MCP

The Browlix MCP server exposes the API v1 catalog as tools, with the same permissions, limits, idempotency and audit as HTTP.

• Endpoint: https://browlix.com/api/public/mcp

• Transport: Streamable HTTP, stateless — every POST carries the API key, so reconnecting is just the next request.

• Protocol revisions (newest first):

• 2026-07-28 — modern (per-request metadata, server/discover, no handshake)

• 2025-11-25, 2025-06-18, 2025-03-26 — legacy (initialize handshake)

A client that speaks 2026-07-28 sends the revision and its capabilities with every request (MCP-Protocol-Version, Mcp-Method, Mcp-Name headers mirror the body) and may call server/discover first; there is no handshake. A client on an older revision starts with initialize and gets its own revision back when it is listed, otherwise the newest legacy one. An unknown revision is refused with error -32022 and the supported list.

Connect a client

Put the key in an environment variable or the client's secret prompt — never in a config file.

export BROWLIX_API_KEY="bxk_…"
claude mcp add --transport http --scope user browlix https://browlix.com/api/public/mcp \
  --header 'Authorization: Bearer ${BROWLIX_API_KEY}'

Cursor reads ~/.cursor/mcp.json; VS Code reads .vscode/mcp.json in the workspace. The web app shows the same snippets under Settings → API keys.

Tools

• One tool per operation, named after its id with underscores: sessions.start → sessions_start. Inputs are the path, query and body fields side by side.
• tools/list returns only the tools the key may use right now.
• Read tools are marked read-only; tools that remove or end something are marked destructive.
• docs_search and docs_fetch read these docs for any valid key.
• Every tool publishes an outputSchema (from revision 2025-06-18), and a successful call returns the same JSON as HTTP in structuredContent and as text. Fields can be added at any time — ignore the ones you do not know.
• With 2026-07-28, tools/list is cacheable for 5 minutes per key (cacheScope: private).

Writes and lost replies

Write tools accept a command_id. If the agent omits it, the server generates one and returns it in the result. After a disconnect, call receipts_get with that id instead of calling the write again.

Errors and feedback

Failures come back as tool results with isError: true and the same error.code as HTTP. For failures Browlix could improve (UNKNOWN, VALIDATION_FAILED, CONFLICT, UNSUPPORTED…) the result adds a hint about feedback_submit — see Feedback from agents.

Clients checked

Each row was run against the Browlix MCP server; a client that could not be exercised says so.