Cline · VS Code · MCP Servers panel

An agent that plans its own work should be able to make its own assets.

Cline breaks a request into steps and executes them. Without image tools, every step that needs artwork becomes a note in the summary. With them, it becomes a step like any other.

Get a free API key Generic MCP docs

Add the server

Open the MCP Servers panel from the Cline toolbar and edit the configuration. Cline persists it to cline_mcp_settings.json, so editing that file by hand has exactly the same effect.

If your Cline build only accepts stdio servers, use the npx mcp-remote form shown underneath — it bridges the hosted endpoint into a local process and works everywhere.

remote
{
  "mcpServers": {
    "imagemcp": {
      "url": "https://mcp.imagemcpserver.com/mcp",
      "headers": { "x-api-key": "sk-img-gen-…" },
      "disabled": false
    }
  }
}
stdio fallback
{
  "mcpServers": {
    "imagemcp": {
      "command": "npx",
      "args": ["-y", "mcp-remote",
        "https://mcp.imagemcpserver.com/mcp",
        "--header", "x-api-key:sk-img-gen-…"]
    }
  }
}

Auto-approve by cost, not by convenience

An autonomous agent with a spending tool needs a policy. The useful line to draw is not between “image tools” and “other tools” — it is between calls with a fixed price and calls that route to a model.

get_user_infolist_models

free

Safe to auto-approve. These are how the agent finds out what it can afford.

compress_imageconvert_format

1 credit

Deterministic and cheap. Auto-approving these makes optimisation passes painless.

remove_backgroundtext_to_svgupscale_image

5–15 credits

Fixed prices, no model routing. Auto-approve once you trust the workflow.

generate_imageedit_imagegenerate_transparent_imagemulticall

model-priced

These are the ones that can run up a bill in a loop. Keep approval on.

Give it a budget in words

Cline follows written constraints well. This is the block worth keeping in your custom instructions.

Image work uses the imagemcp MCP server.

1. Call get_user_info before any batch of
   image work and report the balance.
2. If the plan needs more than 100 credits,
   stop and ask me first.
3. Draft on a fast model. Only re-run on a
   premium model once I approve the draft.
4. Compress everything to WebP before writing
   it into the repository.
5. Never call upscale_image on an image that
   still has a background — cut it out first.

Cline questions

How do I add an MCP server to Cline?

Open the MCP Servers panel from the Cline toolbar and edit the server configuration. It writes a cline_mcp_settings.json file behind the scenes, so editing that file directly does the same thing.

Cline runs autonomously — how do I stop it burning credits?

Three things. Tell it in your instructions to call get_user_info before any batch, keep approval prompts on for tool calls while you are learning what it does, and check the request log afterwards. Every call is recorded with its exact cost.

Can Cline chain image tools without me approving each step?

That depends on your auto-approve settings, not on the server. A sensible middle ground is to auto-approve the cheap deterministic tools — compress_image, convert_format, get_user_info — and keep approval on for the ones that call a model.

Does it work with the local server instead of the hosted one?

Yes. Point Cline at a local node process with IMAGEMCP_API_KEY in its env instead of the remote endpoint. The tools are identical; only the transport changes.

What happens if a call fails mid-plan?

The tool returns a structured error rather than throwing, so Cline can read the message and adapt — most commonly by retrying with a different model or stopping to tell you the credit balance ran out.