World

Release 0.1-alpha is here! Linux binary tested and works, Windows binary experimental.

Go here to download releases: https://git.skaldursgate.com/SkaldursGate/CanvasOrchestrator/releases

Installing

1. Extract the folder anywhere, e.g. ~/canvas-cli-standalone/.
2. chmod +x canvas-cli bin/inner/canvas-cli-real if the executable bit didn't survive transfer.
3. ./canvas-cli --help — should print the command list with no errors.

Nothing to install system-wide; the whole thing is self-contained inside the folder.

Running your first pipeline against the live cloud

Two terminals, both cd'd into the extracted folder.

Terminal 1 — start the relay (stays running in the foreground; this is your machine's own
outbound client, holding your credentials and keys, never the CLI):

export CANVAS_CLOUD_URL=https://canvas.skaldursgate.com
export CANVAS_RELAY_WORKSPACE_ROOT=~/canvas-workspace   # any directory you own; created below
export CANVAS_RELAY_PROJECT=<ask whoever gave you this binary what project name to use>
./canvas-cli relay start

Confirm it actually started: you should see canvas-relay: control surface listening on ....

Terminal 2 — log in, register, and run:

./canvas-cli login

This prints a verification URL and a code — open the URL in a browser and approve it there. Unlike
local dev, the live cloud does not auto-approve; you're authenticating as yourself, for real, and
nobody else's credential is involved. Confirm with:

./canvas-cli whoami
./canvas-cli worker register
./canvas-cli worker status

Now build the smallest possible pipeline — read one file, write another, nothing repo-specific:

mkdir -p ~/canvas-workspace
echo "hello from $(whoami)" > ~/canvas-workspace/input.txt

cat > ~/canvas-workspace/hello-pipeline.json << 'EOF'
{
  "contractVersion": 1,
  "id": "hello-world",
  "name": "Hello world",
  "description": "Smallest possible pipeline: read one file, write another.",
  "nodes": [
    { "id": "read", "type": "read-file", "typeVersion": 1, "label": "Read input",
      "config": { "path": "input.txt" } },
    { "id": "write", "type": "write-file", "typeVersion": 1, "label": "Write output",
      "config": { "path": "output.txt" } }
  ],
  "connections": [
    { "id": "c1", "from": { "node": "read", "port": "out" }, "to": { "node": "write" } }
  ]
}
EOF

./canvas-cli run start ~/canvas-workspace/hello-pipeline.json

That prints a run id. Watch it:

./canvas-cli run watch <run-id> --live

Once it reports completed, check the result:

cat ~/canvas-workspace/output.txt   # should read "hello from <you>"

These are some tools I'm using for lower usage and better accuracy:

• rtk-ai : Filters and compresses command-outputs AI have to go through. https://github.com/rtk-ai/rtk
• codegraph : Traces function-to-function calls in codebase, reduces exploration. https://github.com/colbymchenry/codegraph
• RAG : Retrieval Augmented Generation, used in documentation for chunk-based text-recognition, reduces exploration. https://www.ibm.com/think/topics/retrieval-augmented-generation
• headroom - Haven't tried yet : Middleware compression-layer, compacts prompts before arriving at LLM. https://github.com/headroomlabs-ai/headroom
• ggrun - Haven't tried yet: llama.cpp model optimization. https://github.com/raketenkater/ggrun

Vendors that are fully hosted in Europe, alternatives to American vendors for safer jurisdiction.

• webmail.SkaldursGate.com - Email-provider.
• git.SkaldursGate.com - Version control platform.
• forum.SkaldursGate.com - Forum platform.
• EuroAuth.com - IdP running OAuth and OIDC.