MCP RTK: cut 90% of MCP server tokens

I use Claude Code every day for development. Like many developers, I've connected several MCP servers (GitLab, Grafana, Sentry...) to give Claude direct access to my tools. The problem: every MCP call injects tens of thousands of tokens into the context, and the bill spirals fast.

I built MCP RTK to fix this. It's an MCP proxy written in Rust that sits between Claude Code and MCP servers, filtering responses before they reach the model. Result: over 267 million tokens saved across my 38,000+ commands.

The problem: oversized MCP responses

The MCP protocol (Model Context Protocol) lets Claude interact with external tools. When Claude calls an MCP tool, the server returns a JSON response. The issue is that these responses often contain:

• empty or null fields that add nothing
• technical metadata (internal timestamps, pagination IDs, HTTP headers)
• very long values (full logs, HTML descriptions, entire diffs)
• arrays with dozens of entries when 5 would suffice

A single list_issues call on GitLab can consume over 180,000 tokens. Claude only needs a fraction to answer the question. The rest is pure waste.

Over a typical work session with 50 to 100 MCP calls, that easily adds up to 500,000 wasted tokens injected into the context.

The solution: an 8-step filtering proxy

MCP RTK sits transparently between Claude Code and MCP servers. No workflow change needed: Claude keeps calling the same tools, but responses pass through a filtering pipeline before reaching the context.

The pipeline has 8 steps:

1. Remove null and empty fields - fields without values are stripped
2. Truncate long strings - values exceeding a configurable threshold are cut
3. Field whitelist - only useful fields are kept
4. Field blacklist - known useless fields are removed
5. Deduplication - identical entries are merged
6. Array compression - long arrays are sampled
7. Technical metadata removal - API-internal fields are stripped
8. Normalization - output format is standardized

Each step is independently configurable. You can enable or disable each filter, adjust thresholds, and define server-specific rules.

Configuration with presets

Configuration uses a TOML file. MCP RTK ships with community presets for popular servers:

[servers.gitlab]
preset = "gitlab"

[servers.grafana]
preset = "grafana"

[servers.sentry]
preset = "sentry"

Each preset defines which fields to keep, which to exclude, and appropriate truncation thresholds for the server. For custom servers, you define rules directly:

[servers.my-api]
whitelist = ["id", "name", "status", "created_at"]
max_string_length = 500
max_array_length = 10

MCP RTK auto-detects installed MCP servers and offers to configure them.

Results

Across my 38,000+ commands, MCP RTK has saved 267 million tokens with an average reduction rate of 87%. On Opus 4.6 ($15/M input tokens), that's roughly $4,000 in savings:

Useful information is preserved. Claude responds with the same accuracy, but consumes far fewer tokens per session. The mcp-rtk gain command lets you track savings in real time:

Tokens saved:      267.1M (86.7%)
Efficiency meter: █████████████████████░░░ 86.7%

Installation

MCP RTK is distributed as a single binary:

cargo install mcp-rtk

One line change in your Claude Code config - wrap the existing MCP command with mcp-rtk --:

{
  "mcpServers": {
    "gitlab": {
      "command": "mcp-rtk",
      "args": ["--", "npx", "-y", "@nicepkg/gitlab-mcp"],
      "env": { "GITLAB_PERSONAL_ACCESS_TOKEN": "glpat-..." }
    }
  }
}

MCP RTK detects the upstream server from the command and loads the matching preset automatically.

Why Rust?

Rust was a deliberate choice. The proxy must process every MCP response with minimal latency to avoid slowing down the workflow. Rust provides:

• near-instant startup (no JVM, no runtime)
• minimal memory footprint (a few MB)
• a single binary with no dependencies to install
• compile-time memory safety guarantees

The code is open source

MCP RTK is published under the MIT license on GitLab (mirror on GitHub). Community presets are maintained by users: anyone can contribute their own configurations for new MCP servers.

The project is part of a tooling ecosystem I'm building around Claude Code, alongside Skill Radar (detecting repetitive patterns in sessions) and Claude Deck (multi-workspace for parallel sessions). To learn how to write your own skills, see the guide on effective skills. The project page has installation links and documentation.

Comments

[Raknaos]

Filtering beats model swaps for cost. Did dropping detail ever cause a wrong tool choice downstream? Token cost is easy to measure, silent reasoning failures are not.