Canary Data
by Canary Data
Screen public equities for accounting, fraud and management red flags from an AI agent. Anthropic lists 1 tool, Canary's docs describe 16, and the server's own OAuth descriptor names 26. Read-only, OAuth 2.0 with PKCE.
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 Canary Data via MCP
https://api.canary-data.com/mcpWorks in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.
Canary Data Tools & Capabilities (26)
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
Limits
- Anthropic's directory listing is 1 tool against 26. It names only
new_accounting_issue. The name is real, but a reader working from the directory would miss the universe screen, the KPI history, the analyst plans and the 13F surface entirely. - Ten tools have a scope and no documentation. The 13F, transactions,
accounting_waterfall,business_qualityandfundamentals_metricsscopes are advertised by the server with no published description, parameters or response shape. What they return is unknown to us. - We could not read tool schemas or safety annotations. The endpoint returned 401 to an anonymous request, so no parameter-level detail beyond Canary's documentation is published here, and no
readOnlyHintordestructiveHintvalues exist on this page. - The scope granularity does not separate risk levels. Twenty-six scopes sounds like fine control, and it is — over which datasets an agent reaches. It draws no line between a routine operation and an irreversible one, because on a read-only server no such line exists. If Canary later adds a write tool, the existing scope pattern would give it its own scope, but nothing in today's descriptor pre-commits to that.
- The authorization server offers a password grant.
grant_types_supportedadvertisespasswordalongsideauthorization_codeandrefresh_token. The resource-owner password credentials grant is deprecated in OAuth 2.1 and hands a user's actual password to the client. MCP clients use the authorization-code flow, so this does not affect a Claude connection, but it is a broader surface than the connector itself needs. - No dynamic client registration. A POST to
/registerreturned 404 on 2026-08-22, and the authorization-server metadata declares noregistration_endpoint. Clients that require DCR-based OAuth may not be able to complete sign-in without a pre-registered client. - Screens are window-bounded by design. Canary states there is deliberately no all-history option on
flag_screen. Long-horizon questions need several calls or an explicitsincedate. - A recent podcast figure can displace an earnings-call one.
kpi_snapshotreturns only the newest value per metric across both source types. Canary documents this and thedocument_typefilter that fixes it, but an agent that does not pass the filter can quietly report a podcast number as the latest reported figure. - Rate limits are 120 requests per minute per token, returning HTTP 429 with
Retry-After. A universe screen followed by per-company drill-downs across many rows will reach that quickly. - This is research data, not advice. Flags are Canary's assessments derived from filings and public sources. Every flag record carries source links; follow them before acting.
Frequently asked questions
Twenty-six, on the best available evidence. Anthropic's directory lists one. Canary's own documentation describes sixteen. The server's OAuth protected-resource descriptor, read on 2026-08-22, advertises twenty-six scopes named one-per-tool. The descriptor is the server speaking about itself, so it is the strongest of the three enumerations.
No. Anthropic's directory records the connector's permissions as read only, and every one of the twenty-six scopes the server advertises names a retrieval operation. No scope and no documented tool carries a create, update, delete, submit or send verb. The connector reads Canary's research datasets and returns them.
Red-flag research on public companies. Canary's documentation describes accounting and disclosure flags, insider trading, fraud and malfeasance, management track record, and fundamental business risk, plus KPI time series, earnings-call topics and bull/bear debates. Each flag record carries a date, severity, description and source links.
Yes to both. Canary's documentation states its universe-wide screen has no market-cap floor, so small caps are included, and lists supported markets as global public equity markets. Country filtering uses ISO 3166-1 alpha-2 codes, and the screen covers active companies only.
One scope per tool, twenty-six in total. Every scope name matches a tool name exactly, which is unusually fine-grained for an MCP server. Because all twenty-six are reads, the granularity limits which datasets an agent reaches rather than separating reading from writing — there is nothing destructive to fence off.
Canary's documentation states data is updated daily and that the server allows 120 requests per minute per access token. Exceeding that returns HTTP 429 with a Retry-After header. Access tokens live one hour and refresh tokens thirty days, rotating on use.
Add the server over HTTP transport, then authorise. Canary documents the command claude mcp add --transport http canary-data https://api.canary-data.com/mcp, followed by running the slash-mcp command in your session to complete the OAuth flow. Claude Desktop and Claude.ai Team and Enterprise add it from the Connectors settings page instead.
The response says so explicitly rather than returning a bare empty list. Canary's documentation describes a coverage object carrying a plain-language hint, the echoed ticker and country, and any active date bounds. It distinguishes no flags in this window from no flags at all, so an agent cannot mistake a narrow query for a clean record.
Sources
- Canary Data MCP developer documentation (retrieved 2026-08-22). The
documentationURL Anthropic's directory publishes for this connector; 4,658 words of plain server-rendered HTML, no client-side payload. Source of all sixteen documented tool descriptions, parameters, response shapes, markets, freshness, rate limits and setup instructions. · retrieved 2026-08-22 - Live RFC 9728 protected-resource descriptor (retrieved 2026-08-22). Source of the 26-scope enumeration. Three control probes at sibling paths all returned an identical-hash 404, so the 200 is a genuine descriptor rather than a catch-all. · retrieved 2026-08-22
- Live OAuth authorization-server metadata (retrieved 2026-08-22). Advertises the same 26 scopes byte-identically, plus the
passwordgrant type and the absence of a registration endpoint. · retrieved 2026-08-22 - Live authentication challenge — an anonymous
initializetohttps://api.canary-data.com/mcpreturned HTTP 401 on 2026-08-22 withwww-authenticate: Bearer resource_metadata="https://api.canary-data.com/.well-known/oauth-protected-resource". This is our own observation of the auth posture. · retrieved 2026-08-22 - Anthropic Connectors Directory entry — , read from our directory snapshot dated 2026-08-16. Source of the partner tier, the single
tool_namesentry, and the "Read only" permissions field. · retrieved 2026-08-16 - Canary Data company site — · Privacy — . The site serves
x-robots-tag: noindexand an emptyrobots.txt;api.canary-data.com/robots.txtand/llms.txtboth return 404, so no crawl policy orContent-Signalwas published on either host. - Canary Data support — <mailto:joe@canary-data.com>
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 StudioServer Info
- Category
- Finance
- Developer
- Canary Data
- Tools
- 26
- Domain
- api.canary-data.com
Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.