How long does the integration take?

Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.

Will this lock us into another vendor?

No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.

What happens when a provider fails?

The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

vision - Prompt - OpenSmartRoute

Name: vision
Author: MichelKerkmeester

Prompt file imported from MichelKerkmeester/skilled-agent-harness_spec-driven-loops (.cursor/commands/vision.md). Fill in {{arguments}} before use. Copyright stays with the author.

/vision

1. PURPOSE

Read an image locally and answer from the returned evidence. The analysis runs on-device through a local model. Nothing is uploaded and no API key is involved.

2. OVERVIEW

Cursor reaches sk-vision by running its CLI, not through a tool. Cursor delivers no prompt-time hook event, so there is no way to place the evidence in front of you automatically the way the OpenCode and Devin adapters do. Running the command is what produces the evidence.

The CLI prints a <SK-VISION EVIDENCE> block containing scene analysis, a caption and exact OCR, or an answer to a question when you pass one.

3. CONTRACT

Input: {{arguments}} — an image path, optionally followed by a question about it.

Output: A concise answer synthesized from the <SK-VISION EVIDENCE> block. Treat OCR as exact evidence. Treat scene and caption as interpretation.

Command:

node .skilled/skills/sk-vision/vision-runtime/dist/vision-cli.js <image-path> [question...]

Exit codes: 0 evidence on stdout, 1 a runtime or usage error on stderr, 2 no image found.

4. INSTRUCTIONS

Extract the image path from {{arguments}}. Ask for one if the user named no path, since the CLI cannot see the conversation.
Run the command above with that path, appending the user's question when they asked one.
Answer from the <SK-VISION EVIDENCE> block. Synthesize it rather than pasting the raw block back.
Treat OCR as exact and verbatim, especially error strings, codes, URLs and UI copy. Hedge claims the scene and caption cannot fully support.
Never invent detail the analysis does not contain.

5. HARD RULES

If the command exits 1, report the SK_VISION_ERROR line in one sentence and stop. Do not fabricate an analysis.
If it exits 2, tell the user the path did not resolve and ask for a correct one.
The first run downloads model weights, roughly 3.9 GB, and takes minutes. Say so rather than letting the user think it hung.
Never claim to see anything the block does not contain.

User request: {{arguments}}

Install vision and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

# /vision ## 1. PURPOSE Read an image locally and answer from the returned evidence. The analysis runs on-device through a local model. Nothing is uploaded and no API key is involved. --- ## 2. OVERVIEW Cursor reaches sk-vision by running its CLI, not through a tool. Cursor delivers no prompt-time hook event, so there is no way to place the evidence in front of you automatically the way the OpenCode and Devin adapters do. Running the command is what produces the evidence. The CLI prints a `<SK-VISION EVIDENCE>` block containing scene analysis, a caption and exact OCR, or an answer to a question when you pass one. --- ## 3. CONTRACT **Input:** `{{arguments}}` — an image path, optionally followed by a question about it. **Output:** A concise answer synthesized from the `<SK-VISION EVIDENCE>` block. Treat OCR as exact evidence. Treat scene and caption as interpretation. **Command:** ```bash node .skilled/skills/sk-vision/vision-runtime/dist/vision-cli.js <image-path> [question...] ``` Exit codes: `0` evidence on stdout, `1` a runtime or usage error on stderr, `2` no image found. --- ## 4. INSTRUCTIONS 1. Extract the image path from `{{arguments}}`. Ask for one if the user named no path, since the CLI cannot see the conversation. 2. Run the command above with that path, appending the user's question when they asked one. 3. Answer from the `<SK-VISION EVIDENCE>` block. Synthesize it rather than pasting the raw block back. 4. Treat OCR as exact and verbatim, especially error strings, codes, URLs and UI copy. Hedge claims the scene and caption cannot fully support. 5. Never invent detail the analysis does not contain. --- ## 5. HARD RULES - If the command exits `1`, report the `SK_VISION_ERROR` line in one sentence and stop. Do not fabricate an analysis. - If it exits `2`, tell the user the path did not resolve and ask for a correct one. - The first run downloads model weights, roughly 3.9 GB, and takes minutes. Say so rather than letting the user think it hung. - Never claim to see anything the block does not contain. User request: {{arguments}}

{ "prompt": "# /vision\n\n## 1. PURPOSE\n\nRead an image locally and answer from the returned evidence. The analysis runs on-device through a local model. Nothing is uploaded and no API key is involved.\n\n---\n\n## 2. OVERVIEW\n\nCursor reaches sk-vision by running its CLI, not through a tool. Cursor delivers no prompt-time hook event, so there is no way to place the evidence in front of you automatically the way the OpenCode and Devin adapters do. Running the command is what produces the evidence.\n\nThe CLI prints a `<SK-VISION EVIDENCE>` block containing scene analysis, a caption and exact OCR, or an answer to a question when you pass one.\n\n---\n\n## 3. CONTRACT\n\n**Input:** `{{arguments}}` — an image path, optionally followed by a question about it.\n\n**Output:** A concise answer synthesized from the `<SK-VISION EVIDENCE>` block. Treat OCR as exact evidence. Treat scene and caption as interpretation.\n\n**Command:**\n\n```bash\nnode .skilled/skills/sk-vision/vision-runtime/dist/vision-cli.js <image-path> [question...]\n```\n\nExit codes: `0` evidence on stdout, `1` a runtime or usage error on stderr, `2` no image found.\n\n---\n\n## 4. INSTRUCTIONS\n\n1. Extract the image path from `{{arguments}}`. Ask for one if the user named no path, since the CLI cannot see the conversation.\n2. Run the command above with that path, appending the user's question when they asked one.\n3. Answer from the `<SK-VISION EVIDENCE>` block. Synthesize it rather than pasting the raw block back.\n4. Treat OCR as exact and verbatim, especially error strings, codes, URLs and UI copy. Hedge claims the scene and caption cannot fully support.\n5. Never invent detail the analysis does not contain.\n\n---\n\n## 5. HARD RULES\n\n- If the command exits `1`, report the `SK_VISION_ERROR` line in one sentence and stop. Do not fabricate an analysis.\n- If it exits `2`, tell the user the path did not resolve and ask for a correct one.\n- The first run downloads model weights, roughly 3.9 GB, and takes minutes. Say so rather than letting the user think it hung.\n- Never claim to see anything the block does not contain.\n\nUser request: {{arguments}}", "variables": [ "arguments" ], "notes": "On-device vision on an image, run locally through the sk-vision CLI. Tools: Read, Bash. Arguments: <image path> [question] — omit the question for a full read.", "metadata": { "source": { "provider": "github-cursor-commands", "repository": "https://github.com/MichelKerkmeester/skilled-agent-harness_spec-driven-loops", "path": ".cursor/commands/vision.md", "ref": "c25e48495ad1ea93b91e9000b748c87eacd359b4", "url": "https://github.com/MichelKerkmeester/skilled-agent-harness_spec-driven-loops/blob/c25e48495ad1ea93b91e9000b748c87eacd359b4/.cursor/commands/vision.md", "key": "MichelKerkmeester/skilled-agent-harness_spec-driven-loops/.cursor/commands/vision.md" } } }

vision

About

/vision

1. PURPOSE

2. OVERVIEW

3. CONTRACT

4. INSTRUCTIONS

5. HARD RULES

Install vision and it becomes one more candidate the router can pick - when it fits.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Use it

Manifest

Reviews