Nidhogg
Log in

Nidhogg MCP for any AI

Plug Nidhogg into the agent you already use and create cinematic images and videos without leaving the conversation.

1
Copy your Nidhogg URL

Grab the endpoint below — you'll paste it in the next step.

2
Open Settings → Connectors

Add a custom connector, name it Nidhogg and paste the URL.

Nidhogg auth login
3
Connect and sign in

Click Add → Connect and sign in with your Nidhogg account — done. Ask Claude to generate an image.

|Github ↗

Prefer a static token? (header-auth clients)

Sign in to generate a token

Manual config (mcp.json)

{
  "mcpServers": {
    "nidhogg": {
      "url": "https://nidhogg.ai/api/mcp",
      "headers": { "Authorization": "Bearer dby_••••••••••••••••" }
    }
  }
}

Canvas as API

Every canvas flow doubles as a callable pipeline — three MCP tools drive it from any agent, shell or CI. Same wallet, same models, hard cost cap built in.

canvas_list

List your canvases — id, name and node count. Pick the canvasId the other two tools need.

canvas_flow_estimate

Dry-run quote: topological plan and per-node credit costs from the live catalog. Executes nothing, charges nothing.

canvas_flow_run

Runs the flow server-side with real billing. Requires a hard maxCredits cap and an idempotencyKey — no cap, no run.

Example: an image pipeline in CI

Wire prompt → LLM rewrite → image generation on a canvas once, then let your release workflow call it: estimate first, gate on the quote, then run with the build number as the idempotency key. A retried job replays the same report instead of paying twice — and the finished art lands in your gallery.

① Quote it first — nothing runs, nothing is charged

curl -s https://nidhogg.ai/api/mcp \
  -H "Authorization: Bearer dby_••••••••••••••••" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{
        "name":"canvas_flow_estimate",
        "arguments":{"canvasId":"YOUR_CANVAS_ID"}}}'

② Run it — hard cost cap + idempotency key

curl -s https://nidhogg.ai/api/mcp \
  -H "Authorization: Bearer dby_••••••••••••••••" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{
        "name":"canvas_flow_run",
        "arguments":{
          "canvasId":"YOUR_CANVAS_ID",
          "maxCredits":40,
          "idempotencyKey":"ci-'"$GITHUB_RUN_ID"'"}}}'

Cost safety

maxCredits is mandatory and checked before anything starts: if the quoted fixed-price total exceeds the cap — or any node has no catalog price — the run is rejected with zero spend. Token-billed LLM nodes are flagged in the quote and bounded by the server's per-turn cost cap.

Idempotency

One idempotencyKey = one run. Replaying the same key within 15 minutes returns the same report — a retried CI step can never start a second run. Each generation slot also derives a database-backed idempotency key, so even a crash-and-retry resumes the same jobs instead of re-charging.

One call is budgeted at up to ~55 s. Jobs still pending at the deadline keep running server-side — results land in your gallery and the canvas stays resumable.

Bridge: command the AI army on your machine

Route canvas Agent nodes to the claude / codex / gemini CLIs installed on your own computer — your subscription, your hardware, billed at ✦0.

1
Generate a token

Use the token button above to mint a dby_ personal access token. Shown once — store it safely.

2
Download the bridge

A single zero-dependency file. All it needs is Node ≥ 18:

3
Run it

While the bridge is online, Agent nodes grow a “Local engines” group. When it goes offline the group disappears — we never pretend it's connected.

How it works

Canvas dispatches a job
Queued in the cloud (10 min TTL)
Bridge long-polls and claims it
Local agent runs it and returns the result

Security

  • ▪Whitelisted commands only: the script can execute exactly three binaries — claude, codex, gemini. The server cannot make it run anything else.
  • ▪Prompts are data, not shell: subprocesses launch with argv arrays, never through a shell — any character in a prompt is just text.
  • ▪Leave anytime: Ctrl-C exits gracefully. The token lives only on your machine and can be revoked in settings.

Local engines on the canvas — ✦0

Local runs spend your own agent subscription and your own compute, so the platform charges nothing — the bridge path contains no billing code at all.

Frequently asked

We've answered the questions that come up most.

Nidhogg uses MCP (Model Context Protocol), an open standard that gives AI agents access to external tools.

YOUR AI STACK JUST LEVELED UP

Connect Nidhogg in 30 seconds

One server URL. Every model. Right inside the agent you already use.

Get started