Help me set up Canonry for my public site.
Use the official Canonry docs:
Agent quickstart: https://github.com/Canonry/canonry#or-use-any-shell-capable-coding-agent
CLI reference: https://github.com/Canonry/canonry/blob/main/skills/canonry/references/canonry-cli.md
Plugin setup: https://github.com/Canonry/canonry/blob/main/docs/plugins.md
MCP setup: https://github.com/Canonry/canonry/blob/main/docs/mcp.md
Pick one path and stay on it. Use connected Canonry tools (plugin or MCP) only if you have them and I have not pointed you at a specific install. A connected tool runs in its own process and never sees CANONRY_CONFIG_DIR, CANONRY_PORT, or anything else you export in a shell, so it always acts on my default install. If I named a config directory, a port, or a sandbox, that is the shell path: use cnry and tell me which path you chose. Never mix the two in one run, and never create a duplicate project. The cnry and canonry commands are interchangeable.

Ask for my public domain, country, and language. Do not create or scan anything yet. Country and language are only applied when a project is created, so if you reuse an existing project, report its country and language instead of changing them.
Shell path only: confirm cnry is on PATH, then run cnry --version. If Canonry is missing, propose npm install -g @canonry/canonry and wait for approval. Then run cnry doctor --format json, which is the command that says whether config, database, and server are in place. cnry --version does not read config and succeeds on a completely unconfigured install, so it cannot answer this. If config is missing, run cnry bootstrap yourself: it is not interactive, takes about a second, and is safe to rerun. Do not hand it to me and wait. The interactive command is cnry init, which is optional provider and OAuth setup that Page Health does not need. Bootstrap prints an API key, so do not repeat its output back to me, and never ask me to paste passwords, API keys, OAuth credentials, or command output.
Confirm the API is reachable. cnry doctor --format json reports it, and any project read exits non-zero with CONNECTION_ERROR when it is not. If it is unreachable, propose cnry start and wait for approval. Use cnry start, which is the background daemon, and not cnry serve, which runs in the foreground and will block you until I stop it, even though some error messages suggest it. Stop anything you started with cnry stop.
List projects with the connected project tool or cnry project list --format json, and reuse one whose domain matches. Confirm the proposed name is not already assigned to a different domain. To find out whether a project has already been scanned, run cnry technical-aeo score <project> --format json with no --run-id, which reports the latest run. Read the hasData field, not the score: this command exits 0 and reports aggregateScore: 0 for a project that has never been scanned, so reading the score alone would have you tell me my site scored zero. If hasData is true and runStatus is completed or partial, read that scan instead of starting a new one. If no project matches, show the exact create operation and wait for approval.
Propose a bounded Site Health scan: --max-pages 100 for a first look, plus whether dead-link checking is on (it is off unless you pass --check-dead-links). Show the connected operation or the exact cnry technical-aeo run <project> --max-pages 100 --wait --format json command with the project name filled in, and wait for separate approval before scanning. --wait polls for up to 15 minutes and returns only the run id and status. If the status it returns is still queued or running, the scan has not finished: do not rerun it or report results, poll cnry technical-aeo progress <project> --run-id <run-id> --format json until it is terminal. If --wait outruns your own tool timeout first, recover the run id with cnry technical-aeo score <project> --format json and poll the same way.
When the run is completed or partial, read cnry technical-aeo crawl <project> --run-id <run-id> --format json first. It is the only one of these commands that carries termination and complete; the score and pages commands do not. Then read cnry technical-aeo score <project> --run-id <run-id> --format json and cnry technical-aeo pages <project> --run-id <run-id> --sort score-asc --limit 10 --format jsonl. Tell me the termination reason in plain words and whether the scan covered the whole site or stopped at a page, link, depth, or time limit. A partial run scored the pages it reached and not my site, so never present it as a full-site result. If it stopped early, the fix depends on the reason, so say which: a page or depth limit needs a larger budget, a time limit needs a smaller scan (a lower --max-pages or --max-depth). If the run failed or was cancelled, inspect the run error and stop.
Summarize only completed evidence, then propose AI Visibility setup. Ask before you add queries, connect providers, start a provider-backed or quota-consuming run, edit files, or publish.

Keep cnry serve active. In a second terminal, create a project and start a bounded scan:

cnry project create my-site --domain example.com --country US --language en
cnry technical-aeo run my-site --max-pages 100 --wait --format json

Read the run ID and status from the output. If the status is completed or partial, read evidence from that run:

cnry technical-aeo score my-site --run-id <run-id> --format json
cnry technical-aeo pages my-site --run-id <run-id> --sort score-asc --limit 10 --format jsonl

--wait polls for up to 15 minutes. If the scan remains active, use cnry technical-aeo progress --run-id --format json. If the scan fails or is cancelled, read the error with cnry run show --format json.

For an existing project, configure a provider with your own key.

| Provider | Key source | Environment variable | |---|---|---| | Gemini | Google AI Studio | GEMINI_API_KEY | | OpenAI | OpenAI Platform | OPENAI_API_KEY | | Meta (Muse) | Meta Model API | MUSE_API_KEY | | Claude | Anthropic Console | ANTHROPIC_API_KEY | | Perplexity | Perplexity settings | PERPLEXITY_API_KEY | | OpenRouter | OpenRouter keys | OPENROUTER_API_KEY | | Exa | Exa dashboard | EXA_API_KEY | | Parallel | Parallel platform | PARALLEL_API_KEY | | Local model | Any OpenAI-compatible endpoint | LOCAL_BASE_URL |

Add the queries that matter to your project:

cnry query add my-site "your first query" "your second query"
cnry run my-site --wait
cnry visibility-stats my-site --by-provider

Canonry supports direct model APIs, OpenRouter, and search answer services such as Exa and Parallel. Direct adapters retain the search and grounding details each provider exposes. OpenRouter offers multiple models through one key. Exa and Parallel search the web and return answers with source evidence.

Each direct engine, OpenRouter model route, and service profile keeps separate results and history. The same model through different APIs can return different answers and citation evidence. Canonry measures brand mentions separately from source citations and keeps missing evidence unknown.

Provider guides describe the evidence and limits of each integration.

Expand your tracked questions. Run cnry discover run --icp "...". This does not change the tracked questions. Preview a completed session with cnry discover promote preview . Promote only after approval.

Research without changing tracking. Run cnry research run "query one" "query two" --wait. For a configured portfolio destination, select one --market or --property ; Canonry saves the destination with the answers without changing the query. Add paired --template-id --template-version only to preserve an already-expanded template's provenance. Research never adds queries to the tracked questions.

| Problem | Next step | |---|---| | Site scan is still active | Read exact counters with cnry technical-aeo progress --run-id --format json. | | Site scan failed | Read the error with cnry run show --format json. Read the last phase and counters with the progress command. | | No visibility results | Inspect existing work with cnry runs --format json, then cnry run show --format json. This does not start another paid run. | | npm install fails on node-gyp | Install build tools for better-sqlite3 (guide). |