trimcp

MCP proxy that cuts LLM token costs by 60–90% through lossless compression and caching.

---

MCP tool outputs are full of noise. Pretty-printed JSON, ANSI color codes, repeated log lines, inline comments — none of it adds information, all of it costs tokens. On a busy session with filesystem and git tools, this quietly burns 30–50% of your context budget before the model even starts thinking.

trimcp sits transparently between your LLM client and every MCP server. It strips the noise, caches identical calls, and reports exactly what it saved.

LLM client (Claude Code, Cursor, claude_desktop…)
        ↓
    trimcp          ← strips ANSI, minifies JSON, deduplicates lines, caches
     ↙    ↓    ↘
  MCP1  MCP2  MCP3

[trimcp] session summary
  Requests    : 47
  Tokens in   : 84,312
  Tokens out  : 31,847  (62% smaller)
  Cache hits  : 12 / 47

Zero configuration required — it works out of the box.

---

Install

cargo install trimcp

---

Quick start

# Register your MCP servers once
trimcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /tmp
trimcp add github    -- npx -y @modelcontextprotocol/server-github

# Point your LLM client at trimcp instead of the server directly
trimcp proxy filesystem

---

LLM client configuration

Replace your server commands with trimcp proxy <name>. Nothing else changes.

Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "filesystem": { "command": "trimcp", "args": ["proxy", "filesystem"] },
    "github":     { "command": "trimcp", "args": ["proxy", "github"] }
  }
}

Cursor / claude_desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "filesystem": { "command": "trimcp", "args": ["proxy", "filesystem"] }
  }
}

---

What gets compressed

Strategy	What it removes	Typical saving
strip_ansi	Terminal color/cursor codes (\x1b[…)	~5%
compact_json	Whitespace in pretty-printed JSON	~30%
dedup	Consecutive identical lines → line (xN)	~40% on logs
minify	Trailing whitespace, extra blank lines	~5%
strip_comments	// full-line comments in code blocks	~10%

All strategies are lossless — no information is truncated or dropped.

---

Configuration

~/.config/trimcp/config.toml — created automatically on first add.

[servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]

[compression]
enabled       = true   # master switch
strip_ansi    = true
compact_json  = true
dedup         = true
minify        = true
strip_comments = false

[cache]
enabled  = true
ttl_secs = 300    # cache lifetime in seconds

[metrics]
enabled  = true   # print summary to stderr at session end
realtime = false  # print running total after each tool call

---

Optional: semantic code context (semtree feature)

Enable the semtree feature to inject relevant code snippets into every tools/call response. The model sees the context it needs without you having to ask for it.

# Cargo.toml
trimcp = { version = "0.1", features = ["semtree"] }

# ~/.config/trimcp/config.toml
[servers.my-server]
semtree_codebase = "/path/to/your/project"
semtree_top_k    = 3

# Index the codebase once
# trimcp semtree-index my-server

---

CLI

trimcp add <name> -- <command> [args…]   Register a server
trimcp remove <name>                      Remove a server
trimcp list                               List configured servers
trimcp proxy <name>                       Run as proxy (used by LLM clients)
trimcp stats                              Show compression statistics
trimcp knowledge                          Inspect the semantic knowledge store
trimcp semtree-index <name>               Index a codebase (--features semtree)

---

License

MIT — see LICENSE