Why

Not every model behind Claude Code understands images. Depending on the backend, an unsupported image request fails in one of two very different ways:

cvh targets the first case. It never touches the original image content block — it only appends a text description generated by a vision model you configure, via Claude Code's additionalContext hook output. That's the whole trick: no schema surgery, no risky content replacement, just an honest text description sitting next to the image the main model can't parse.

✅ Verified against a real Claude Code session. A model that silently ignores images was tested with and without cvh: without it, it guessed the wrong color for a test image; with cvh installed, the hook transcript shows additionalContext being injected and the model answering correctly. See CHANGELOG.md for details.

How it works

flowchart LR
    A["User pastes image /<br/>tool returns screenshot"] --> B{cvh enabled?}
    B -- no --> Z["passthrough, no-op"]
    B -- yes --> C["Extract image bytes<br/>(UserPromptSubmit / PostToolUse)"]
    C --> D{"Cached by<br/>content hash?"}
    D -- hit --> F
    D -- miss --> E["Ask your vision model<br/>(via AI SDK)"]
    E --> F["additionalContext<br/>injected into Claude Code"]

• UserPromptSubmit hook scans ~/.claude/image-cache/<session_id>/ for newly pasted images.
• PostToolUse hook (matcher: "*") runs a generic recursive extractor over any tool's tool_response — verified against three real, structurally different shapes: Read's discriminated object, MCP's content-block array, and Bash/PowerShell's flat isImage + data-URI string.
• Both hooks only ever output additionalContext — never updatedToolOutput. This keeps cvh simple and safe: no per-tool schema replication, no risk of a silently-dropped hook output because the replacement shape didn't match.
• Vision inference goes through the Vercel AI SDK, so you can point cvh at OpenAI (Chat Completions or Responses), Anthropic, or Gemini — or any OpenAI/Anthropic-compatible gateway.
• Descriptions are cached on disk by image content hash (~/.claude/cc-vision-hook/cache/), shared globally, with a 7-day lazy-expiring TTL — the same image is only ever described once, no matter where it came from.

Scope & limitations

cvh only helps with models that silently ignore images. If your model hard-rejects image input at the protocol level (the request itself fails), cvh cannot help — it never replaces or removes the original image content block, so the upstream will still see (and reject) it. Run cvh doctor or send a manual test request to figure out which category your model falls into before enabling cvh.

Other known limitations:

1. Hard-rejection models are out of scope (see above).
2. Support for "user pastes an image" relies on an unofficial implementation detail of Claude Code (~/.claude/image-cache/ directory scanning). This could break in a future Claude Code release; cvh does not attempt version detection or graceful degradation.
3. cvh does not auto-detect model capabilities. It's a plain enable/disable switch — you decide when to turn it on.
4. Claude Code only. No support for other agent hosts.
5. Images are sent to whichever third-party vision provider you configure. You are responsible for that data flow — review your provider's data handling policy if your screenshots may contain sensitive information.

Install

npm install -g cc-vision-hook

cvh install                            # create config + register hooks (idempotent)
cvh config set provider anthropic      # or: oai / responses / gemini
cvh config set model <your-vision-model>
cvh config set apiKey <your-api-key>
cvh doctor                             # sanity-check config + connectivity
cvh enable

Commands

Every command supports --json output for scripting.

Configuration

~/.claude/cc-vision-hook.json:

{
  "enabled": true,
  "provider": "oai",
  "model": "gpt-4o-mini",
  "baseUrl": "https://api.openai.com/v1",
  "apiKey": "sk-...",
  "timeoutMs": 45000,
  "maxTokens": 1200,
  "cache": { "ttlDays": 7 }
}

The API key is stored in plaintext; the file permission is automatically set to 0600 (owner read/write only).

Environment variable overrides (take priority over the config file): CVH_ENABLED / CVH_PROVIDER / CVH_MODEL / CVH_BASE_URL / CVH_API_KEY / CVH_TIMEOUT_MS / CVH_MAX_TOKENS.

Roadmap

• cvh init — interactive setup wizard.
• MCP tools (vision_ask, vision_describe_image, vision_describe_data_url) so an agent can follow up on a previously seen image. The on-disk cache already stores everything needed — the MCP surface just hasn't been wired up yet.

Troubleshooting

• Installed but nothing happens — Claude Code reads $HOME/.claude/settings.json (not $HOME itself). Make sure the path cvh install wrote to matches the one Claude Code actually loads. Run cvh status and confirm both hooks are true and enabled is true.
• doctor reports a connectivity failure — run cvh test-image <local-image> first to isolate whether the problem is in the vision-model call itself or in hook wiring.
• The model errors out entirely / the whole turn fails when an image is involved — you're using a hard-rejection model. cvh cannot help here; see Scope & limitations.

Development

bun install
bun run typecheck   # tsc --noEmit
bun run build       # emit dist/
bun test            # fixture-driven unit tests, no real network calls

For local/integration testing, override ~/.claude with the CVH_CLAUDE_HOME environment variable so you never touch your real Claude Code configuration:

export CVH_CLAUDE_HOME=/tmp/some-isolated-dir/.claude
node dist/cli.js install

When wiring this up against a real Claude Code session, set CVH_CLAUDE_HOME to $HOME/.claude (not $HOME itself), and also set HOME=<isolated dir> so the claude process reads the same settings.json.

Contributing

Issues and PRs are welcome. Please run bun run ci (typecheck + test + build) before submitting.

Contributions that add or change behavior should include a changeset:

bun run changeset

Releasing

Releases are published to npm via Trusted Publishing (OIDC, no long-lived tokens) whenever a v* tag is pushed. See docs/releasing.md for the full process and docs/release-checklist.md for the pre-flight checklist.

License

MIT