io.github.cyanheads/browser-compat-mcp-server

@cyanheads/browser-compat-mcp-server

Browser compatibility and Baseline status for any web feature — offline, from MDN's browser-compat-data, web-features, and caniuse. STDIO or Streamable HTTP.

5 Tools

Public Hosted Server: https://browser-compat.caseyjhand.com/mcp

Overview

Browser support and Baseline status for any web platform feature, from MDN's @mdn/browser-compat-data, web-features, and caniuse-lite usage data. Check whether a feature is safe to ship against your browserslist targets. All data ships in the package: no network calls, no API key, no rate limit.

Tools

Capability reference

browsercompat_list_reference tool

• Required topic: bcd_namespaces (12), bcd_browsers (17), browserslist_agents (19), baseline_states (4), groups (104), or snapshots (11). Group and snapshot ids feed the search filters.
• Entries carry id, label, and detail, plus count, reported, bcd_browser, usage_percent, maps_from, or spec_url where applicable. An agent's bcd_browser: null means support comparisons cannot evaluate it and report it in unchecked_targets.

browsercompat_get_feature tool

• One feature, 1–200 characters: a BCD key (css.selectors.has) or web-features id (has). resolve: true (off by default) also accepts a name or notation when it matches exactly one feature. include_runtimes: true adds bun, deno, nodejs, and oculus support rows.
• outcome is found, no_compat_data, or miss; a miss returns found: false with guidance, and whitespace-only input returns invalid_feature_input. Each support row gives the newest release's statement, with the other BCD statements for that browser in other_statements.
• A multi-key id such as grid returns feature_support (one row per Baseline core browser) instead of support, status, mdn_url, and spec_urls; call one of its compat_keys for a key-level answer. compat_keys and subkeys page 100 at a time with compat_keys_offset and subkeys_offset (default 0). limiting_browser is the Baseline core browser with the newest required release.

browsercompat_check_baseline tool

• BCD keys or web-features ids, 1–200 characters each. Up to 100 are accepted and the first 20 are evaluated in input order; the rest come back verbatim in omitted_features. More than 100 is rejected, and a whitespace-only entry returns invalid_feature_input.
• Results carry Baseline state and dates, limiting_browser, deprecated, experimental, and discouraged, plus usage_percent_excluded and usage_source where caniuse covers the feature. Usage is left out when unavailable, never reported as zero.
• all_widely_available is true only when every entry resolves at widely and none was omitted; it does not assess deprecation.

browsercompat_search_features tool

• query, 1–100 characters, accepts names, keywords, or notation such as Array.prototype.at, display: grid, or <dialog>. Combine namespace (12 BCD namespaces), baseline (widely, newly, limited, not_mapped), group (including nested groups), and snapshot (such as ecmascript-2023).
• matched_on says why a result ranked where it did, and support_summary covers the seven Baseline core browsers (— unsupported, ? unknown). Typed failures: invalid_query (zero searchable tokens), unknown_group, unknown_snapshot.
• Page with limit (default 10, at most 50; larger values are served as 50 and reported in cap) and offset (default 0); totalCount and nextOffset drive paging. Zero hits succeed with a notice naming a filter to drop.

browsercompat_compare_support tool

• Up to 20 features per call against a required targets browserslist query, such as defaults or > 0.5%, last 2 versions; local browserslist config is never used. Up to 100 features are accepted and the rest come back in omitted_features, which makes all_clear false. Typed failures: invalid_target_query, no_targets_resolved, invalid_feature_input.
• Each verdict is clears, fails, inconclusive, miss, or ambiguous. A feature clears only when every target in the query was evaluated and supports it, so a query that includes a browser with no compatibility data, as defaults does, never clears. all_clear is true only when every feature clears.
• Results report support_source, failing_ranges, failing_total, and evaluated and unchecked totals with coverage percentages. Target rows page with target_offset (default 0) and target_limit (default 10, at most 10; larger values are served as 10 and reported in cap); verdicts and totals are the same on every page, and totalCount and nextOffset drive paging.

Data sources

CC BY 4.0 requires attribution wherever the caniuse data travels, so every response carrying a usage figure carries this string: Usage data from caniuse.com, © Can I Use contributors, CC BY 4.0. Figures are a share of the ~97.3% of global traffic caniuse tracks. Full license texts and notices are in THIRD_PARTY_NOTICES.md.

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Browser-compat-specific:

• All four datasets are bundled and loaded in process — no runtime network calls, no API key, no rate limit, and nothing to configure
• Baseline is read per browser-compat-data key from status.by_compat_key, never rolled up from the feature level, because keys under one feature legitimately disagree
• Every tool resolves a feature the same way: exact BCD key, then web-features id, then a moved redirect
• Target versions follow browser-compat-data's release order, with caniuse spellings normalized (safari 16.0 ↔ 16)

Agent-friendly output:

• Every response echoes data_version, the version of each bundled dataset behind the answer
• A pass is never claimed for a browser that was not evaluated; targets without data land in unchecked_targets and make the verdict inconclusive
• Misses are results, not failures: found: false with guidance naming the next call, and typed errors with recovery hints
• Usage figures state the population they are a share of and carry the caniuse attribution

Getting started

Public Hosted Instance

A public instance is available at https://browser-compat.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "browser-compat-mcp-server": {
      "type": "streamable-http",
      "url": "https://browser-compat.caseyjhand.com/mcp"
    }
  }
}

Self-Hosted / Local

Add the following to your MCP client configuration file:

{
  "mcpServers": {
    "browser-compat-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/browser-compat-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "browser-compat-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/browser-compat-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "browser-compat-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "MCP_TRANSPORT_TYPE=stdio",
        "ghcr.io/cyanheads/browser-compat-mcp-server:latest"
      ]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No API keys, accounts, or network access required — every dataset ships with the package.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/browser-compat-mcp-server.git

1. Navigate into the directory:

cd browser-compat-mcp-server

1. Install dependencies:

bun install

1. Configure environment (optional):

cp .env.example .env
# edit .env if you want to override transport or logging defaults

Configuration

There are no server-specific environment variables: no API keys, no base URLs, and deliberately no browserslist configuration variable — the target query is always a tool input rather than ambient state. Framework transport, logging, and telemetry settings remain configurable.

See .env.example for the full list of optional framework overrides.

Running the server

Local development

# One-time build
bun run rebuild

# Run the built server
bun run start:stdio
# or
bun run start:http

bun run devcheck   # Lint, format, typecheck, security
bun run test       # Vitest test suite
bun run lint:mcp   # Validate MCP definitions against spec

Docker

docker build -t browser-compat-mcp-server .
docker run --rm -p 3010:3010 browser-compat-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/browser-compat-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

The generated file tree is docs/tree.md.

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools directly in src/index.ts
• Data integrity: read the bundled datasets as they are and preserve their uncertainty; never fabricate a support fact the data does not carry

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.