Add a keyless Parallel Search MCP example

[georgeatparallel]

All Submissions:

• Notify authors: @agentforce314, @ericleepi314

New Feature Submissions:

• Explained the change and why it is useful.

Description

Adds opt-in free, keyless Parallel Search MCP for ClawCodex. Merge the demo server into your existing user configuration and use your configured model provider. Fixes native agent startup unpacking the MCP loader's (configs, validation_errors) result and surfaces configuration notices. The helper fetches a URL returned by search. Defaults, permission prompts and project approval rules remain unchanged; anonymous access has rate limits.

I work at Parallel. For context: Artificial Analysis Search API benchmarks.

Tested: 95 focused regressions and native agent-server discovery, ToolSearch, explicit approvals, real anonymous search, result-selected fetch and fetched content in the next controlled-provider response. The fixture proves wiring, not production model quality. Offline denial and configuration-error regressions pass, as does clean runtime installation.

Full-suite limits: 10,878 tests pass, with the same 19 failures as the unchanged base (npm registry access and root/full-access assumptions). The targeted permission tests pass with IS_SANDBOX=1. Source mypy and broad Black/isort checks also fail on the base; this change adds no diagnostics or failing formatting paths.

[georgeatparallel]

Updated the native agent path: MCP startup now unpacks the loader's configuration/validation tuple and displays its notices. Added a regression through the real agent-server loader, HTTP transport, ToolSearch, permission round trips and query loop. Anonymous Parallel search returned a URL that the next tool call fetched, and fetched content reached the following controlled-provider response. The fixture proves wiring, not production model quality. Permission denial blocks execution, and pending interactive project servers remain excluded. The helper also selects its fetch URL from search results.

Validation: 95 focused regressions pass. The full suite has 10,878 passes and the same 19 failures as the unchanged base: one unavailable npm registry smoke test and 18 root/full-access assumptions. The 89 targeted permission tests pass with IS_SANDBOX=1. Broad mypy/Black/isort failures also reproduce on the base, with no new diagnostics or failing files. I work at Parallel.