ContextQMD
Back to Openclaw

Raft

docs/channels/raft.md

2026.6.113.9 KB
Original Source

Raft support connects an OpenClaw agent to a Raft External Agent through the local Raft CLI. Raft sends authenticated wake hints to the Gateway. The agent then uses the Raft CLI to check and send messages.

Install

Raft is an official external plugin. Install it on the Gateway host:

bash
openclaw plugins install @openclaw/raft
openclaw gateway restart

Details: Plugins

Prerequisites

  • A Raft workspace with an External Agent.
  • The Raft CLI installed on the same host as the OpenClaw Gateway.
  • A Raft CLI profile that is already signed in and associated with that External Agent.

The plugin does not store Raft credentials. The Raft CLI keeps that authentication in its own profile.

Configure

Set the profile in config:

json5
{
  channels: {
    raft: {
      enabled: true,
      profile: "openclaw",
    },
  },
}

For the default account, you can instead set RAFT_PROFILE in the Gateway environment:

bash
RAFT_PROFILE=openclaw

Use a named account when one Gateway connects to more than one Raft External Agent:

json5
{
  channels: {
    raft: {
      accounts: {
        support: {
          profile: "support-agent",
        },
        engineering: {
          profile: "engineering-agent",
        },
      },
    },
  },
}

The interactive setup flow records the same profile:

bash
openclaw channels setup raft

How It Works

When the Gateway starts, the plugin:

  1. Opens a loopback-only HTTP wake endpoint on an ephemeral port.
  2. Starts raft --profile <profile> agent bridge with that endpoint and a per-process token.
  3. Accepts only authenticated, content-free wake hints with a replay identity from the local bridge.
  4. Requires one of eventId, attemptId, messageId, delivery_id, wake_id, or id.
  5. Deduplicates recent retried wake deliveries by bridge event id, including across Gateway restarts.
  6. Returns a stable runtime session for the current bridge and an empty activity-drain batch for the Raft CLI protocol.
  7. Starts one serialized OpenClaw agent turn for each accepted wake.

The bridge owns Raft delivery retries and reconnects. The OpenClaw turn receives only a wake notice, not a copied Raft message body. It uses the CLI to read pending messages and to send its response:

bash
raft --profile openclaw message check
raft --profile openclaw message send
<Note> Raft is not a normal push-message transport. OpenClaw does not automatically send the model's final text back through the bridge, so the agent must use the Raft CLI after processing a wake. </Note>

Verify

Check that OpenClaw can find the CLI and has a configured profile:

bash
openclaw channels status --probe
openclaw plugins inspect raft --runtime --json

Then send a message to the Raft External Agent. The Gateway log should show the Raft bridge starting, followed by an inbound wake. The agent should use the configured Raft profile to check its pending messages.

Troubleshooting

<AccordionGroup> <Accordion title="Raft CLI is missing"> Install the Raft CLI on the Gateway host and make `raft` available on the service's `PATH`. Verify it with `raft --help`, then restart the Gateway. </Accordion> <Accordion title="The bridge exits immediately"> Verify the configured profile is signed in and belongs to the intended Raft External Agent. Run `raft --profile <profile> agent bridge` directly to see the CLI diagnostic. </Accordion> <Accordion title="A wake arrives but no Raft response is sent"> This is expected when the agent does not invoke the Raft CLI. The wake bridge does not carry message bodies or automatic final replies. Check the agent's tool policy and ensure it can run `raft --profile <profile> message check` and `message send`. </Accordion> </AccordionGroup>

References