DSHPlugins. EN Independent community

Krea AI MCP with DeepSeek Harness

Krea hosts an MCP server at https://api.krea.ai/mcp for image and video generation. DSH cannot run Krea's interactive consent screen, so the practical path is the API-token header — and that choice also decides which balance pays for the work.

Last checked 2026-10-13 · English-only guide · checked against the sources listed at the end

Short answer. Authenticate with an API token in a streamable-HTTP row: an Authorization: Bearer header, no interactive sign-in. The billing difference is real: OAuth connections bill compute units to the workspace chosen on the consent screen, while API-token connections bill that workspace's API balance. Either way generation costs money, so start with list_models, not generate.

Scope: DeepSeek Harness only. The row below is a Cordis patch for DSH's MCP client. Claude Desktop, Cursor, Windsurf and other clients use their own config format, so follow that service's own setup page for those — these steps do not apply there.

At a glance

Server
Krea MCP (Krea developer docs → Model Context Protocol server)
Endpoint
https://api.krea.ai/mcp
Transport
Remote MCP over streamable HTTP
Authentication
OAuth by default; for clients without OAuth, Authorization: Bearer KREA_API_TOKEN from krea.ai/app/api/tokens
Who can issue a token
Workspace owners and admins
Billing
OAuth → compute units on the workspace selected during consent; API token → that workspace's separate API balance
Tools
list_models, get_model_schema, generate, execute_node_app, get_job, cancel_job, get_upload_url
DSH fit
Direct — streamable-HTTP row with the bearer header
Last checked
2026-10-13

Step 1 — Issue an API token, and know which balance it spends

Krea's docs say you do not need an API key when the client can run OAuth; the consent screen lets you pick a workspace and charges that workspace's compute units. DSH cannot drive that consent flow, so mint an API token instead at krea.ai/app/api/tokens. Only workspace owners and admins can create one, and API-token usage is billed to the workspace's API balance — the same pool as direct API calls, not the compute units.

Krea's own guidance is to store API-token MCP credentials the way you store API keys, and never to commit an MCP configuration file that contains a live token.

Step 2 — Add the DSH row

# cordis.patch.yml — passed with: dsh web --patch ./cordis.patch.yml
- insert:
    - id: mcp-krea
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: krea
        transport: streamable-http
        url: https://api.krea.ai/mcp
        headers:
          Authorization: Bearer KREA_API_TOKEN

Put the real token only in the local file, keep that file out of version control, and restrict it (chmod 600). Tools then appear as mcp__krea__<tool>.

Step 3 — Discover models first, generate second

Model discovery is free and it is also how you avoid guessing at model ids. Ask for a list, then for one schema:

“List the available Krea image models, then show the input schema for Krea 2. Do not generate anything yet.”

When you are ready to spend credits, generation is asynchronous: generate submits a job and the output carries a job_id, which you poll with get_job and can abandon with cancel_job. execute_node_app runs Krea node apps and is likewise billable. get_upload_url returns a short-lived presigned URL for sending a local file as a model input.

Troubleshooting

401 or unauthorized
The bearer header is missing, malformed, or the token was revoked. Confirm the header reads Authorization: Bearer <token> and that the token still exists in the workspace's API tokens page. Only owners and admins can mint tokens.
The wrong workspace is paying
API tokens are workspace-scoped, so check which workspace the token belongs to. If you connected with OAuth somewhere else, the workspace picker on the consent screen is what binds that session, and switching later means disconnecting and reconnecting.
Upload-based tools fail
get_upload_url needs a bound workspace. For OAuth sessions Krea notes that upload tools only work once a workspace is bound; reconnect and pick the workspace if you are unsure which one your session uses.
Generation seems to hang
Submit, then poll get_job — do not read a slow submission as a failure. If the submission call itself times out, raise the client's per-call timeout.
A model id stops working
Model discovery returns deprecation metadata for retired models. Re-read the model list instead of pinning an old id from a tutorial.

Cost and third-party notes. We did not run a generation, so this page does not quote a price per call: check Krea's billing pages for current rates before wiring an agent that generates unattended. A community server, krea-mcp by keugenek, also exists as an independent Apache-2.0 project published as @vmosaic/krea-mcp-server over stdio; it is not maintained by Krea, and its feature set can lag the hosted server.

FAQ

Do I need an API key for Krea MCP?

Not if your client can complete the OAuth flow: Krea states you can add the server URL and sign in, with no API key required. DSH cannot run that interactive flow, so use the API-token header instead, which a workspace owner or admin creates at krea.ai/app/api/tokens.

Which balance pays for Krea MCP usage?

It depends on how you authenticate. OAuth deducts compute units from the workspace selected on the consent screen, while API-token connections are billed to that workspace's separate API balance.

How do I check availability without spending credits?

Ask the agent to call list_models and get_model_schema. Those are discovery calls. Reserve generate and execute_node_app for work you intend to pay for, and poll get_job rather than resubmitting.

Can I use a community Krea MCP server instead?

Yes, if you prefer a stdio server. krea-mcp by keugenek is an independent Apache-2.0 project published as @vmosaic/krea-mcp-server. It is not maintained by Krea, and the hosted endpoint documented here is the one Krea supports.

Sources

Every claim above was checked against these pages on 2026-10-13. Third-party material is labelled where it is used.

Related MCP guides

New to MCP in DSH? Start with the DSH MCP setup walkthrough, which covers transports, the patch file and tool naming for every server below.

Superhuman MCP

Superhuman Mail runs a remote MCP server at https://mcp.mail.superhuman.com/mcp. This guide covers what that server requires, why DSH cannot complete its interactive sign-in on its own, the stdio bridge Superhuman documents for desktop clients, and how to check the connection without letting an agent touch your inbox.

Stitch MCP Server

Google Stitch exposes an MCP server at https://stitch.googleapis.com/mcp. It is remote and header-authenticated, which makes it one of the simpler servers for DSH to call directly. This guide covers the API-key path, the Cordis row, the timeout that decides whether a slow generation looks like a broken connection, and the npm package you do not need.

Alpha Vantage MCP

Alpha Vantage maps its market-data endpoints onto MCP tools and publishes both a hosted endpoint and a local server. For DSH the local server is the documented route, because the hosted endpoint authenticates through a sign-in page — and the data tier, not the connection, is what limits you.

Verification status

These guides document configuration; they are not reports of a live run. Endpoints, prerequisites, tool names and billing rules were read from the sources listed above on 2026-10-13. We did not sign in to any of the four services, did not run a paid generation, and did not send mail, so treat every config row as a documented recipe to test in your own environment. If a service changes an endpoint or its plan terms, that operator's own page wins.