Exa Agent Research Skill
Multi-step web research, list-building, enrichment, and structured outputs via Exa Agent.
Evidence shows the skill requires user API key or OAuth, and the skill documentation explicitly suggests using previousRunId for continuation, but does not disclose data flow. It encourages using evidence URLs and confidence fields but does not require explicit user confirmation or restrict permissions. Given dependence on external service (Exa) and unverified publisher, deductions include lack of explicit user confirmation mechanism and data flow disclosure.
Instructions are consistent, but no executable tests exist to validate the skill path. The documentation provides clear error-handling steps but no actual run verification. Therefore, deduction for lack of reproducible verification and edge-case testing.
Skill clearly targets Exa Agent usage with clear scenarios, but boundaries and non-fit cases are not explicit. For mainland China users, Exa service may be unreachable, limiting applicability. Thus deduction.
Documentation is well-structured with schema examples and decision tree, but lacks explicit disclosure of known limitations (e.g., network reachability) and no version history or changelog. Deduction.
Core task is well-defined, but no actual execution evidence to verify output is directly usable. Documentation claims structured and grounding outputs but no concrete examples. Hence deduction.
CI workflow and tests exist, but tests mainly cover MCP server functionality, not the skill path itself. The agent_run tool mentioned in skill file has no corresponding test coverage. Therefore deduction.
- This skill depends on the external Exa service, which may be inaccessible from mainland China; verify network reachability before use.
- The Exa Agent tool requires user API key or OAuth; ensure keys are securely stored to avoid leakage.
- The skill documentation does not explicitly disclose data flow or require user confirmation; be mindful of data privacy when using.
What does this skill do, and when should you use it?
This skill leverages the Exa Agent through the Exa MCP server to perform complex web research tasks, including list-building, company/person discovery, enrichment, and structured output. It provides a clear decision framework for choosing between scripted batch processing and Agent-driven research, along with detailed schema guidance, coverage validation, and continuation patterns. The skill integrates with multiple Exa Connect data providers (e.g., Similarweb, financial_datasets) for enriched results.
Invokes the agent_run tool to execute multi-step research. It defines an objective, universe, segments, and outputSchema before creating a run. It uses input.data for known rows, input.exclusion to avoid prior results, and previousRunId for continuations. It reads structured and grounding outputs, validates coverage, and generates follow-up runs when needed. For batch enrichment tasks, it guides writing deterministic scripts with bounded concurrency, exponential backoff, and checkpoints.
- A market researcher needs to build a list of companies meeting specific criteria, with evidence URLs and confidence ratings.
- An investment analyst wants to enrich a list of candidate companies with website traffic estimates from Similarweb.
- A recruiter uses the fiber provider to find contact information for people at B2B companies.
- A researcher seeks to validate coverage of a literature review by segmenting searches and excluding previously found results.
- A product manager uses multi-hop research to build a prospect list and tracks progress with structured schemas.
What are this skill's strengths and limitations?
- Tailored for multi-step, open-ended research tasks beyond simple search.
- Emphasizes structured output (outputSchema) and evidence tracking.
- Supports continuation and exclusions for iterative research.
- Integrates multiple data providers (fiber, similarweb, etc.) for enriched results.
- Clearly separates scripted batch vs. Agent-driven approaches to avoid context waste.
- Requires Exa MCP server and API key; not standalone.
- Depends on external service, which may incur costs or rate limits.
- Guidance may be overkill for simple scripted tasks.
- ZDR customers face streaming limitations, possibly unsuitable for very long runs.
- No official test suite; reliability is unverified.
How do you install this skill?
Place the skill folder (containing SKILL.md) into your agent skills directory (e.g., ~/.claude/skills/ for Claude Code). Configure the Exa MCP server with endpoint https://mcp.exa.ai/mcp and an Exa API key. Enable the agent_run tool by adding tools=agent_tools to the MCP URL.
How do you use this skill?
For instance, prompt the agent: 'Find 10 fast-growing B2B SaaS companies and return monthly visits using Similarweb data.' The skill guides you to define objective, universe, segments, and outputSchema, then call agent_run. If the run returns running, poll with the runId until outputReady is true. Then validate coverage and continue as needed.