Clear-Code and the LeanMCP AI Gateway: seeing what Claude Code sends

What Clear-Code actually is, and what it is not

Two things share one README, and conflating them is the main source of confusion. The first is a curated guide: the README calls Clear-Code "a comprehensive, community-driven resource hub dedicated to cataloging, comparing, and promoting the best open-source AI coding assistants," with entries for Cline, OpenCode, OpenHands, Aider, Continue, Tabby, Void and Goose, plus a comparison table and a section on choosing a tool. The second is a hosted logging service. The repository is written in Shell, licensed MIT, and its top-level entries are .gitignore, LICENSE, README.md, claude-code-skills/ and tests/. There is no proxy binary, no server directory, no Dockerfile. The logging itself runs at aigateway.leanmcp.com, not in the repository.

That distinction matters for anyone evaluating it. If you want a comparison of open source coding assistants you can read the README and the claude-code-skills/ directory. If you want to see what your own Claude Code session is sending, you are signing up for a third-party gateway, and the repository is only the documentation and the redirect snippet. Nothing in the layout suggests the logging code is published here.

The redirect: how the gateway captures Claude Code traffic

The mechanism is an endpoint substitution rather than an agent, a plugin or a wrapper process. Claude Code reads ANTHROPIC_BASE_URL to decide where to send its API calls. Point that variable at the LeanMCP gateway and every request the tool makes, including the file contents it reads and the tokens it spends, passes through a server that can record it. The README frames the offer as "Every request. Every file it reads. Every token it burns," and describes the setup as "One URL change. No code required."

Because the redirect sits at the HTTP layer, it is client-agnostic in principle: anything that respects ANTHROPIC_BASE_URL and an API key will be routed the same way. The trade-off is equally structural. The gateway is in the request path, so it sees prompt content, and it is a network dependency, so a gateway outage or a DNS problem becomes a Claude Code outage. The README does not describe a fallback path or a local buffer, which means the failure mode is a failed request rather than a degraded one. It also does not state whether the gateway forwards to Anthropic unchanged or rewrites anything, so treat the request path as opaque until you have confirmed it yourself.

Installing the redirect and reading your first log

There is nothing to install from the repository. The README gives two exports and points at a web app for the key and the dashboard. Set the variables in the shell that launches Claude Code, not in a file the tool does not read. The README shows exactly this:

export ANTHROPIC_BASE_URL=https://aigateway.leanmcp.com/v1/anthropic
export ANTHROPIC_API_KEY=your_leanmcp_key

The first line moves the base URL to the gateway's /v1/anthropic path. The second replaces your provider key with a LeanMCP key, which the README says you obtain at app.leanmcp.com/billing. Note that the variable name stays ANTHROPIC_API_KEY even though the value is no longer an Anthropic key; if your client validates key format, that is worth checking before you debug anything else.

With both exported, run any Claude Code task in the same shell. The README's instruction is to "Run any Claude Code task" and then open app.leanmcp.com to see the logs. If the dashboard stays empty, the first thing to confirm is that the task inherited the exported variables, since a shell that was already running before the export will not have them. The README offers no troubleshooting section, so that check is on you.

The catalogue half: Cline as the reference agent

The featured list is the more developed part of the README, and Cline gets the longest entry. The description is specific about what separates it from a completion engine: "Cline doesn't just suggest code, it can create and edit files, run terminal commands, use the browser, and execute multi-step tasks autonomously," with each action approved through a human-in-the-loop GUI. The README also lists model-agnostic operation across OpenAI, Anthropic, Google, Mistral, Ollama, LM Studio and any OpenAI-compatible API, first-class Model Context Protocol support, project-wide context reading, and browser automation with screenshots and clicks.

The README states Cline is licensed Apache 2.0. Other entries are thinner. OpenCode is described as terminal-native and the section is truncated in the README text, and the remaining tools (OpenHands, Aider, Continue, Tabby, Void, Goose) are catalogued with less detail. Treat the comparison table as an index into each project's own documentation rather than as an evaluation. The Quick Start block does give install commands for four of them, including `pip install aider-chat` for Aider and `pipx install goose-ai` for Goose, which is the most immediately usable content in the catalogue.

Where the gateway is the wrong tool

The clearest limitation is that the logs are not yours. The README's own framing of the problem with closed tools applies to this service as well: it warns that with proprietary assistants "You don't know what data is being collected, how your code is being processed, or what happens to your proprietary business logic when it's sent to a remote server." Routing Claude Code through a hosted gateway puts prompt content, and therefore source code, through a third party by design. The README does not document retention, deletion, regional storage or an on-premises deployment for the gateway, so a team that cannot send code to a vendor cannot use this feature regardless of how useful the dashboard is.

There is a second, quieter gap. The README says nothing about rollback. Removing the two exports returns you to your previous endpoint, but the README does not say whether logs already captured are deletable, whether the key can be revoked from the billing page, or what happens to in-flight requests when you switch back. For a change that touches every API call your editor makes, that silence is a real cost. Finally, the repository is Shell and the logging is remote, so anyone hoping to run this inside an air-gapped network has the wrong project.

A real alternative: local proxying with a self-hosted tool

If the goal is visibility without a vendor in the path, the difference is where the recording happens. A local proxy runs on your machine, listens on a loopback port, forwards to Anthropic, and writes request and response pairs to disk you control. The Claude Code setup looks the same from the outside, since it is still an ANTHROPIC_BASE_URL change, but the value points at localhost and the logs never leave the host. That is the approach to take when the constraint is data residency rather than convenience.

The trade-off runs the other way. A local proxy gives you raw files and no dashboard, no cross-machine view, and no hosted key management; you own the storage, the rotation and the failure handling. The LeanMCP gateway's advantage is precisely that it is one URL and a signup, with the UI already built. The honest comparison is not which is more capable but which cost you can absorb: operating a proxy yourself, or sending prompt content to a service whose retention policy the README does not state.

Maintenance, licence and what the repository does not say

The repository is not archived and is licensed MIT, which covers the README, the claude-code-skills/ directory and the tests/ directory. MIT is permissive: you can copy and adapt the documentation and the skills package, and there is no copyleft obligation. The hosted gateway is a separate service and its terms are not in the repository, so the MIT grant should not be read as covering the API or the dashboard.

On maintenance, no last-push date is available for this repository, so there is no basis for calling the project actively developed or regularly updated, and the release list is empty. The practical consequence is that the catalogue will drift as the listed tools ship changes, and the gateway documentation is the only part that has to track a live service. Before depending on either half, check the commit history yourself and, for the gateway, check whether the docs at docs.leanmcp.com/ai-gateway/claude-code have moved past what the README shows.

Editorial conclusion

Adopt the gateway if you are already running Claude Code with an Anthropic-compatible endpoint and you want a hosted view of which files and how many tokens each task consumes, and if you accept that the logs live on LeanMCP's servers rather than yours. Do not adopt it if your code cannot leave your network, if you need the logging to work offline, or if you came for a self-hosted logging tool: the repository is a Shell-based catalogue, not a proxy you run yourself. Before committing, verify three things: that ANTHROPIC_BASE_URL accepts the /v1/anthropic path your client will append, that your LeanMCP key is scoped the way you expect, and that the dashboard retains logs for as long as your review process needs. The README documents the redirect and nothing about retention, so that last point has to be answered by the vendor, not by the repository.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository