social-intel

A fixed two-call, read-only web-search bundle. It sends the same explicit query to Exa semantic search and Tavily advanced search, then returns two independent raw responses for the calling agent to compare and synthesize. The manifest does not merge, deduplicate, score, or summarize the results itself.

When To Use

Use it for requests such as:

• "Give me a sourced web brief on agent payments and cross-check two engines."
• "Compare what independent web searches surface about this product launch."
• "Research this brand and separate repeated reporting from single-engine finds."

Do not use it as a substitute for direct Reddit, X/Twitter, LinkedIn, Instagram, or TikTok data. It has no engagement metrics, account graph, social listening, or sentiment-measurement endpoint. It also cannot establish that something is "trending" without an explicit time window and appropriate platform evidence.

Rails

Both provider-facing payment challenges currently resolve to the same rail:

• Exa (api.exa.ai/search): the live probe resolves routed x402 via Circle Gateway (2026-10-04). It used to resolve routed MPP on Tempo. The provider challenge advertises several payment options. The free live quote is authoritative for the route SELAT will use.
• Tavily (x402.tavily.com/search): routed x402 via Circle Gateway.

Either way, you pay from whichever chain holds your funded Gateway balance.

Protocol, route, and price can change. Always run a fresh free live probe before approval; do not infer the active route from the hostname.

Workflow

1. Require one explicit topic. Add a date range or recency phrase to the query when freshness matters.
2. Validate locally: selat skill validate ./skills/social-intel
3. Obtain a fresh free quote: selat skill verify ./skills/social-intel --topic "<topic>" --live-probe
4. Show the user every step, route, expected total, and the sum of per-step caps. Obtain explicit approval before adding --pay or running the installed skill. Paid runs need an armed session budget (selat budget start --amount <approved-amount>, then selat budget stop).
5. Execute exactly once. The CLI runs both independently capped calls in order and may continue after one step fails.
6. Inspect every per-step success or failure and payment history before any retry. A paid application error may still have been charged.
7. Synthesize the raw results with claim-level source URLs, publication dates when present, and explicit uncertainty.

The fixed pipeline is:

1. Exa semantic web context — POST /search, ten results, with up to 4,000 characters of page text per result.
2. Tavily advanced web corroboration — POST /search, advanced depth, up to ten results.

Inputs And Outputs

Output: two independent raw JSON responses. Exa supplies ranked result metadata, URLs, and requested page text; Tavily supplies ranked result metadata, URLs, and snippets. The calling agent performs deduplication and synthesis.

A defensible brief should:

• cite the underlying page URL for each material claim, not the search engine;
• treat the same page found by both engines as one source, not two independent confirmations;
• label agreement between retrieval engines separately from independent factual corroboration by different primary or authoritative sources;
• preserve conflicts and credible single-engine evidence instead of discarding it merely because the other engine did not retrieve it;
• distinguish publication date from retrieval date and avoid freshness claims when dates are absent.

Gotchas

• This is a fixed two-call bundle. Both paid steps run; it is not a menu or cheapest-success waterfall.
• Both steps are POST. The query belongs in the JSON body.
• Required means required. There is no fallback topic; missing input must fail before any network call.
• Caps are not prices. Live verification on 2026-10-04 quoted $0.007 for Exa and $0.0105 for Tavily (expected total $0.0175). Step caps are $0.01 and $0.02, summing to $0.03. The top-level maxAmount is only a per-step fallback, not a run cap. Tavily's cap sits above the 0.016 second offer in its 402 challenge (see references/endpoints.md). Re-quote before every paid run.
• Free verification is limited. A successful 402 probe verifies the payment challenge, route, quote, and cap compatibility; it does not prove that the post-payment business response will succeed or contain useful evidence.
• Paid failures may charge. Never auto-retry. Inspect per-step output and payment history, obtain a fresh quote, and request new approval first.
• Search-engine agreement is not ground truth. Both engines can index the same page or repeat the same reporting.
• No direct social-platform coverage. Route platform-specific requests to an appropriate dedicated skill rather than describing web results as social data.

Validation

• Static: selat skill validate ./skills/social-intel
• Missing-input gate: selat skill verify ./skills/social-intel --live-probe must fail before probing either endpoint.
• Live gate (free): selat skill verify ./skills/social-intel --topic "stablecoin payments for AI agents October 2026" --live-probe
• Single-step probes (free): see references/endpoints.md.
• Paid confirmation: only after showing a fresh quote and receiving explicit approval, add --pay to the verified command. Never reuse approval after a route, price, input, or cap change.

References

• manifest.json — the machine-readable payment recipe this skill runs.
• references/endpoints.md — the catalogue endpoints, rails, and live prices.
• ../../references/agent-skill-authoring-sop.md — authoring standard.
• selat-pay — https://github.com/SELAT-AI/selat-pay

Exa and Tavily are third-party services. Their names and trademarks belong to their respective owners and are used only to identify the endpoints called.