@open-neko/plugin-parallel-search

Web search + page fetch for OpenNeko, backed by Parallel.ai's Search MCP.

The plugin connects to https://search.parallel.ai/mcp over Streamable HTTP and exposes two action kinds.

Install

# From the official OpenNeko marketplace (verified integrity hash):
openneko install @open-neko/plugin-parallel-search

# Or, bypassing every marketplace (e.g. while testing a local build):
openneko install @open-neko/plugin-parallel-search --unverified

The anonymous tier works without any API key — try it first. To use the authenticated tier (higher rate limits), set your key in either:

• the manifest env: "env": { "PARALLEL_API_KEY": "..." } (env-injection ships in v2)
• the per-call payload: payload.api_key (works today; key is stored in action_request.payload)

Actions

web_search

| Field | Value | |---|---| | Scope | external | | Payload | { query: string, api_key?: string, mcp_url?: string } | | Result | { text: string, bytes: number } — concatenated text content from the MCP tool's content[] blocks (~25 KB cap upstream) |

mcp_url defaults to https://search.parallel.ai/mcp. Override to https://search.parallel.ai/mcp-oauth if your operator policy requires OAuth.

web_fetch

| Field | Value | |---|---| | Scope | external | | Payload | { url: string, api_key?: string, mcp_url?: string } | | Result | { text: string, bytes: number } — markdown for the URL, token-efficient per upstream docs |

Capabilities (manifest)

network:
  - search.parallel.ai

The OpenNeko loader translates this declaration into the plugin VM's network policy. Any attempt by the plugin to reach a different host is blocked at the VM boundary.

Local development

pnpm install
pnpm test
pnpm build           # → dist/run.js (bundled, ~980 KB; bundles @modelcontextprotocol/sdk)
node dist/run.js register '{}'

License

Apache-2.0