sipcode

Sipcode

Sip your tokens — don't gulp them. Keep Claude Code's context clean: fewer tokens spent, sharper answers, and the savings measured so you can trust them.

Visit the site → · npm · benchmark methodology · changelog

Token savings (illustrative 30-min Claude Code session)

Operation	How often	Without Sipcode	With Sipcode	Savings
git status	10×	8,000	1,500	−81%
git log	5×	12,500	2,000	−84%
git diff / git show	4×	16,000	6,000	−62%
grep -r	6×	12,000	2,400	−80%
ls / find	8×	12,000	3,200	−73%
cat (large files)	5×	15,000	6,000	−60%
npm ls	2×	10,000	1,600	−84%
Grep tool	4×	10,000	2,000	−80%
Glob tool	3×	4,500	1,500	−67%
Total		~100,000	~26,200	~−74%

⚠️ These are illustrative estimates for a tool-heavy session, not a promise about yours — savings depend entirely on how many commands your session runs. Two numbers you can trust instead: the reproducible benchmark below (npx sipcode benchmark → 62.6% median on a locked corpus), and your own real number after using it (sipcode proxy --stats).

What Sipcode is — in one picture

Sipcode is two things, and it matters which surface you're on:

	Claude Code (terminal / VS Code / JetBrains)	Claude Desktop (the chat app)
The valve — actually reduces tokens by rewriting commands	✅ Yes — this is where it works	❌ No (chat has no shell commands to optimize)
The meter — audit, cost prediction, savings proof, score	✅ via CLI and MCP tools	✅ via MCP tools

Plain-English version: if you use Claude Code in your terminal, install the valve and it quietly saves you tokens. If you only use the Claude Desktop chat app, you get the meter (analytics + cost tools) inside chat, and you can use it to turn on the valve for whenever you do use Claude Code.

Why clean context matters — not just cost

A bloated, redundant context doesn't only cost more tokens — it makes the model worse. As stale junk piles up, Claude starts hedging, forgetting fixes you already made, and mixing up files. People call it context rot. Keeping context lean is both a measured token win and — per Anthropic's own published research — a quality win (leaner context ≈ 29% better outcomes; well-maintained context ≈ 40% fewer agent errors).

So "sip, don't gulp" cuts two ways: a gulp is a bloated context that's expensive and makes your agent sloppy; a sip is clean, measured context that's cheaper and sharper. Sipcode cleans context automatically (the valve), measures it (the meter), and warns you when it starts to rot (drift, below). The token savings are your proof the cleaning is real.

We cite the reliability numbers above from published research — we don't claim them as Sipcode's own measurement. The part Sipcode measures and proves is the token/context reduction.

Get started

A) Claude Code — turn on the valve (saves tokens)

Copy-paste this one block. That's the whole setup:

npm install -g sipcode      # 1. install the toolkit
sipcode proxy --install     # 2. turn ON the optimizer (writes a Claude Code hook)

Then restart Claude Code and work exactly as you normally do. Done.

🔑 Important: step 1 alone does nothing to your tokens — it just puts the tool on your machine. The optimizer only turns on after step 2 (sipcode proxy --install). To turn it off at any time: sipcode proxy --uninstall.

Don't have Node.js? Install it once from nodejs.org (the LTS button). That's the only prerequisite.

B) Claude Desktop — add the meter (analytics in chat)

• Install the toolkit: npm install -g sipcode

• Open your Claude Desktop config file:

  Your OS	File to open
  Windows	%APPDATA%\Claude\claude_desktop_config.json
  macOS	~/Library/Application Support/Claude/claude_desktop_config.json
  Linux	~/.config/Claude/claude_desktop_config.json

• Add the sipcode block inside "mcpServers" (if the file is empty, paste the whole thing):

  Windows:

  {
    "mcpServers": {
      "sipcode": { "command": "cmd", "args": ["/c", "sipcode-mcp"] }
    }
  }
  

  macOS / Linux:

  {
    "mcpServers": {
      "sipcode": { "command": "sipcode-mcp" }
    }
  }
  

  💡 Windows needs the cmd /c part because of how Windows runs the program. Copy the Windows block exactly.

• Fully quit and reopen Claude Desktop. Then just ask Claude things like "use sipcode to show my token stats" or "turn on the sipcode proxy" — it will call the right tool.

How it works

Claude Code asks to run a command. Before it runs, Sipcode's hook swaps it for a leaner equivalent that returns the same information in fewer tokens. Because that output sits in Claude's context for the rest of the conversation, the saving compounds on every later turn.

WITHOUT Sipcode                          WITH Sipcode (proxy installed)

Claude ── "git status" ──▶ shell         Claude ── "git status" ──▶ [Sipcode hook]
       ◀── ~800 tokens ────                                              │ rewrites to
           (full output)                                                 ▼ "git status -s"
                                                            shell ──▶ runs the short form
                                                 Claude ◀── ~150 tokens (same info, compact)

Every rewrite is safe and reversible — Claude can always re-run the full command with explicit flags. Sipcode never touches commands that are already compact, already limited, or chained with &&, ||, ;, or |.

What it optimizes

Git

git status        →  git status -s              # short format
git log           →  git log --oneline -n 20    # 20 one-liners, not full history
git diff / show   →  … | head -200              # caps huge diffs

Files & search

ls, find          →  … | head -50 / -100        # caps long listings
grep -r           →  grep -c -r                 # per-file match counts
cat <file>        →  size-aware: full file if small, head+tail only if >300 lines

Package & build tools

npm ls / list     →  --depth=0                  # top-level deps only
cargo build/test  →  --quiet                    # drops "Compiling…" noise

Claude Code's built-in tools (an edge RTK doesn't have — these go through our hook too)

Grep tool         →  head_limit=50              # Claude Code's default is 250
Glob tool         →  head_limit=100             # caps path floods

There is intentionally no Read-tool rewrite: Claude Code's Read already caps at 2000 lines, so adding a limit would save nothing. We don't book savings we don't deliver.

Check that it's working

Claude Code (the valve)

sipcode proxy --diff      # preview exactly what the install will change (writes nothing)
sipcode proxy --stats     # after using Claude Code for a bit: shows what it rewrote + tokens saved

Quick manual proof: with the proxy installed, run git log in any repo with more than 20 commits — you'll get exactly 20 one-line entries instead of the full history. That's the rewrite firing.

Claude Desktop (the meter)

In a Claude Desktop chat, ask: "What sipcode tools do you have?" Claude will call get_sipcode_info and list them — if you see the version and tool list, the MCP server is wired up correctly. Then try "use sipcode to check the proxy status."

Troubleshooting

Symptom	Fix
Installed the proxy but tokens look the same	Did you restart Claude Code after sipcode proxy --install? Hooks load at startup. Then run sipcode proxy --stats after some use to confirm rewrites are happening.
sipcode: command not found	The global install didn't land on your PATH. Re-run npm install -g sipcode; if it persists, reopen your terminal. No Node.js? Install it from nodejs.org.
Claude Desktop says the sipcode server "is down" / won't start	99% of the time it's the config. On Windows you must use "command": "cmd", "args": ["/c", "sipcode-mcp"]. Confirm npm install -g sipcode succeeded, then fully quit and reopen Claude Desktop (not just close the window).
A command's output got cut off and I needed the rest	That's by design — re-run it with an explicit flag (e.g. git diff <file>, grep -rn …). If a rewrite ever hides something it shouldn't, that's a bug — please file it.
Want to turn it all off	sipcode proxy --uninstall (removes the hook, fully reversible).

The meter — see and prove your savings

Works in your terminal (CLI) and inside Claude Desktop (MCP tool, in parentheses):

sipcode why          # where your last session burned tokens   (audit_latest_session)
sipcode impact       # your spend before vs after Sipcode       (verify_sipcode_impact)
sipcode stats        # cross-session totals + top spenders      (get_session_stats)
sipcode estimate "…" # predict a task's cost per model          (estimate_task_cost)
sipcode score        # 24-check agent-friendliness audit        (get_agent_score)
sipcode proxy --stats# what the valve has rewritten             (get_proxy_stats)

The meter reads only the transcripts Claude Code already writes to ~/.claude/projects/. Zero network calls — your code never leaves your machine (Privacy).

Catch context drift

sipcode drift watches whether your recent Claude Code sessions are quietly getting more expensive or context-bloated — and stays silent unless something actually regressed. A smoke alarm, not a dashboard you have to babysit.

sipcode drift          # e.g. ⚠ cost/turn up ~34%; cache hit rate 71% → 48%
sipcode drift --json   # machine-readable   (also the get_drift_report MCP tool)

It compares your latest session against the median of your recent ones — conservative thresholds, so no false alarms — and names what moved: cost/turn up, cache-hit-rate down, or re-read waste up. These cost tokens and degrade answer quality (context rot), so catching them early keeps your agent both cheap and sharp.

The benchmark — the one number you can reproduce

npx sipcode benchmark

62.6% median token reduction across a locked 20-task corpus (range 37.4%–80.6%). This is the optimization methodology measured on captured transcripts — anyone can reproduce it in under 90 seconds, no signup. It is not a promise about your sessions (use sipcode impact for that) and not a measurement of the live proxy. Full method: benchmark/METHODOLOGY.md.

MCP tools (13, for Claude Desktop & Claude Code)

Meter (read-only): get_sipcode_info, verify_sipcode_impact, list_recent_sessions, audit_latest_session, get_project_manifest, estimate_task_cost, get_agent_score, get_session_stats, get_proxy_stats, get_proxy_status, get_drift_report Valve control (writes settings.json): install_proxy, uninstall_proxy

So a Desktop user never has to touch a terminal: ask Claude to install the proxy, check savings, audit a project, or predict a cost — all from chat. Full docs: docs/MCP.md.

What Sipcode does NOT do

Stating this plainly because trust is the product:

• It does not optimize the Claude Desktop chat itself — the valve only rewrites Claude Code tool calls. Desktop gets the meter.
• Installing the npm package alone changes nothing — you must run sipcode proxy --install to turn on the valve.
• The meter never reduces tokens — auditing is observation; only the valve reduces.
• No network calls from any core path — enforced by a CI test. Your code never leaves your machine.
• No guaranteed savings multiplier — it's workload-dependent; measure your own.
• It never silently loses data — every rewrite is recoverable by re-running with explicit flags.

Privacy

Local-first by engineering, not by promise. A CI test fails the build if node:http/https/net/dns is ever imported into a core path, and the proxy's rewriter functions are guarded to import no filesystem or network modules at all. Nothing is uploaded or phoned home.

Command reference

Command	What it does
sipcode proxy --install / --uninstall / --diff / --stats	The valve — runtime command optimizer
sipcode why	Where your last session burned tokens
sipcode impact	Before/after your own spend
sipcode stats	Cross-session analytics
sipcode estimate "<task>"	Per-model cost prediction
sipcode score	Agent-friendliness audit + badge
sipcode drift	Flag context/cost regressions vs your baseline (silent unless it regressed)
sipcode benchmark	Reproducible corpus (--hardest, --vs-rtk)
sipcode receipt	Shareable PNG of a session's savings
sipcode hygiene	Read-once rules + context-pressure hooks
sipcode rules	Output-compression rules in CLAUDE.md
sipcode manifest / init	Project manifest; multi-agent (--agent cursor)

Updating

npm install -g sipcode@latest

No auto-updater. Monthly is plenty unless a security note is pinned here or on Releases. The proxy hook points at your global install, so it picks up updates automatically — no need to re-run proxy --install.

Contributing & author

Bugs, edge cases, or a rewrite that hides something it shouldn't → open an issue. Built by Anuj Ojha — solo dev, also author of Answerable. MIT licensed.