Building one agent is easy. Making several agents collaborate is where the time goes. Once your system has three to five agents, the real cost isn’t model calls — it’s the unglamorous plumbing around them: who has which capability, who can take this task, how context gets shared, and how results get tracked.
Ant Group’s inclusionAI team has open-sourced Avernet V0.1 (Apache 2.0) to attack exactly that collaboration layer. It is not another agent framework; it is infrastructure that lets heterogeneous agents join one collaboration network — with native support for OpenClaw. A single script brings up a five-bot “virtual company” of OpenClaw agents.
The production credentials are real: as of early July 2026, Avernet-based multi-agent deployments covered 12 business units across Ant Group, with multi-agent workflow task completion above 90%. An open-source project with production mileage is worth an hour of your time.
The Core Idea: Findability, Alignment, Speed, and State
Avernet’s positioning in one sentence: it does not replace your agents’ reasoning — it turns the collaboration process into infrastructure. Registration, discovery, invitations, group management, sessions, routing, consensus, and traceable execution are all standardized.
The architectural core is a BCS (Bot Collaboration Service) that handles connections, registration, routing, message delivery, session management, collaboration state, and multi-bot network management. A frontend workbench sits on top, with a bcs-cli command-line tool for onboarding and inspection.
Two integration paths exist:
- Plugin integration: OpenClaw instances and local agent runtimes connect into BCS proactively — the most practical path today, and how the bundled demo bots work.
- Gateway integration: Avernet dispatches tasks to an existing bot platform, which schedules the work and returns results — the right fit for brownfield systems.
The connection chain reads plainly:
Local OpenClaw gateway
-> openclaw-channel-bcn plugin
-> ws://127.0.0.1:21000/ws/bot
-> BCS
-> bcs-cli onboard
-> bot appears in the Avernet networkStep 1: One Script to Bring Up the Local Environment
Clone the repository and run the script. It installs anything missing — Node.js 22, Rust/Cargo, protoc, OpenClaw:
git clone https://github.com/inclusionAI/Avernet.git
cd Avernet
# preflight + auto-install missing tools
./scripts/singlebox.sh install-tools
# start BCS + frontend workbench + 5 OpenClaw demo bots
./scripts/singlebox.shWhen it finishes, open the frontend workbench at http://127.0.0.1:8000/. The script launches five OpenClaw demo bots with distinct personas: ceo, product-manager, engineering, verification, and customer-service — a CEO that delegates, engineers that execute, and customer service as the safety net. A functioning virtual company, onboarded automatically through the BCN plugin.
Verify the setup:
# BCS health check (default port 21000)
curl -fsS http://127.0.0.1:21000/health
# list bots in the network
./src/bcs/target/debug/bcs-cli --url http://127.0.0.1:21000 listIf you see Bots in network (...) followed by the five bot names, you’re up.
Step 2 (Optional): Wire In a Real Model
Avernet’s basics — connecting, onboarding, collaboration testing — require no model API key. To make the demo bots produce real responses, set three OpenAI-compatible environment variables. All three must exist together; a partial set is silently ignored:
export OPENCLAW_OPENAI_BASE_URL=<model-api-base-url>
export OPENCLAW_OPENAI_API_KEY=<model-api-key>
export OPENCLAW_OPENAI_MODEL_ID=<model-id>Put them in .env.local (already git-ignored) or your current shell — never inside a tracked file.
Step 3: Onboard Your Own OpenClaw Instance
To attach an extra local OpenClaw instance, first build the BCN plugin:
cd src/bcs/crates/plugins/openclaw-channel-bcn
npm install && npm run build
# verify the artifact
test -f dist/esm/index.jsThen generate an isolated OpenClaw configuration — strongly preferred over editing ~/.openclaw/openclaw.json directly. The key part is registering the bcs channel:
cat > .openclaw-host-bcn/openclaw.json <<'EOF'
{
"channels": {
"bcs": {
"enabled": true,
"bcsUrl": "ws://127.0.0.1:21000/ws/bot",
"botId": "host-openclaw",
"botName": "Host OpenClaw",
"capabilities": {
"summary": "Local OpenClaw gateway on host machine",
"domains": ["local", "openclaw"],
"skills": ["openclaw"],
"scopes": ["local"]
},
"heartbeatIntervalMs": 60000,
"reconnectIntervalMs": 5000,
"connectionTimeoutMs": 30000
}
},
"gateway": {
"port": 18789,
"mode": "local",
"bind": "loopback"
},
"plugins": {
"load": { "paths": ["/path/to/Avernet/src/bcs/crates/plugins/openclaw-channel-bcn"] },
"entries": { "openclaw-channel-bcn": { "enabled": true } }
}
}
EOFStart the gateway (keep the terminal open). Once the plugin connects to BCS, a bot token lands in .openclaw-host-bcn/.bcs/session.json:
OPENCLAW_DATA_DIR="$(pwd)/.openclaw-host-bcn" \
OPENCLAW_CONFIG_PATH="$(pwd)/.openclaw-host-bcn/openclaw.json" \
openclaw gateway run --port 18789Finally, use the token to onboard — registering the bot’s name, capabilities, and visibility:
BOT_DATA_DIR="$(pwd)/.openclaw-host-bcn" \
./src/bcs/target/debug/bcs-cli --url http://127.0.0.1:21000 onboard \
--name "Host OpenClaw" \
--summary "Local OpenClaw gateway on host machine" \
--domains "local,openclaw" \
--skills "openclaw" \
--scopes "local"
# make it appear in the collaboration list
BOT_DATA_DIR="$(pwd)/.openclaw-host-bcn" \
./src/bcs/target/debug/bcs-cli --url http://127.0.0.1:21000 visibility set --value publicPractical Advice
- For a first run, just use singlebox.sh. Watch how the five bots go from plugin build to symlinked extensions to generated profile to automatic onboarding — that chain is the template for attaching your own agents.
- Keep your own OpenClaw in an isolated directory. The documented
.openclaw-host-bcn/approach leaves~/.openclawuntouched — model keys and personal configs stay safe. - Mind the ports: BCS on
21000, frontend on8000, demo bots on30001-30041. If anything collides, changeBCS_PORT/FRONTEND_PORTin.env.local, or pass--bcs-port/--frontend-portat startup. - Don’t treat this as a production deployment guide. The local scripts run in debug mode with mock authentication — right for proving the pipeline and integration testing. Production needs the security, audit, and governance capabilities on the roadmap.
- Set expectations by version. This is V0.1: memory, orchestration, evaluation, and evolution are still listed as planned. What’s worth studying today is the architecture, the open protocol, and the integration model — those are the durable parts.
Resources
- GitHub repository: github.com/inclusionAI/Avernet (Apache 2.0)
- Quick start:
docs/quick-start.zh-CN.md - Manual OpenClaw integration guide:
docs/openclaw-bcn-local.zh-CN.md - Bot-platform gateway integration:
docs/bot-provider-integration.zh-CN.md
