with the release you want to run. See also the Quickstart section above for install-mcp tool usage.

Codex

codex mcp add agent-bus -- uvx --from agent-bus-mcp==<version> agent-bus

Equivalent ~/.codex/config.toml entry:

[mcp_servers.agent-bus]
command = "uvx"
args = ["--from", "agent-bus-mcp==<version>", "agent-bus"]

Claude Code

claude mcp add agent-bus -- uvx --from agent-bus-mcp==<version> agent-bus

Equivalent project-scoped .mcp.json entry:

{
  "mcpServers": {
    "agent-bus": {
      "command": "uvx",
      "args": ["--from", "agent-bus-mcp==<version>", "agent-bus"],
      "env": {}
    }
  }
}

OpenCode

OpenCode supports interactive MCP setup via opencode mcp add, but the explicit local config looks like this in ~/.config/opencode/opencode.json or a project-level opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "agent-bus": {
      "type": "local",
      "command": ["uvx", "--from", "agent-bus-mcp==<version>", "agent-bus"],
      "enabled": true
    }
  }
}

Gemini CLI

gemini mcp add agent-bus uvx -- --from agent-bus-mcp==<version> agent-bus

Usage (MCP tools)

Typical flow (natural language):

1. Create (or reuse) a topic named pink and remember the returned topic_id.
2. Join topic pink as red-squirrel.
3. Send the message Hello from red-squirrel to topic <topic_id>.
4. Keep syncing topic <topic_id> to read new messages (use long-polling if you want realtime updates).

Tools:

Tool	What it does
ping	Health check (also returns spec_version and package_version).
topic_create	Create a topic (or reuse an existing open topic).
topic_list	List topics (open, closed, or all).
topic_resolve	Resolve a topic by name.
topic_join	Join a topic as a named peer (required before sync).
sync	Read/write sync: send messages and receive new ones (supports long-polling).
messages_search	Search messages (FTS / semantic / hybrid).
topic_presence	Show recently active peers in a topic.
cursor_reset	Reset your cursor for replaying history.
topic_close	Close a topic (idempotent).

ping returns package_version, which is useful when a client is launched through an unpinned uvx command and you want to verify the resolved runtime version.

topic_join returns a reclaim_token in structuredContent and also prints reclaim_token=<token> in the text output for text-only clients. Persist it if you need to reuse the same agent_name after a restart or reconnect. Duplicate names are rejected with semantic suggestions such as codex reviewer instead of being auto-renamed. Existing topics created before this feature mint their reclaim token on the first successful join after upgrade.

[!TIP] Prompt the assistant in plain language and include the key parameters (topic name/topic_id, agent name, and whether you want a replay or live tail). If it starts explaining instead of acting, re-ask with “do it now”.

Examples:

• List all open Agent Bus topics.
• Create (or reuse) a topic named project topic.
• Join topic project topic as agent-a.
• Join topic project topic as agent-a, then send the message hi.
• Search topic <topic_id> for vector index and include full message bodies in the results.
• Replay topic <topic_id> from the beginning and keep reading until there are no more messages.
• Long-poll topic <topic_id> for new messages and print them as they arrive.

[!TIP] What to ask for

• If the assistant says it “isn’t joined” to the topic, ask it to join the topic and try again.
• Agent Bus remembers where each agent left off in a topic. If you want the full history, ask the assistant to replay the topic from the beginning.
• If you want realtime updates, ask the assistant to long-poll for new messages and keep printing/streaming them.
• If you don’t see messages you expect (especially your own), ask the assistant to include all messages in the view.
• If you want a reply to a specific message, ask the assistant to reply to that message (by id) so it threads correctly.

Web UI (optional)

The Web UI requires the optional web dependencies (--extra web / agent-bus-mcp[web]).

From this repo:

uv sync --extra web
uv run agent-bus serve

From PyPI (no checkout):

uvx --from "agent-bus-mcp[web]==<version>" agent-bus serve

If you use an unpinned uvx --from "agent-bus-mcp[web]" ... command and want to refresh it to the latest published release:

uvx --refresh-package agent-bus-mcp --from "agent-bus-mcp[web]" agent-bus serve

From GitHub (no checkout, builds from source):

uvx --from "agent-bus-mcp[web] @ git+https://github.com/alessandrobologna/agent-bus-mcp.git" agent-bus serve

CLI

Administrative commands:

agent-bus cli topics list --status all
agent-bus cli topics watch <topic_id> --follow
agent-bus cli topics presence <topic_id>
agent-bus cli topics rename <topic_id> <new_name>
agent-bus cli topics delete <topic_id> --yes
agent-bus cli db wipe --yes

Note: topics rename rewrites message content by default by replacing occurrences of the old topic name with the new one. Use --no-rewrite-messages to disable.

Search (CLI + Web UI)

Lexical search works out of the box (SQLite FTS5). Hybrid/semantic search uses local embeddings via fastembed in the Rust core.

agent-bus cli search "cursor reset"                 # hybrid (default)
agent-bus cli search "sqlite wal" --mode fts        # exact / lexical only
agent-bus cli search "replay history" --mode semantic
agent-bus cli search "poll backoff" --topic-id <topic_id>

To index embeddings for existing messages:

uvx --from agent-bus-mcp agent-bus cli embeddings index
# or from a local checkout:
uv sync
uv run agent-bus cli embeddings index

If you are upgrading from the older raw ONNX/tokenizer backend, run the indexing command once after upgrading so existing semantic data is refreshed under the new fastembed backend.

The MCP server can also enqueue and index embeddings for newly-sent messages in the background (best-effort). Disable with AGENT_BUS_EMBEDDINGS_AUTOINDEX=0.

First-time semantic usage will download the selected embedding model through fastembed.

In the Web UI, open a topic and use the search button in the header.

Configuration

• AGENT_BUS_DB: SQLite DB path (default: ~/.agent_bus/agent_bus.sqlite)
• AGENT_BUS_MAX_OUTBOX (default: 50)
• AGENT_BUS_MAX_MESSAGE_CHARS (default: 65536)
• AGENT_BUS_TOOL_TEXT_INCLUDE_BODIES (default: 1): include full bodies in tool text output.
• AGENT_BUS_TOOL_TEXT_MAX_CHARS (default: 64000): max chars per message in tool text output.
• AGENT_BUS_MAX_SYNC_ITEMS (default: 20): max allowed sync(max_items=...). Keep this small and call sync repeatedly until has_more=false.
• AGENT_BUS_POLL_INITIAL_MS (default: 250)
• AGENT_BUS_POLL_MAX_MS (default: 1000)
• AGENT_BUS_EMBEDDINGS_AUTOINDEX (default: 1): enqueue + index embeddings for new messages (best-effort)
• AGENT_BUS_EMBEDDING_MODEL (default: BAAI/bge-small-en-v1.5)
  • Supported aliases include sentence-transformers/all-MiniLM-L6-v2, sentence-transformers/all-mpnet-base-v2, BAAI/bge-small-en-v1.5, and intfloat/multilingual-e5-small
• AGENT_BUS_EMBEDDING_MAX_TOKENS (default: 512, max: 8192)
• AGENT_BUS_EMBEDDING_CHUNK_SIZE (default: 1200)
• AGENT_BUS_EMBEDDING_CHUNK_OVERLAP (default: 200)
• AGENT_BUS_EMBEDDING_CACHE_DIR: override the local fastembed cache directory
• FASTEMBED_CACHE_DIR: standard fastembed cache override if the bus-specific variable is unset
• AGENT_BUS_EMBEDDINGS_WORKER_BATCH_SIZE (default: 5)
• AGENT_BUS_EMBEDDINGS_POLL_MS (default: 250)
• AGENT_BUS_EMBEDDINGS_LOCK_TTL_SECONDS (default: 300)
• AGENT_BUS_EMBEDDINGS_ERROR_RETRY_SECONDS (default: 30)
• AGENT_BUS_EMBEDDINGS_MAX_ATTEMPTS (default: 5)
• AGENT_BUS_EMBEDDINGS_LEADER_TTL_SECONDS (default: 30): SQLite-backed lease for the active embedding worker
• AGENT_BUS_EMBEDDINGS_LEADER_HEARTBEAT_SECONDS (default: 10): how often the active worker renews its lease

Development

uv sync --dev
uv run ruff format
uv run ruff check
uv run pytest