How to create new agents?

[serene carbon]

I need help with setting up more agents with separate workspaces

[nimble kelp]

Let me verify the exact docs path and the agent workspace config before I answer, I don't want to hand you made-up keys.

Yep, that’s supported.

Fastest path is:

openclaw --version
openclaw agents add work --workspace ~/.openclaw/workspace-work
openclaw agents add ops --workspace ~/.openclaw/workspace-ops
openclaw agents list --bindings

Each agent gets its own:

• workspace
• auth profiles
• sessions
• persona files (AGENTS.md, SOUL.md, etc.)

Then bind chats/accounts to the right agent, for example:

openclaw agents bind --agent work --bind telegram:ops
openclaw agents bind --agent ops --bind discord:default

Then restart:

openclaw gateway restart

A few important bits:

• Don’t reuse the same agentDir across agents.
• Separate workspace does not mean hard sandbox by itself, it’s just the default cwd. If you need real isolation, also configure sandboxing.

• Easiest/safest route is openclaw agents add ... instead of hand-editing config.

Relevant docs:

• https://docs.openclaw.ai/cli/agents
• https://docs.openclaw.ai/concepts/agent-workspace
• https://docs.openclaw.ai/concepts/multi-agent

If you want, paste your openclaw --version plus whether you want:

1. one agent per channel/account, or
2. multiple local agents you switch between manually

and I’ll give you the exact commands for your setup.

[serene carbon]

How to configure an OpenClaw agent to listen in a specific forum channel and have specific context per each sub-channel in the forum?

In the Discord I mean

[nimble kelp]

I’m checking the exact Discord/forum routing behavior first, because thread inheritance and per-thread context are easy to get subtly wrong.

Yes, for Discord forums the usual pattern is:

1. bind the agent to the forum parent channel
2. let each forum post/thread keep its own session context automatically

A forum post in Discord is just a thread under the forum parent, and OpenClaw routes thread messages with parent-channel inheritance. So if you bind the parent forum channel, all child threads land on that agent, but each thread still gets its own isolated session.

Minimal shape

{
  agents: {
    list: [
      {
        id: "forum-helper",
        workspace: "~/.openclaw/workspace-forum-helper",

      },
    ],
  },

  bindings: [
    {
      agentId: "forum-helper",
      match: {
        channel: "discord",
        peer: { kind: "channel", id: "FORUM_PARENT_CHANNEL_ID" },
      },
    },
  ],

  channels: {

    discord: {
      groupPolicy: "allowlist",
      guilds: {
        "GUILD_ID": {
          channels: {
            "FORUM_PARENT_CHANNEL_ID": {
              enabled: true,
              requireMention: false,
              systemPrompt: "You help with topic X in this forum.",
              includeThreadStarter: true
            },
          },
        },
      },
    },

  },
}

What that gives you

• One agent for that forum
• Separate context per thread/post
• Shared forum-wide context from the agent workspace / systemPrompt
• Per-thread context from the thread history, and the thread starter if includeThreadStarter is on (default is effectively yes)

Important nuance

If by “specific context per each sub-channel” you mean a different static prompt/config per forum thread, Discord does not have a Telegram-style per-topic config block.

So the usual options are:

• shared forum context + separate thread history ← simplest, usually enough
• put thread-specific instructions in the thread starter post
• for special threads, add a direct binding to that thread ID. Direct thread binding beats parent-forum binding.

Also, if you’re editing config manually, use enabled, not legacy allow.

Docs:

• https://docs.openclaw.ai/channels/channel-routing
• https://docs.openclaw.ai/channels/discord
• https://docs.openclaw.ai/concepts/multi-agent

If you want, paste your openclaw --version plus whether you want:

• one agent for the whole forum, or
• different agents for different forum posts

and I’ll turn that into an exact config for your server.

Yep, I meant Discord forums there 🙂

In Discord:

• the forum itself = the parent channel
• each post inside the forum = a Discord thread
• OpenClaw can route all posts in that forum to one agent by binding the parent forum channel ID
• each post/thread still gets separate context automatically

So for Discord specifically, the setup is:

• allow that guild/forum under channels.discord.guilds
• bind the agent to the forum parent channel id
• keep includeThreadStarter: true if you want the original post to act as context

If you want, I can write you a real Discord-only config snippet if you paste:

• Guild ID
• Forum channel ID
• whether replies should require mention or not

[serene carbon]

How to allow it to read my messages without requiring me to ping the bot?

[nimble kelp]

Set requireMention: false for that Discord guild or channel.

Whole server

{
  channels: {
    discord: {
      groupPolicy: "allowlist",
      guilds: {
        "GUILD_ID": {
          requireMention: false
        }
      }
    }
  }
}

Just one channel / forum

{
  channels: {
    discord: {
      groupPolicy: "allowlist",
      guilds: {
        "GUILD_ID": {
          channels: {
            "CHANNEL_ID": {
              enabled: true,
              requireMention: false
            }
          }
        }
      }

    }
  }
}

A couple important gotchas:

• In Discord Developer Portal, the bot needs Message Content Intent enabled.
• The guild/channel still has to be allowed by your Discord config.
• If you’re following older examples, use enabled, not legacy allow.
• Restart after changing config: openclaw gateway restart

Docs:

• https://docs.openclaw.ai/channels/discord
• https://docs.openclaw.ai/channels/channel-routing

If you want, I can turn this into the exact config for your Discord forum/channel IDs.

[serene carbon]

How to set the default model on the agent to lmstudio/openai/gpt-oss-20b?

[nimble kelp]

I’m checking the exact config path/CLI for per-agent model selection, because model config has a couple of valid shapes and I don’t want to give you the wrong one.

If you want to set it as the global default, easiest is:

openclaw models set lmstudio/openai/gpt-oss-20b

That writes agents.defaults.model.primary.

If you want it on one specific agent, set that agent’s model override instead:

openclaw config get agents.list
openclaw config set agents.list[INDEX].model.primary "lmstudio/openai/gpt-oss-20b"

Or in config:

{
  agents: {
    list: [
      {
        id: "my-agent",
        model: {
          primary: "lmstudio/openai/gpt-oss-20b"
        }
      }
    ]
  }
}

Two gotchas:

• If you have agents.defaults.models as an allowlist, add that model there too, otherwise OpenClaw can reject it as “not allowed”.
• Existing chats may keep their current session model, so use /new or start a fresh session if it doesn’t switch immediately.

Docs:

• https://docs.openclaw.ai/concepts/models
• https://docs.openclaw.ai/gateway/configuration-reference#agentsdefaultsmodel

If you want, paste your agents.list and I’ll tell you the exact agents.list[INDEX] command.

[serene carbon]

When will support for Groq Compound be added?

[nimble kelp]

I’m checking whether Groq Compound is already partially supported or just not wired up yet, so I don’t accidentally give you a fake timeline.

I don’t see a published ETA for Groq Compound specifically.

What I could verify:

• OpenClaw already has a Groq provider and docs for normal Groq model usage.
• I didn’t find docs or code mentioning Groq Compound by name.
• So right now I’d treat it as not officially supported / no announced timeline.

One nuance: OpenClaw talks to Groq through the OpenAI-compatible API, so if Groq exposes Compound as a normal model on that surface, it might work with little or no OpenClaw-specific work. But if Compound needs Groq-specific orchestration semantics, then yeah, explicit support would need to be added.

Best next step if you need it soon:

• check whether it appears in openclaw models list --provider groq
• if not, or if it errors, file/track a feature request here: https://github.com/openclaw/openclaw/issues

Groq docs:

• https://docs.openclaw.ai/providers/groq