weave-openclaw

This is an OpenClaw plugin that sends a record of what your agents do to W&B Weave, so you can see and search it all in one dashboard. Once it's running, every agent run, model call, and tool call shows up in the Agents tab of your Weave project, along with the full conversation, token usage, and cost. By default it sends that content, unredacted, to W&B (a third-party service).

Data and privacy

This plugin sends a trace of your agent activity to W&B Weave, a hosted service. With captureContent on (the default) that includes the full, unredacted text of prompts, replies, and tool inputs and results. Set captureContent: false to send only structure and token and cost totals.

Requirements

• Node.js >= 22
• OpenClaw >= 2026.4.25
• A free W&B account and a project. Sign up at wandb.ai if you don't have one.

Install

Install it with OpenClaw's plugin manager. This works from any directory, so you don't have to be inside a particular project:

openclaw plugins install weave-openclaw

Use the full name weave-openclaw; weave on its own is the W&B SDK, not this plugin. This registers the plugin with your gateway. You don't import it in your own code; the gateway loads it from the config you set up next.

Quickstart

1. Get a W&B API key at wandb.ai/authorize and export it:

   export WANDB_API_KEY=<your-key>
   

2. Add the plugin to your OpenClaw gateway config (by default ~/.openclaw/openclaw.json; run openclaw onboard to create one):

   {
     diagnostics: { enabled: true },
     plugins: {
       allow: ["weave"],
       entries: {
         weave: {
           enabled: true,
           config: { entity: "your-team", project: "your-project" },
           hooks: { allowConversationAccess: true },
         },
       },
     },
   }
   

3. Restart the gateway. In any OpenClaw chat, run /weave status to confirm the plugin says running. Your traces will appear at wandb.ai/<entity>/<project>/weave/agents.

hooks.allowConversationAccess: true is easy to miss. Without it, OpenClaw won't share prompts and replies with a third-party plugin, so model-call spans arrive without prompts, responses, or per-call token counts. Trace structure, tool calls, and run-level cost and token totals still come through. (diagnostics.enabled is on by default; setting it to false disables tracing.)

Configuration

Only entity and project are required. Everything else has a sensible default.

{
  plugins: {
    entries: {
      weave: {
        enabled: true,
        config: {
          entity: "your-team",        // your W&B team or username
          project: "your-project",    // your W&B project name

          // Leave apiKey out to use the WANDB_API_KEY environment variable.
          // To read the key from a file instead:
          //   apiKey: { source: "file", provider: "default", id: "/run/secrets/wandb" }
          // A plain key string works too, but keeping secrets out of config is safer:
          //   apiKey: "your-wandb-api-key"
          apiKey: { source: "env", provider: "default", id: "WANDB_API_KEY" },

          serviceName: "openclaw-agent",   // shown in /weave status
          // These help group and label your agents in the dashboard.
          agentName: "my-agent",
          agentVersion: "v1.0",
          agentDescription: "What my agent does.",

          // On by default. Set to false to stop recording the actual message
          // text (for example, to meet a privacy or retention policy). The
          // plugin records text as-is and does not hide sensitive values, so
          // remove them beforehand if you need to.
          captureContent: true,

          // How often (in milliseconds) traces are sent.
          flushIntervalMs: 1000,
        },
        hooks: { allowConversationAccess: true },
      },
    },
  },
}

Where the plugin finds your API key

It checks these in order and uses the first one it finds:

1. The apiKey you set in config (from an environment variable or a file)
2. A plain apiKey string in config
3. The WANDB_API_KEY environment variable
4. Your ~/.netrc, which wandb login fills in for you

You can also put WANDB_API_KEY=... in ~/.openclaw/.env. OpenClaw loads that file into the gateway's environment at startup, so it feeds option 3 above. A key already set in your shell takes precedence.

Most people don't need to change where traces are sent. If you're on a dedicated or self-hosted W&B install, set the WANDB_BASE_URL environment variable to your install's address; the plugin and the wandb command-line tools read it the same way.

What gets recorded

When everything is connected, each trace includes:

• Every agent run, with its name and a running total of cost and tokens
• Every model call, with the model name and how many tokens went in and out
• Every tool call, with the tool name and its result

With content recording on (the default), the full input and output messages and the tool inputs and results are saved too. Extra details like subagents, retries, and context compaction are recorded as well.

Viewing your traces

After you restart the gateway, run an agent through OpenClaw. Within a few seconds (set by flushIntervalMs) your traces appear at:

https://wandb.ai/<entity>/<project>/weave/agents

Open the Agents tab in the left nav for the conversation view and per-agent grouping, or the Traces tab for the full step-by-step view. The complete Weave docs are at weave-docs.wandb.ai.

Troubleshooting

The plugin is loaded but nothing shows up

1. Run /weave status. If it doesn't say running, the plugin didn't start. The usual causes are a missing entity or project, or an out-of-date plugin. The gateway log says which.
2. Diagnostics is on by default; check you haven't set diagnostics.enabled: false.
3. Make sure entity and project match the project you're looking at. /weave status prints them as project=<entity>/<project>.
4. Check which key is being used. /weave status prints auth=.... If it points at the wrong place, fix the key.
5. If runs appear but model calls or messages are empty, set hooks.allowConversationAccess: true (see below).

Traces show up but the messages are blank

OpenClaw is hiding the conversation text. Set hooks.allowConversationAccess: true (under plugins.entries.weave) and restart the gateway. Structure, tool I/O, and run-level totals come through regardless; only prompts, responses, and per-call token counts need this setting. The gateway log shows the blocked hooks.

Traces aren't reaching W&B