Bowmark

What Bowmark MCP server does

The Bowmark MCP server turns precomputed web navigation recipes into callable functions for agent workflows. It is designed for public websites and lets an agent inspect available functions before interacting with a live site. The server does not expose a broad browser-control API in the supplied material; its workflow centers on reading a typed library, running a short script, and reporting problems.

get_library accepts a task or site query and returns the relevant typed function library. It is read-only and does not contact the site being described. If the query is not recognized, the tool returns the library index instead of failing. This gives an agent a way to discover the available surface before writing a script.

How it works

A typical interaction starts with get_library. The agent uses the returned definitions to construct a short asynchronous JavaScript program, then sends that program to run. Bowmark executes the script against live sites and returns an object containing ok, result, logs, error, and ms fields. These fields provide both the script outcome and basic execution diagnostics.

When a function or execution does not work as expected, report sends feedback to Bowmark. Reports record what failed or which result was incorrect; they do not automatically retry the operation. The material describes this feedback path as part of the service rather than as another site-operation tool.

The Bowmark MCP server is available through an MCP endpoint, and the repository also documents equivalent HTTP calls for environments without an MCP client. The library is retrieved with a GET request to /v1/library?query=…, while execution uses POST /v1/run. Both HTTP calls use a bearer token.

Setup and configuration

The repository provides an npm skill package that can be installed with npx skills add bowmark-ai/skill. It states that the skill standard works with Claude Code, Codex, Cursor, GitHub Copilot, OpenCode, and other compatible hosts. Claude Code and Codex can instead install the Bowmark plugin, which bundles the skill and wires the MCP connection.

Direct MCP setup instructions are available through Bowmark's installation page and Smithery. The README also provides install links for Cursor, VS Code, and LM Studio. For direct HTTP use, set BOWMARK_API_KEY and send it as Authorization: Bearer $BOWMARK_API_KEY.

Tools and capabilities

• get_library: Reads typed functions for a task or site without touching the site.
• run: Executes a short asynchronous JavaScript script against live sites.
• report: Records an inability or incorrect result without retrying.
• HTTP access: Provides library and execution endpoints for clients that do not use MCP.

Limitations and notes

The supplied description focuses on public websites and precomputed recipes. It does not document arbitrary browser navigation, persistent sessions, authentication flows, or a list of supported sites. run executes short scripts, so the material does not establish support for long-running jobs or unrestricted JavaScript.

An unrecognized library query returns an index, which is useful for discovery but is not a successful task-specific lookup. Feedback submitted through report is recorded only; it does not cause an automatic retry. The OpenAI Apps variant has separate packaging details and uses OAuth rather than an API key, according to the repository notes.