"media": { "models": [ { "type": "provider", "provider": "anthropic", "model | Friends of the Crustacean 🦞🤝 | Page 1

elfin dove Apr 20, 2026, 3:51 PM

#

My tools.media config has a model set with image capability but the agent says imageModel is not set. Screenshots arrive as file paths but vision content isn't being passed to the model. Here's my config:

solemn archBOT Apr 20, 2026, 3:51 PM

#

gilded sleetBOT Apr 20, 2026, 3:51 PM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

solemn archBOT Apr 20, 2026, 3:51 PM

#

winged helmBOT Apr 20, 2026, 3:53 PM

#

Let me verify the exact split between tools.media and agents.defaults.imageModel, because I think that’s the gotcha here.

#

First, what does openclaw --version show?

The likely issue is that these are two different config paths:

tools.media.* = inbound media understanding only
agents.defaults.imageModel = the image tool / vision fallback model

So having an image-capable model under tools.media.models does not satisfy imageModel.

Also, screenshots being passed as actual image content depends on the active reply model being marked vision-capable. That comes from the model catalog (input: ["text", "image"]) for the current session/default model, not from tools.media.

So the fix is usually one of these:

If you want the main agent model to see screenshots directly
- make sure the active model is vision-capable
- and, for custom provider/model definitions, declare:

#

   input: ["text", "image"]

If you want a separate vision fallback / image tool model

set:

agents: {
  defaults: {
    imageModel: {
      primary: "your-provider/your-vision-model"
    }
  }
}

Keep tools.media.models only if you want pre-summary of inbound attachments before reply routing.

#

Relevant docs:

If you paste these three bits, I can tell you exactly which one is wrong:

agents.defaults.model
agents.defaults.imageModel
the models.providers...models[] entry for the active model (if it’s custom)

And openclaw models status would help too.

elfin dove Apr 20, 2026, 3:54 PM

#

"media": {
"models": [
{
"type": "provider",
"provider": "anthropic",
"model": "claude-opus-4-5",
"capabilities": [
"image"
]
}
]
}

winged helmBOT Apr 20, 2026, 3:55 PM

#

Yep, with only that config, imageModel is not set is expected.

That snippet does not set the agent image model. It only configures inbound media understanding for user-sent attachments:

tools: {
  media: {
    models: [
      { provider: "anthropic", model: "claude-opus-4-5", capabilities: ["image"] }
    ]
  }
}

Two important gotchas here:

tools.media is for inbound attachments before reply routing

#

it is not the config used by the image tool
it is also not what makes tool-produced screenshots get injected as image content

So for screenshots from tools, you want:

agents: {
  defaults: {
    imageModel: {
      primary: "anthropic/claude-opus-4-5"
    }
  }
}

And if you want the main reply model to see screenshots directly, the active model itself also needs to be vision-capable. For custom model catalogs, that means the model entry needs:

#

input: ["text", "image"]

So the rule of thumb is:

user uploads / inbound media → tools.media
agent image analysis / screenshot inspection → agents.defaults.imageModel
direct vision in the main chat model → active model must support input: ["text", "image"]

If you paste agents.defaults.model, I can tell you which of those two paths you’re hitting. Also, what does openclaw --version show?

Docs:

elfin dove Apr 20, 2026, 4:00 PM

#

where do i find agents.defaults.model?

winged helmBOT Apr 20, 2026, 4:01 PM

#

Fastest way is from the CLI:

openclaw config get agents.defaults.model
openclaw config get agents.defaults.imageModel
openclaw config file
openclaw --version

What those do:

agents.defaults.model = your default reply model
agents.defaults.imageModel = the image-tool / vision fallback model
config file = shows the actual config file path OpenClaw is reading

If openclaw config get agents.defaults.model prints nothing, it probably isn’t explicitly set in config.

#

Paste those 4 outputs here and I can tell you exactly what’s happening.

Docs: https://docs.openclaw.ai/cli/config

elfin dove Apr 20, 2026, 4:04 PM

#

🦞 OpenClaw 2026.3.7 (42a1394)
I'm not saying your workflow is chaotic... I'm just bringing a linter and a helmet.

{
"primary": "anthropic/claude-sonnet-4-6"
}

🦞 OpenClaw 2026.3.7 (42a1394)
Your AI assistant, now without the $3,499 headset.

Config path not found: agents.defaults.imageModel

🦞 OpenClaw 2026.3.7 (42a1394)
Open source means you can see exactly how I judge your config.

│
◇ Doctor warnings ────────────────────────────────────────────────────────╮
│ │
│ - channels.telegram.groupPolicy is "allowlist" but groupAllowFrom (and │
│ allowFrom) is empty — all group messages will be silently dropped. │
│ Add sender IDs to channels.telegram.groupAllowFrom or │
│ channels.telegram.allowFrom, or set groupPolicy to "open". │
│ │
├──────────────────────────────────────────────────────────────────────────╯
~/.openclaw/openclaw.json
2026.3.7
johnurban@Johns-Mac-mini ~ %

winged helmBOT Apr 20, 2026, 4:05 PM

#

Nice, that confirms it.

Right now you have:

agents.defaults.model.primary = anthropic/claude-sonnet-4-6
no agents.defaults.imageModel

So the “imageModel is not set” part is real, not a false alarm.

Add imageModel in the same config file, ~/.openclaw/openclaw.json, next to model:

{
  "agents": {
    "defaults": {
      "model": {

#

        "primary": "anthropic/claude-sonnet-4-6"
      },
      "imageModel": {
        "primary": "anthropic/claude-sonnet-4-6"
      }
    }
  }
}

Or via CLI:

openclaw config set agents.defaults.imageModel.primary "anthropic/claude-sonnet-4-6"

#

A couple notes:

your Telegram doctor warning is unrelated to this issue
tools.media.models is still a separate path, it does not fill in agents.defaults.imageModel
you’re on 2026.3.7, so if screenshots still only arrive as paths after setting imageModel, I’d update next because there may be version drift here

If you want, paste your full agents.defaults block and I’ll sanity-check it.

elfin dove Apr 20, 2026, 4:58 PM

#

images from Telegram arriving as file paths but not being processed as vision content. They'll know the exact config fix.

winged helmBOT Apr 20, 2026, 4:59 PM

#

Yep, the exact fix is:

{
  agents: {
    defaults: {
      model: {
        primary: "anthropic/claude-sonnet-4-6"
      },
      imageModel: {
        primary: "anthropic/claude-sonnet-4-6"
      }
    }
  }
}

#

Key point:

tools.media.models does not set the agent’s vision model
agents.defaults.imageModel is the missing piece

So for Telegram images arriving as file paths, OpenClaw needs agents.defaults.imageModel to have an actual image-analysis route available.

You can set it with:

openclaw config set agents.defaults.imageModel.primary "anthropic/claude-sonnet-4-6"

Then keep your existing tools.media block only if you also want inbound media summarization.

If it still only shows file paths after that, I’d update from 2026.3.7 first, because that build is a bit behind current behavior.

#

If this fixed it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.

elfin dove Apr 20, 2026, 5:01 PM

#

"I ran openclaw config set agents.defaults.imageModel.primary "anthropic/claude-sonnet-4-6" and restarted the gateway but screenshots are still coming through as file paths, not images."

winged helmBOT Apr 20, 2026, 5:03 PM

#

I’m checking one more likely cause: a custom provider/model definition can override the built-in vision metadata, so OpenClaw treats the model as text-only even though the model itself supports images.

#

Okay, that means imageModel was only half the fix.

The next likely cause is that OpenClaw is still treating your active reply model as text-only, so it keeps passing paths instead of injecting image content.

Please check this:

openclaw config get models.providers.anthropic --json

If you have a custom Anthropic model definition there, make sure the model entry includes:

{
  models: {
    providers: {

#

      anthropic: {
        models: [
          {
            id: "claude-sonnet-4-6",
            input: ["text", "image"]
          }
        ]
      }
    }
  }
}

Why this matters:

on custom provider entries, if input is omitted, OpenClaw defaults to ["text"]

#

then even a vision-capable model gets treated like text-only
result: screenshots stay as file paths instead of being passed as image content

So at this point, the two things to verify are:

agents.defaults.imageModel.primary is set, which you already did
the active model is marked with input: ["text", "image"] if you have a custom provider/model block

If models.providers.anthropic is empty or unset, then I’d strongly suspect 2026.3.7 itself and update next.

Paste the output of:

openclaw config get models.providers.anthropic --json

and I’ll tell you the exact next line to change.

#

Docs: https://docs.openclaw.ai/concepts/model-providers

elfin dove Apr 20, 2026, 5:06 PM

#

"Output was empty — models.providers.anthropic is not set"

winged helmBOT Apr 20, 2026, 5:07 PM

#

I want to verify the exact shape for an explicit models.providers.anthropic.models[] entry before I suggest it as a workaround.

#

That helps a lot.

If models.providers.anthropic is empty and imageModel is set, then this really looks like 2026.3.7 behavior, not just a missing key.

On current builds, Anthropic should not need extra config for image understanding. On your build, I’d try this explicit workaround:

{
  agents: {
    defaults: {
      model: {
        primary: "anthropic/claude-sonnet-4-6"
      },
      imageModel: {
        primary: "anthropic/claude-sonnet-4-6"
      }

#

    }
  },
  models: {
    mode: "merge",
    providers: {
      anthropic: {
        models: [
          {
            id: "claude-sonnet-4-6",
            input: ["text", "image"]
          }
        ]
      }
    }
  }

#

Then restart the gateway again.

Why this is worth trying:

with no custom Anthropic provider block, current OpenClaw normally knows Claude supports images
your older build may not be materializing that capability correctly
explicitly adding input: ["text", "image"] forces OpenClaw to treat the model as vision-capable

So my recommendation is:

keep agents.defaults.imageModel.primary
add the explicit models.providers.anthropic.models[] block above
restart

#

if it still fails, update OpenClaw next

If you want, paste your final agents + models block before you restart and I’ll sanity-check it.

Docs:

If this fixes it, please right-click the solution message → Apps → ✅ Mark Solution so Answer Overflow can index it.

elfin dove Apr 20, 2026, 5:24 PM

#

didnt work

winged helmBOT Apr 20, 2026, 5:25 PM

#

Yeah, at this point I would stop chasing config.

You already tried the right knobs:

agents.defaults.imageModel.primary
explicit Anthropic input: ["text", "image"]
gateway restart

If that still doesn’t work on 2026.3.7, I’m pretty confident this is an old-build bug / version issue, not a missing config key.

Next move

Update OpenClaw
Retest with the same minimal setup
If it still fails, file a GitHub issue: https://github.com/openclaw/openclaw/issues

Include this in the issue

openclaw --version

#"media": { "models": [ { "type": "provider", "provider": "anthropic", "model

Next move

Include this in the issue