Lowdown Proxy

What the lowdown-proxy MCP server does

The lowdown-proxy MCP server has two related roles. In proxy mode, it sits between an MCP client and another MCP server, forwarding protocol traffic while recording interaction metadata to Lowdown. In lookup mode, the lowdown-mcp command runs as a separate stdio MCP server that lets agents query the Lowdown service directly.

Recorded fields include the actor, target label, tool name or protocol method, outcome, latency, node ID, and, when available, the target server's version and tool-schema hash. Tool arguments and results are not recorded. Recording can be disabled with LOWDOWN_RECORD=0.

The package also supports watching application log files. After a tools/call, matching lines written since the previous call can be added to the returned tool result. Log content remains local and is not sent to Lowdown.

How it works

Start a wrapped server with a command such as npx lowdown-proxy -- npx some-mcp-server. A node identity is created at ~/.lowdown/node.json on first use. The proxy uses the public Lowdown API for recording from version 0.1.17 onward, so a database key is not needed.

The lookup process is configured as an MCP stdio entry using npx -y -p lowdown-proxy lowdown-mcp. Its tools retrieve interaction history, compare two to five targets by reported success rate, record an outcome, and return contribution statistics for the local node. Reputation responses can include separate Scanner observations for servers in the official MCP Registry. Scanner data comes from initialization and tool-list requests only, and is not included in success-rate calculations.

Success classification follows the MCP error-handling categories. Successful tool calls count as successes; tool results marked isError count as failures. Certain request errors are tracked as rejected requests rather than rated failures, while other server errors count against the rate. Protocol requests are tracked separately.

Setup and configuration

The basic setup requires Node.js package execution through npx; the README does not specify a required environment variable. Add the proxy command to an MCP client's server configuration and replace the original command with the wrapped command. The target label defaults to a package or file name from the command, not the complete command line.

Useful options include --actor, --target, --source, --test, --watch-log, and --log-pattern. Multiple --watch-log flags can monitor several files. Paths expand %VAR%, $VAR, and ~. The default log pattern matches terms such as error, exception, fatal, uncaught, unhandled, and failed; a custom case-insensitive regular expression can replace it.

Set LOWDOWN_API_URL to use another Lowdown deployment. Set LOWDOWN_RECORD=0 for pass-through operation without recording. The public API requires no authentication, while supplying an x-lowdown-node-id header returns contributor-specific data.

Tools and capabilities

The lowdown-proxy MCP server's lookup interface provides:

• get_lowdown for reported interactions and, where applicable, MCP Registry Scanner observations.
• compare_tools for side-by-side results for two to five targets.
• record_interaction for submitting a tool or service outcome.
• get_node_stats for checking the current node's contribution statistics.

Log reporting is bounded to the latest 20 matching lines per file and 300 characters per line. Bearer tokens, API keys, JWTs, token= values, and long hexadecimal strings are masked. Pending matching lines are deferred when a JSON-RPC error produces no content.

Limitations and notes

Lowdown records observed interaction history rather than ratings or reviews. Organic records determine success_rate; seeded and synthetic records are reported separately. Scanner observations are limited to connection, initialization, and tool-list behavior, and a reported HTTP status is preserved as observed rather than converted into a reachability judgment.

Only lines written after the proxy starts are considered, and each matching line is reported once. Log lines are not uploaded. The --test option marks a session as synthetic so intentional failure tests do not affect the server's organic results. Fuzzy target matching supports short names such as brave-search, fetch, and github.