Connect over MCP

Wire Bixel into Claude, Cursor, or any MCP client — OAuth sign-in or key headers, and the 12 tools.

The MCP server lives at https://api.bixel.com/mcp (streamable HTTP). Tool calls work on every tier, including Free — they just need an identity: OAuth sign-in or any API key. Calls draw your account's monthly credits at the operation's own weight (current-state tools 1 credit, history/diff tools 3).

Claude (custom connector, OAuth)

In Claude, add a custom connector with URL https://api.bixel.com/mcp.
Claude offers sign-in (the server's 401 carries the OAuth discovery metadata). Signing in lands on Bixel's consent page, which shows the app name, the exact redirect URI, and plain-word scopes before you approve.
Approve, and the connector carries your account identity — all 12 tools work; no key is pasted anywhere. Pro unlocks full history depth on the history tools.

Without any sign-in the server answers 401 with the OAuth discovery metadata — that response is what makes clients offer sign-in. Any account works, Free included.

Key-header clients (Claude Code, Cursor, mcp-remote)

Any client that can send a header authenticates with an API key — the free tier's key works:

claude mcp add --transport http bixel https://api.bixel.com/mcp \
  --header "Authorization: Bearer bx_your_key_here"

Cursor and similar (JSON config):

{
  "mcpServers": {
    "bixel": {
      "url": "https://api.bixel.com/mcp",
      "headers": { "Authorization": "Bearer bx_your_key_here" }
    }
  }
}

Clients that only speak stdio can bridge with mcp-remote:

npx mcp-remote https://api.bixel.com/mcp --header "Authorization: Bearer bx_your_key_here"

x-api-key: bx_... works everywhere Authorization: Bearer does.

The 12 tools

Tool	What it answers
`search`	Resolve a free-text name to canonical domains
`list_categories`	Every category with member counts
`get_category`	One category's members with per-dimension rollups
`get_company`	A company's full current record
`get_company_facts`	Current facts, filterable by dimension/provenance
`get_company_jobs`	Live ATS job postings with compensation breakouts
`query_changes`	The cross-company change feed
`get_company_history`	One company's dated change log
`get_company_trajectory`	Direction over 30/90/365-day windows
`check_company_claims`	Claimed vs observed cross-read
`get_fact_evidence`	The provenance trail for one fact
`get_fact_proof`	The cryptographic proof bundle for one fact

Prompts (skills)

The server ships four curated prompts — pre-built workflows your client can surface directly (Claude shows them in the connector's prompt picker):

Prompt	What it produces
`state_of_category`	A dated, cited market report over one category — roster with coverage tiers, per-dimension comparison, recent confirmed changes
`company_brief`	One company's dated brief, scoped to its captured dimensions
`verify_claims`	A claimed-vs-observed audit with evidence and proof pointers
`compare`	Side-by-side over shared dimensions, coverage-fair (never scores a thin record against a deep one)

Each encodes the working method below, so an agent following one cites dates and sources by construction. The same method is available as a readable resource at bixel://methodology.

Patterns that work well

Resolve first. Company arguments are apex domains; run search before assuming one.
Filter by dimension. get_company returns all ten dimensions and is token-heavy; get_company_facts with a dimension argument answers most questions at a fraction of the size and cost.
Cite what the payload gives you. Every fact carries as_of and source_url; a report that quotes them is checkable, one that does not is not.
Read the derivation label. Change events labeled historical are archive diffs dated "by then", not "on that day"; confirmed events are debounced live changes. See the data model.
Mind the meter. Each tool call costs its operation's weight — current-state tools 1 credit, history/diff tools 3 (each tool's description states its cost, and the response's X-Bixel-Request-Cost confirms it). Dimension-filtered calls cost the same 1 credit while keeping your agent's context small enough to need fewer calls.