Kepler MCP server icon

Kepler

by Kepler

HIPAA CompliantSOC2 ReadyISO 27001 Ready
Productivity9 tools

Run citation-backed research on public company financials from an AI agent, and export the resulting Excel models as CSV. Nine tools, all confined to your own research runs. OAuth 2.1 sign-in, identity-only scopes.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect Kepler via MCP

https://mcp.kepler.ai/api/mcp/v1

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

Kepler Tools & Capabilities (9)

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • No application scopes exist. Kepler's RFC 9728 descriptor and its authorization server both advertise exactly four scopes — openid, profile, email, offline_access — every one an OpenID identity claim. Nothing at the OAuth grant separates looking up a company from starting a research run. The boundary that matters here does not fall between the cheap call and the expensive one; it does not exist at all.
  • A run's cost is unknowable from inside the surface. Kepler publishes no per-run rate, and scanning all nine tool names for credit, balance, usage, quota and billing returns nothing. An agent can start an arbitrary number of runs and has no tool that could check what one costs first. Kepler's own docs say a run takes minutes because it reads primary sources, which is real compute. Do not blanket-approve run_financial_research and continue_research the way you might approve a read tool; approve them per call until you know what your account is charged.
  • No idempotency key is documented. Neither write tool documents one, and both return a new conversation_id. A timed-out run_financial_research retried blind may therefore start a second run rather than rejoin the first. Kepler's documented recovery path is to wait — get_run_result and continue_waiting both resume an existing run — rather than to restart.
  • Access is gated per account, separately from having a login. MCP access is rolled out individually and requires a sign-out and sign-in to take effect. A working Kepler account is not sufficient.
  • Coverage is public companies only. Kepler states it cannot research private companies. Its sources are SEC filings, earnings transcripts and market data, so a name that does not file is out of scope regardless of how the question is phrased.
  • Not every figure is guaranteed cited. Kepler ships uncited_figures and fully_cited in every result, which exist precisely because an answer can contain numbers without sources. Read the sourcing report before relying on a figure.
  • A prompt discrepancy is unresolved. Kepler documents three prompts; Anthropic's directory lists prompt_names: []. We could not run prompts/list to arbitrate because the endpoint requires authentication. Directory prompt listings are known to drift in both directions, so this is most likely a stale listing rather than a missing feature — but we did not verify it.
  • We could not read tool schemas or safety annotations. The endpoint returned 401 to an anonymous request, so parameter types below come from Kepler's documentation, not from the server's own inputSchema.
  • An interactive run panel is documented and unverified. Kepler describes a live run view in supporting hosts. We did not read its bundle, and the directory entry carries no has_mcp_app field. Treat its behaviour as unknown rather than as absent.
  • kepler.ai's robots.txt contradicts itself. It opens with a group disallowing ClaudeBot, GPTBot, Google-Extended and others, then a later group allows exactly those agents. The documentation host, docs.kepler.ai, is unambiguous and permissive, and is where this page's synthesis comes from.

Frequently asked questions

Kepler the AI financial research agent, a New York company whose legal name is Keru AI, founded in 2025. It is not Kepler.gl, Uber's open-source geospatial visualization library, and not Kepler Communications, the satellite operator. The connector reads SEC filings, earnings transcripts and market data, and has nothing to do with maps or satellites.

No. None of the nine tools carries a delete, remove or archive verb, and Kepler documents connector access as read and run management only. The one stop-like tool, cancel_run, is documented as non-destructive: the conversation and any partial results survive, and a follow-up can re-engage it with full context.

Identity claims only. Kepler's RFC 9728 resource descriptor advertised openid, profile, email and offline_access on 2026-08-23, and its authorization server advertises the same four. None of them names a Kepler object or verb, so the consent screen offers no way to grant looking up a company without also granting starting a research run.

Kepler publishes no per-run rate, and no tool in the surface can observe spend. No tool name contains credit, balance, usage, quota or billing, so an agent cannot check a cost before starting a run. Access is granted per account rather than metered publicly, so ask your Kepler contact what a run consumes before allowing unattended runs.

Because Kepler rolls MCP access out per account rather than to everyone with a login. Kepler's documentation states you should ask your Kepler contact to enable it, then sign out of Kepler and back in so your session picks up the change. A Kepler account alone does not grant connector access.

Usually a few minutes, and longer for larger models, because Kepler reads primary sources rather than summarizing from memory. Two tools block while waiting: get_run_result holds the connection open and streams progress, and continue_waiting resumes when a very long run returns still running.

No. Kepler's documentation states coverage is limited to publicly listed companies, reached through their SEC filings and earnings call transcripts plus market data. To check one company before committing to a run, lookup_company resolves a name, ticker or SEC CIK and lists available filings and transcripts instantly, with no run required.

Not through any of the nine tools. Kepler documents access as scoped to your own conversations and results, and no tool shares, publishes or emails. Kepler's app has features that do reach outside — shareable chat links, a forwarding inbox, a SharePoint connector — but none of them appears as a connector tool.

Kepler's documentation describes three: build a financial model, recap an earnings call, and compare companies. Anthropic's directory advertises zero prompt names for Kepler. We could not resolve the discrepancy because the endpoint requires authentication, so treat three as the documented figure rather than an observed one.

Sources

Use in Agentman

Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio

Server Info

Category
Productivity
Developer
Kepler
Tools
9
Domain
mcp.kepler.ai

Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.