CtxSlim

Put your MCP servers on a diet. One proxy entry replaces your whole MCP stack — your agent keeps every tool, your context keeps 26k more tokens per request.

npx -y ctxslim init          # detects Claude Desktop / Cursor / Windsurf / VS Code / Claude Code
npx -y ctxslim init --client cursor --yes    # wires itself in, timestamped backup included

Your servers stay exactly where they are. CtxSlim reads them (read-only), connects to all of them itself, and your client only ever sees one MCP server: Slim.

Why

Every MCP server dumps its entire tool catalog into every single LLM request — names, descriptions, JSON schemas, the lot. Six servers × twelve tools ≈ 36k tokens per request, before your prompt even starts. That's slower answers, bigger bills, and a worse agent.

Measured through real MCP transports (npm run bench). Bigger stacks save more.

Related MCP server: mcp-proxy

Features

• Adaptive context budget — ranked tools are admitted by exact token cost, not just count; the BudgetGuard enforces a hard ceiling so an oversized tool cannot silently blow the budget

• Cost-exact compression — admission cost is calculated from the exact compressed representation that will be exposed, eliminating estimate/exposure drift

• Predictive tool affinity — recent tool sequences create a local affinity graph, so tools commonly used after the current tool receive a bounded routing boost

• Top-K exposure — BM25 ranking + your usage history pick the tools that matter per turn; the rest stay one search_tools away, never blocked

• Schema compression — boilerplate stripped, descriptions budgeted, dead $defs dropped

• Adaptive ranking — tools you actually call get a boost, learned across sessions

• Progressive disclosure — opt-in stub listings (~40 tokens/tool) with on-demand full schemas

• Lazy connect — instant startup; upstream servers boot in the background

• Output diet — safely minify JSON text results, truncate giant text results, and downsample screenshots (opt-in, per server)

• Spend audit — every call metered; ctxslim audit prices it per task in dollars

• Auto-tune — ctxslim doctor --tune reads your real usage and suggests config fixes (never writes)

• 100% local — no API keys, no telemetry, no cloud. Works with stdio and Streamable HTTP servers

How it works

flowchart LR
    C[MCP Client<br/>Claude · Cursor · Codex] <--> S[CtxSlim<br/>one stdio entry]
    S <--> A[files]
    S <--> B[database]
    S <--> D[browser]
    S <--> E[…]
    style S fill:#16a34a,stroke:#14532d,color:#fff

1. Startup — Slim answers your client instantly, boots your servers in the background, and announces each as it lands.

2. Listing — every tools/list scores all tools (search relevance + usage + local tool affinity + pins − globs), compresses them once, then admits the highest-value representations through a hard token budget. Oversized candidates are skipped instead of overflowing the budget.

3. Discovery — search_tools and describe_tools pull full schemas on demand; hidden tools stay directly callable. Nothing is hard-blocked, ever.

4. Calls — routed transparently to the right server; results pass through your squeezes (output.maxChars, images) before entering context.

5. Learning — every call feeds local stats; slim_stats shows the session, stats the lifetime, audit the money, doctor --tune the next config fix.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp"],
      "include": ["browser_*"],          // per-server globs; exclude wins
      "exclude": ["*_debug*"],
      "output": { "maxChars": 4000 },    // truncate this server's text results
      "images": { "scale": 0.5, "format": "jpeg", "quality": 70 }  // needs optional `sharp`
    }
  },
  "slim": {
    "mode": "auto",              // auto | manual (allowlist + pins) | off (pure aggregation)
    "maxTools": 24,              // hard cap on exposed tools
    "contextBudget": 6000,        // adaptive token budget for exposed tool definitions
    "adaptive": true,            // learn from usage across sessions
    "disclosure": false,         // stub listings + on-demand schemas
    "pins": ["playwright::browser_navigate"],
    "descriptionBudget": 280,
    "connectTimeout": 15000,
    "stats": true
  }
}

ctxslim                       start the proxy
ctxslim init                  wire into a client config (--client, --yes, backup included)
ctxslim --config <path>       use a specific config
ctxslim --mode <mode>         auto | manual | off
ctxslim --max-tools <n>       override top-K
ctxslim --no-stats            no stats, usage, or audit files
ctxslim --quiet               minimal logging
ctxslim stats [--json]        lifetime savings
ctxslim audit [--gap] [--model] [--prices] [--json]
                              per-task dollar spend across 10 real models
ctxslim doctor [--tune] [--json]
                              validate config (+ suggest-only tuning)

Every routed call is metered to ~/.ctxslim/audit.jsonl (sizes + hashes, never content). ctxslim audit groups calls into tasks by idle gap and prices them:

tasks               12
tool calls          148
tool-output spend   claude-sonnet-5 $0.0156  claude-opus-5 $0.0390  gemini-3.8-flash $0.0059
definitions/request ~9.2k tokens  claude-sonnet-5 $0.0184/$0.0018 (full/cached)
duplicate waste     $0.0021 (claude-sonnet-5)

Tool outputs are priced as input tokens; definitions show full + prompt-cached cost. Rates are indicative (as of 2026-09-08) for claude-fable-5-1, claude-opus-5, claude-sonnet-5, claude-haiku-4-5, gpt-6-astra, gpt-5.6-sol/terra/luna, gemini-3.1-pro, gemini-3.8-flash — override anytime with --prices.

Context engine architecture

CtxSlim treats context as a constrained resource rather than a pile of JSON.

MCP servers
    │
    ▼
Context Router ── relevance + usage + tool affinity
    │
    ▼
Tool Optimizer ── schema compression + metadata preservation
    │
    ▼
BudgetGuard ──── exact admission cost + hard ceiling
    │
    ▼
MCP client context
    │
    ▼
Result Optimizer ── JSON compaction + truncation + images

Experimental techniques

• BudgetGuard: one deterministic admission controller owns the context ceiling. Pinned tools no longer bypass the ceiling by accident.

• Cost-Exact Compression: a candidate is compressed once and its actual serialized representation is used for admission accounting.

• Predictive Tool Affinity: successful tool sequences build a bounded in-memory transition graph. If browser_navigate → browser_click is common in a session, the second tool receives a small routing boost on the next listing.

• Progressive Discovery: tools that do not fit remain callable and discoverable through search_tools, avoiding the classic top-K dead end.

The design goal is not merely a high percentage reduction. It is bounded context growth: adding MCP servers should increase discovery space without forcing the full catalog into every request.

FAQ

Yes. search_tools returns full schemas for matches, and even a direct call to a known-but-hidden tool gets routed. Nothing is ever hard-blocked.

Because you installed them for a reason. The problem isn't having tools — it's paying for all of them in every single prompt.

131 of them (npm test) — compressor, ranking + adaptive scoring, globs, truncation, disclosure, images, lazy connect, persistence, audit/pricing/tune, plus an integration suite speaking real MCP over real transports.

Contributing

Genuinely welcome — see CONTRIBUTING.md. Small, strict, comment-free codebase; a nice one to read. Say hi via issues.

Star history

License

MIT © 2026