tamp

# Tamp for OpenClaw Save 3-50% on input tokens by routing API requests through [Tamp](https://github.com/sliday/tamp) — a local HTTP proxy that compresses `tool_result` blocks before they reach Anthropic. ## Prerequisites - Node.js 18+ - Anthropic API key — Tamp proxies requests to Anthropic. Your `ANTHROPIC_API_KEY` must be set in your OpenClaw config. Tamp itself does not store or read this key — it forwards the `x-api-key` header from incoming requests unchanged. ## 1. Install & Run ```bash # Install a pinned version (recommended) npm i -g @sliday/tamp@0.3.8 # Or run without installing npx @sliday/tamp@0.3.8 -y ``` Start with default stages: ```bash TAMP_STAGES=minify,toon,strip-lines,whitespace,dedup,diff,prune tamp -y ``` Verify: ```bash curl http://localhost:7778/health # {"status":"ok","version":"0.3.8","stages":["minify","toon",...]} ``` > Tamp is [open source (MIT)](https://github.com/sliday/tamp). Audit the source, build from git, or run from a local clone: `git clone https://github.com/sliday/tamp && cd tamp && npm install && node bin/tamp.js -y` ## 2. Run as systemd service Create `~/.config/systemd/user/tamp.service`: ```ini [Unit] Description=Tamp token compression proxy After=network.target [Service] # Adjust path: node -e "console.log(require('child_process').execFileSync('which', ['tamp']).toString().trim())" ExecStart=/usr/local/bin/tamp Restart=always RestartSec=5 Environment=TAMP_PORT=7778 Environment=TAMP_STAGES=minify,toon,strip-lines,whitespace,dedup,diff,prune Environment=TAMP_LOG=true [Install] WantedBy=default.target ``` ```bash systemctl --user daemon-reload systemctl --user enable --now tamp.service journalctl --user -u tamp -f # live compression logs ``` ## 3. Configure OpenClaw Add a provider in your OpenClaw config: ```json5 { models: { providers: { "anthropic-tamp": { baseUrl: "http://localhost:7778", apiKey: "${ANTHROPIC_API_KEY}", // Forwarded to upstream, not stored by Tamp api: "anthropic-messages", models: [ { id: "claude-opus-4-6", name: "Claude Opus 4.6 (compressed)" }, { id: "claude-sonnet-4-6", name: "Claude Sonnet 4.6 (compressed)" } ] } } } } ``` Set as primary model: ```json5 { agents: { defaults: { model: { primary: "anthropic-tamp/claude-opus-4-6" } } } } ``` Restart the gateway. All requests now flow through Tamp. ## How it works ``` OpenClaw → POST /v1/messages → Tamp (localhost:7778) → compresses JSON body → Anthropic API ← streams response back unchanged ``` Tamp intercepts the request body, finds `tool_result` blocks in `messages[]`, and compresses their content. Headers (including `x-api-key`) are forwarded unchanged. The response streams back untouched. ## 7 Compression Stages | Stage | What it does | Lossy? | |-------|-------------|--------| | minify | Strip JSON whitespace | No | | toon | Columnar encoding for arrays (file listings, deps, routes) | No | | strip-lines | Remove line-number prefixes from Read tool output | No | | whitespace | Collapse blank lines, trim trailing spaces | No | | dedup | Deduplicate identical tool_results across turns | No | | diff | Delta-encode similar re-reads as unified diffs | No | | prune | Strip lockfile hashes, registry URLs, npm metadata | Metadata only* | \* Prune removes fields like `integrity`, `resolved`, `shasum`, `_id`, `_from`, `_nodeVersion` from JSON — npm registry metadata not needed by the LLM. To keep full provenance, remove `prune` from `TAMP_STAGES`. ## What to expect | Scenario | Savings | |----------|---------| | Chat sessions (short turns) | 3-5% | | Coding sessions (file reads, JSON) | 30-50% | | Lockfiles | up to 81% | | Subagent tasks | 20-40% | ## Security Notes - **API key handling:** Tamp forwards the `x-api-key` / `Authorization` header from the incoming request to upstream. It does not store, log, or read API keys. - **Local only:** Tamp binds to localhost by default. It does not accept external connections unless you change the bind address. - **No telemetry:** Tamp does not phone home, collect analytics, or make any outbound connections except to the configured upstream API. - **Fallback:** Add Anthropic direct as a fallback model in OpenClaw. If Tamp is down, requests bypass it automatically. ## Resources - ~70MB RAM, <5ms latency per request - No Python needed — all 7 stages run in Node.js - [Source (MIT)](https://github.com/sliday/tamp) · [tamp.dev](https://tamp.dev) · [white paper](https://tamp.dev/whitepaper.pdf)