Connect your agent to Research Packs.

Connect once. Keep the research state.

1. Add EquityLayer as the research layer.

   claude mcp add --transport http equitylayer https://equitylayer.io/mcp

   Copy the command

   Codex, Hermes, and other MCP clients: point them at https://equitylayer.io/mcp (HTTP transport). It lists the Pack loop and the embedded Investment Research Harness. Start with list_registry. The full tool catalog at /mcp/full belongs to this same service. Keep one connection.

2. Give your agent the setup prompt.

   Set up EquityLayer as the fundamental research layer for this workspace.
   
   Use this remote MCP server:
   - Name: equitylayer
   - URL: https://equitylayer.io/mcp
   - Transport: HTTP
   
   If you can add MCP servers, add this server now. Ask for approval if this client requires it.
   If you cannot add MCP servers, show the exact command or configuration for this client. Then wait for me.
   
   After the MCP connection works, guide me through a short setup. How to guide me:
   - Ask one question at a time. Offer at most four numbered options, mark one "(recommended)" with a one-line reason, and let me answer in my own words instead. If this client has a question tool with options, use it.
   - Before each question, say in one or two plain sentences what it decides. Explain a term the first time it appears, for example CPO, segment revenue, wrong_if, baseline or withheld.
   - Answer in my language. Keep source values, dates and identifiers as returned.
   - Remember my answers in this workspace (sector, whether I had a view, how often to check, language), so next time you skip what you already know.
   
   1. Start from the EquityLayer Research Registry. Call list_registry.
      If it lists entries, show each entry's title, status, and freshness.next_check_at, and offer "Follow a published Research Pack" first (steps 11 to 14). Ask one question if the scope is unclear.
      If it lists no entry, say so and offer to track my own thesis (steps 2 to 10).
   2. Ask which sector I want to research:
      1. CPO and optical networking (recommended): live readings.
      2. Space infrastructure: variables are being added; today I would mostly see gaps.
      3. Custom AI chips (ASIC): live readings, tracking only, no Research Pack.
      4. Something else: for example Physical AI (variables listed, most readings withheld) or my own topic; plan_thesis_tracking says what the catalog cannot cover.
      Pass theme_id when one fits: optical_cpo, space_infrastructure, custom_asic, physical_ai.
   3. Ask whether I already have a view:
      1. I have a thesis: I write it in my own words.
      2. Help me form one: ask me up to three short questions about the sector (what happens, by when, who shows it first). Turn my answers into one sentence and let me edit it. Never write a view for me or present one as EquityLayer's.
   4. Ask which companies or terms I mean. Offer up to four that fit the sector, and accept others.
   5. Call plan_thesis_tracking. Show the candidate variables in plain words: what each shows, what it cannot show (its caveat), and how often it updates. Show the gaps. Use only listed ids. Recommend three to five.
      Call get_variable on each variable I choose, to see its unit and recent level. A withheld value is no data; never guess it.
   6. Propose thresholds around the latest levels that cover the whole range, and say they are my judgment. I can change them. A link with no rules decides nothing.
   7. Explain wrong_if: what would make me admit I am wrong. Suggest one numeric condition counted from today, and one that only I can judge.
   8. Show hypothesis, why, tracking and wrong_if, and wait for my yes. Then call check_thesis_harness. Report each link's status and reason, any wrong_if that fired, and everything in needs_human.
   9. Save the harness and next_baseline in this workspace once I have read the result. Ask how often to check again: 1. every day at 09:00 my time (recommended), 2. every Monday, 3. only when I ask. If this client can schedule tasks, schedule it and tell me whether scheduling worked. Next time, pass the saved next_baseline and report only what changed.
   10. Close in three lines: what is saved and where, when the next check runs, and that you never tell me what to buy, sell, size or time.
   
   To follow a published Research Pack (from step 1):
   11. Read the entry I choose with its read call. Then call get_pack_scorecard and tell me how its claims resolved and whether it is fresh.
   12. If freshness is stale, say that no newer verified release exists yet. Never present that as "nothing changed".
       An entry with status in_review cannot be read or followed yet. Tell me its title and check list_registry again later.
   13. Follow the entry with its follow call (create_research_follow). Save follow_card and caller_action.next_check_at in this workspace.
   14. If this client can schedule tasks, schedule check_research_follow with the saved card for next_check_at. Otherwise tell me the date so I can ask you then.
       If an entry's access is "entitlement", its claims need access I have bought; say so instead of guessing them.
   
   Always:
   - Do not add a source or access private data without my permission.
   - Never send private or portfolio data to EquityLayer.
   - The full thesis loop is at https://equitylayer.io/skills/thesis-harness.md.
   
   All of these tools are listed on the same /mcp connection.
   https://equitylayer.io/mcp/full lists every tool of the same EquityLayer server. It is not a separate product or a second connection.
   
   About EquityLayer:
   Fundamental research layer for AI agents. Two ways to use it: track the user's own thesis, or follow a Research Pack EquityLayer publishes.
   Connect once at https://equitylayer.io/mcp with HTTP transport.
   The same service embeds EquityLayer's Investment Research Harness.
   /mcp/full lists all tools from this service. It does not require a second connection.
   For a Research Pack, start with list_registry. Use each entry's returned read, scorecard, and follow calls.
   Current product industries: CPO, Physical AI, and space infrastructure. Publication requires human approval.
   Coverage today for the user's own thesis: CPO and custom ASIC variables have readings; custom ASIC is tracking coverage only, with no Research Pack. Space variables are being added. Physical AI variables are listed, but most readings are withheld pending a source-rights review.
   A thesis outside these industries is accepted; plan_thesis_tracking reports what the catalog cannot cover.
   An empty registry does not establish coverage. Explain the gap.
   Read the years-long regime, the falsifiable hypothesis, and the company inclusion and exclusion framework.
   Read how value transmits, where the judgment lands, and the cycle-level wrong_if conditions.
   Keep hypothesis, why, tracking, wrong_if, and exposure connected to their evidence.
   Revisit the same framework: event → insight → track.
   Read each event against registered thesis variables and the Pack's watches.
   Report the registered reading and evidence limits. A person reviews the insight.
   A long Pack holds the framework. A short update records a delta of the same Pack.
   Use its revisions, editions, and Follow Card. Do not create a separate event pipeline.
   A one-day move, one contrary quarter, or expired fresh_until does not change the thesis state.
   Review agreement across subjective and quantitative signals. Do not compute a new cycle score.
   Only the Research Lead changes the published thesis state, with a recorded reason.
   Save the complete Follow Card locally. Check it on the returned next_check_at.
   When review is required, preserve the old card. Follow caller_action after review.
   Schedule checks only when the client supports scheduling. Report whether scheduling succeeded.
   For the user's own thesis, use plan_thesis_tracking, then get_variable on each chosen variable, then check_thesis_harness.
   A careful model or a person selects variables. The caller keeps the harness and baseline.
   Private research does not become an approved EquityLayer Pack.
   Preserve source rights and exact revision grants. Do not reconstruct withheld evidence.
   tw-monthly-revenue is historical compatibility data. It is not a current thesis or launch sample.
   The PCB Pack remains withdrawn. Do not use it as a current product.
   Exposure describes operating effects. Do not provide trade recommendations.
   
   Evidence and timing:
   - reported_fact: A primary source reports this fact directly.
   - management_claim: Management states this claim, which requires independent verification.
   - inference: The researcher draws this conclusion from evidence and states the assumptions.
   - unknown: Available evidence does not establish this point.
   - next_check_at: This timestamp specifies a proposed next check after creation and does not confirm a scheduled task.
   
   Do not ask for email, account, broker, wallet, or execution access.
   If the MCP is unavailable, say so and do not fabricate results.

   Copy the onboarding prompt
3. Verify the connection, then start research.

   Ask your agent to call list_registry. A copied URL does not confirm a successful connection.

   Your agent uses approved sources and the financial connectors already in its workspace. It returns source-linked changes, affected companies, and the next check.

Machine-readable index: /llms.txt. Everything below is the reference contract behind the Pack and harness tools.

Models change. Research contracts should persist.

A useful capability has a bounded job, explicit inputs, structured outputs, source provenance, a data-as-of timestamp, known limitations, and a maintainer. MCP may be the interface, but the managed research service is the product.

Every entry in the source index is published against this contract, and every research note shows how sources become dated, testable claims.

Suggested output envelope

{
  "question": "What changed since the prior quarter?",
  "data_as_of": "ISO-8601 timestamp",
  "facts": [{ "claim": "...", "source": "..." }],
  "inferences": [{ "claim": "...", "confidence": 0.72 }],
  "assumptions": ["..."],
  "risks": ["..."],
  "invalidation_conditions": ["..."],
  "capability_version": "semis-earnings@0.3.1"
}

Trust should be inspectable.

The first version should avoid one composite “accuracy” score. Show separate signals for source coverage, point-in-time integrity, schema reliability, latency, cost, maintenance recency, benchmark task results, and user reuse.

• Never infer commercial rights from “open source” alone.
• Separate data-provider terms from repository license.
• Version reputation by capability release, not only creator.
• Evaluate bounded tasks before subjective full reports.