[georgeatparallel]

Summary

Adds Parallel Search to the bundled integrations so users can run openfang add parallel for web search and page extraction without a Parallel account or API key.

OpenFang's default auto path tries Tavily first when its key is configured, with basic search depth. That's the comparison I focused on. With no search providers configured, it falls back to DuckDuckGo, which Artificial Analysis doesn't benchmark.

In Artificial Analysis's August 31 results (select all providers to see Tavily), Parallel advanced scores 77% on BrowseComp versus Tavily basic's 59%, an 18-point lead. Its overall Search Index is 75 versus 66. Parallel fast also scores 73 overall, with $8.41 in search costs per 1,000 benchmark tasks versus Tavily's $126, about 93% less.

For setups where auto currently picks Tavily, I'd consider evaluating Parallel as a replacement for that first choice if better research answers are the goal. These are Search API results in Artificial Analysis's harness, not measurements of this free MCP preset or OpenFang. This PR leaves the auto routing unchanged.

I work at Parallel, which operates this service.

Changes

• Add a keyless Streamable HTTP preset using the existing registry, installer and MCP client.
• Include setup, removal, agent access and data-sharing guidance in the template and MCP guide. Once enabled, agents with access may send queries, URLs and supplied context or metadata to Parallel. Free access is rate limited.
• Preserve the built-in search provider and existing permission controls. Add a regression for installation without credentials, saved configuration, disabling and removal; update catalog checks for the new entry.

Testing

• cargo +1.95.0 fmt --all --check and git diff --check pass.
• Parsed all 26 bundled TOML files and checked unique IDs, category counts, registration and top-level setup instructions.
• Live HTTP client check with the official OpenFang v0.6.9 macOS release, verified against its published SHA-256. An isolated temporary configuration with the preset's name and transport, no credentials and empty auth headers discovered both tools. Calls through OpenFang's /mcp tool runner returned 10 search results and one fetched documentation page with no per-URL errors.
• cargo clippy --workspace --all-targets -- -D warnings passes
• cargo test --workspace passes
• New compiled bundle and installer regression executed

No repository build or Rust tests were run locally. The prebuilt check exercises the released configuration loader and MCP client, not the new compile-time catalog entry; compilation and the checked-in installer regression still need CI.

Security

• No new unsafe code
• No secrets or API keys in diff
• Uses existing configuration and tool input validation boundaries