สรุป · Sarup

Thai-first context compression for Claude Code. An MCP server that actually shrinks Thai (50–78% fewer tokens) and caches every original so nothing is ever lost. Measured for what matters too: whether Claude still answers correctly.

         

สรุป means "to summarize." Headroom routes Thai through noop (0% savings) because its whitespace tokenizer can't find Thai word boundaries. Sarup uses PyThaiNLP segmentation, so it compresses Thai as well as English — and caches every original so nothing is ever lost.

Contents

• Highlights

• Why it's safe — the two-tier guarantee

• How it works

• Tools

• Compression modes

• Measured results

• Example

• Install

• Register with Claude Code

• Auto-compression hook

• Privacy & data

• Configuration

• Project structure

• Tech stack & techniques

• Testing

• Roadmap

• License

Related MCP server: cctx

Highlights

• 🇹🇭 Real Thai compression — PyThaiNLP newmm word segmentation, not whitespace.

• ♻️ Lossless by guarantee — every compress caches the original; verified: true proves a byte-for-byte round-trip.

• 🎚️ Five modes: from offline 1 ms TF-IDF to a 78%-savings cascade.

• 🧠 Optional local LLM — embeddings + rewrite via Ollama, with automatic offline fallback.

• 📏 Honest metrics: a real tokenizer (tiktoken) at runtime; Claude's own tokenizer and a Thai QA eval for the published numbers (Does Claude still get it right?).

• 🕳️ Visible gaps: […] marks every place sentences were dropped, so Claude retrieves instead of mistaking a nearby fact for the answer.

• 🔌 Content-aware — JSON compaction, log dedup, and verbatim code-fence preservation built in.

• 🛟 Can't break Claude — it's an MCP tool, not an API proxy; if the server is down the tools just go away and Claude keeps working.

Why it's safe — the two-tier guarantee

Aggressive lossy compression is safe because the original is always one sarup_retrieve(hash) away. This is how "maximum savings" and "100% accuracy" coexist — they live in different tiers.

How it works

Two entry points feed one engine: a cheap compressed view the model reads, and a lossless retrieval store that can restore the original byte-for-byte.

flowchart TD
    M["🧑 Manual<br/>sarup_compress()"]:::entry --> R
    A["⚙️ Automatic<br/>PostToolUse hook<br/>(Read · Bash · Grep)"]:::entry --> R

    R{"Sarup compress<br/>extractive · semantic · abstractive · pipeline"}:::engine
    R -- "compressed view<br/>50–78% fewer tokens" --> V["📄 Model context"]:::lossy
    R -. "cache original" .-> S[("🗄️ Retrieval store<br/>hash → original")]:::lossless

    V -. "need full detail?" .-> RET["🔑 sarup_retrieve(hash)"]:::lossless
    RET --> S
    S == "byte-for-byte ✓" ==> V

    classDef entry fill:#e0e7ff,stroke:#6366f1,color:#111
    classDef engine fill:#fde68a,stroke:#d97706,color:#111
    classDef lossy fill:#fef3c7,stroke:#f59e0b,color:#111
    classDef lossless fill:#bbf7d0,stroke:#16a34a,color:#111

• Manual — the model calls sarup_compress / sarup_retrieve itself.

• Automatic — the hook intercepts large tool outputs, caches the original to SARUP_DB_PATH, and substitutes the compressed view + a retrieval hash. Source code is skipped; small outputs pass through untouched.

Tools

sarup_compress arguments

Compression modes

¹ Measured on a 10-sentence Thai paragraph (522 tokens), rewrite model gemma4:12b. Every mode stays 100% recoverable via the store; Ollama modes degrade gracefully to extractive when the backend is down. (gemma4:e4b: pipeline 82.2%; June 2026, gemma3:12b: 88.1%.)

Extractive and semantic output is a verbatim subset of the input sentences, with […] on its own line wherever sentences were dropped. Sentences carrying specifics (numbers, Thai digits, identifiers like X-Api-Key) score higher than general statements, since those are what questions ask about.

Measured results

$ .\.venv\Scripts\python.exe bench\benchmark.py

sample                      before   after   savings   verify
Thai prose                     522     247    52.7%       OK
Thai prose (aggressive)        522     189    63.8%       OK
English prose                  105      60    42.9%       OK
JSON (lossless)                 67      44    34.3%       OK
Logs                           563     300    46.7%       OK
TOTAL                         1779     840    52.8%    ALL OK   → 100% recoverable

Mode comparison (Thai prose, 522 tok):
  extractive 52.7% (1ms) · semantic 63.8% (2.1s) · auto 63.8% (2.1s)
  abstractive 44.4% (8.3s) · pipeline 78.2% (5.8s)      ← gemma4:12b

Token counts via tiktoken cl100k_base. That is OpenAI's tokenizer and undercounts Claude, most of all on Thai; bench\benchmark.py --claude adds a column from Claude's own tokenizer (Anthropic count_tokens, needs an API key).

Does Claude still get it right?

Fewer tokens only help if the facts survive. bench\eval_qa.py compresses 8 Thai documents (incident report, meeting notes, API docs, HR policy, product spec, support chat, README, sales report; half line-per-sentence, half dense paragraphs) and asks Claude Opus 5.5 32 questions from the compressed view alone, graded deterministically:

$ .\.venv\Scripts\python.exe bench\eval_qa.py        # uses your local `claude` login

mode               claude tok   saved  correct  abstain  wrong
──────────────────────────────────────────────────────────────
original                 4601    0.0%     100%       0%      0
extractive:0.5           3204   30.4%      69%      31%      0
extractive:0.3           1969   57.2%      44%      56%      0
semantic:0.5             3214   30.1%      62%      38%      0
extractive:0.5+q         3207   30.3%     100%       0%      0
abstractive:0.5          2697   41.6%      91%       9%      0     ← gemma4:12b
pipeline:0.5             1854   59.8%      59%      41%      0     ← gemma4:12b

(gemma4:e4b scored the same: abstractive 91%, pipeline 59%, wrong 0.)

• wrong = 0 is the safety property: when a fact was dropped, Claude said so (abstain) and in real use would call sarup_retrieve. It did not guess from what was left.

• abstain is the real cost of blind compression: about a third of specific questions need a retrieval round-trip at 0.5. Know what you're looking for? Pass it as query (+q row).

• abstractive keeps the most facts per token (91% at 42% saved) because a rewrite can keep every number while dropping words; extraction can only drop whole sentences. It costs seconds per document, so use it through sarup_compress, not the hook.

• In both evals no paraphrase changed a fact into a wrong one, but 8 documents can't prove a rewrite never will, which is why the original stays one sarup_retrieve away.

• Token counts here are Claude's own (from the answer calls' usage) and include the ~40-token [sarup: …] footer, which weighs heavily on these short (~570-token) documents and fades on the large outputs the hook actually compresses.

• 8 documents is a small sample; differences of a couple of questions between modes are noise.

Before the fixes this eval drove (Oct 2026): dense Thai paragraphs were not compressed at all (the sentence splitter's model wasn't installed), extractive:0.5 answered 59% with one confident wrong answer, and semantic:0.5 47%.

Example

A real sarup_compress call on a Thai paragraph (mode="auto", Ollama up → semantic):

// → sarup_compress(content="…518-token Thai paragraph…", mode="auto")
{
  "compressed": "จุดเด่นที่สำคัญที่สุดคือมันไม่มีทางทำให้ Claude พัง…",
  "hash": "caa568140bec0ff734937cf5",
  "original_tokens": 518,
  "compressed_tokens": 154,
  "tokens_saved": 364,
  "savings_percent": 70.3,
  "transforms": ["semantic_extractive", "embeddings", "thai"],
  "lossy": true,
  "verified": true,                    // round-trip proven byte-for-byte
  "token_method": "tiktoken:cl100k_base"
}

The model keeps working on the 154-token view; the full 518-token original is one call away:

// → sarup_retrieve(hash="caa568140bec0ff734937cf5")
{ "content": "…the exact original text, restored byte-for-byte…" }

Install

As a Claude Code plugin (easiest)

Needs uv: the plugin runs Sarup through uvx, which fetches Python and the dependencies on first use. Nothing to clone. In Claude Code:

/plugin marketplace add PHUICMT/sarup
/plugin install sarup@sarup

That installs the MCP server, the auto-compression hook and the /sarup:sarup-setup skill. The cache lives in the plugin's data directory, so it survives plugin updates. The first start takes a few seconds while uvx builds the environment. From a shell instead: claude plugin marketplace add PHUICMT/sarup, then claude plugin install sarup@sarup.

Ollama is optional. Without it everything works offline: the hook and the default extractive mode never touch it, and asking for semantic, abstractive or pipeline falls back to extractive with a setup field that names what to install. sarup_stats shows what is available, and /sarup:sarup-setup walks through adding Ollama and the models below.

From PyPI (any MCP client)

The package is sarup on PyPI. With uv, nothing to install by hand; point any MCP client at uvx sarup:

{
  "mcpServers": {
    "sarup": {
      "command": "uvx",
      "args": ["sarup"],
      "env": { "SARUP_DB_PATH": "/home/me/.sarup-cache.db" }
    }
  }
}

In Claude Code without the plugin: claude mcp add sarup -s user -- uvx sarup. The hook is uvx --from sarup sarup-hook --db <same file as SARUP_DB_PATH>. pip install sarup works too and puts sarup and sarup-hook on PATH.

From a clone

One command (creates the venv, installs everything, registers the MCP server for all projects, idempotent):

.\scripts\setup.ps1 -All      # Windows  (-All also adds the hook, the /sarup-setup skill, pulls Ollama models)
./scripts/setup.sh --all      # Linux / WSL / macOS

Tip: -All/--all installs a global /sarup-setup skill, so on any other machine you can just type /sarup-setup in Claude Code and it walks through the install. (Or run scripts/install-skill.ps1 / install-skill.sh on its own.)

Uninstall just as cleanly (only removes what Sarup added; -Purge/--purge also deletes the venv + cache):

.\scripts\uninstall.ps1       # Windows
./scripts/uninstall.sh        # Linux / WSL / macOS

py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[dev]"

Optional local-LLM modes (semantic / abstractive / pipeline) need Ollama:

ollama pull nomic-embed-text     # embeddings → semantic mode
ollama pull gemma4:12b           # rewrite → abstractive / pipeline, 7.6 GB (or gemma4:e4b / gemma3:12b)

Register with Claude Code

One-command setup (recommended). Detects this machine's paths, probes Ollama (picks the best mode + models), and merges into .mcp.json / .claude/settings.json without clobbering anything already there (a .bak is written first):

.\.venv\Scripts\python.exe scripts\install.py --with-hook --pull

• No Ollama? It configures offline extractive mode — still fully works.

• Ollama up? It auto-selects nomic-embed-text (semantic) + the best installed rewrite model (gemma4:12b → gemma4:e4b → gemma3:12b → qwen3) for sarup_compress. The hook stays extractive either way (it runs on every tool call). --pull fetches any missing models.

• Idempotent — safe to re-run; --global writes to ~/.claude instead.

Manual — or add it yourself to your MCP config (e.g. .mcp.json or ~/.claude.json). Replace <SARUP_DIR> with the absolute path where you cloned this repo (the installer above fills these in for you):

{
  "mcpServers": {
    "sarup": {
      "command": "<SARUP_DIR>/.venv/Scripts/python.exe",
      "args": ["-m", "sarup.server"],
      "env": { "SARUP_DB_PATH": "<SARUP_DIR>/.sarup-cache.db" }
    }
  }
}

On Linux/macOS the interpreter is <SARUP_DIR>/.venv/bin/python.

Or run it directly over stdio:

.\.venv\Scripts\python.exe -m sarup.server

Auto-compression hook

Skip manual tool calls entirely: install the PostToolUse hook and large Read/Bash/Grep outputs are compressed before they enter context, with the original cached for retrieval. Source-code reads are skipped for safety. Full setup in hooks/README.md.

Fixed in October 2026. Earlier versions returned the compressed text as a plain string. Claude Code only accepts updatedToolOutput in the tool's own shape ({stdout, stderr, …} for Bash, {type, file: {content, …}} for Read) and silently dropped the string, so the hook never applied on any surface. It now keeps each tool's shape; verified on the CLI (2.1.289): the model receives the compressed view and the retrieval hash. To see what the hook does on your build, set SARUP_HOOK_LOG to a file path (one JSON line per call).

The plugin wires this up for you. From a clone, replace <SARUP_DIR> with your clone path, or run install.py --with-hook:

{
  "hooks": {
    "PostToolUse": [
      { "matcher": "Read|Bash|Grep",
        "hooks": [{ "type": "command",
          "command": "<SARUP_DIR>/.venv/Scripts/sarup-hook --db <SARUP_DIR>/.sarup-cache.db" }] }
    ]
  }
}

Privacy & data

To guarantee recovery, Sarup caches the original content in the store. Two things to know:

• With SARUP_DB_PATH set, originals are written to that SQLite file in plaintext (no encryption). Treat it like a cache of whatever you compressed.

• If you compress tool outputs that contain secrets (e.g. a .env dump or credentials in a log), those land in the cache too. The auto-hook skips source-code/config file reads, but Bash output is fair game — review what you point it at.

*.db is git-ignored, so the cache never gets committed. For zero on-disk footprint, leave SARUP_DB_PATH unset (memory-only; the MCP server then loses the cache on restart, and the hook will not substitute — see the hook docs).

Configuration

Project structure

sarup/
├── src/sarup/
│   ├── server.py       # MCP stdio server — 3 tools
│   ├── compressor.py   # router + modes (extractive/semantic/abstractive/pipeline/auto)
│   ├── thai.py         # PyThaiNLP tokenization, sentence split, TF-IDF
│   ├── semantic.py     # embedding centrality + cosine dedup
│   ├── llm.py          # optional Ollama backend (generate + embed)
│   ├── tokens.py       # real token counting (tiktoken)
│   └── store.py        # CCR store: hash → original (memory + SQLite)
├── hooks/
│   ├── sarup_hook.py   # PostToolUse auto-compression hook
│   └── README.md       # hook install guide
├── bench/benchmark.py  # before/after measurement
├── tests/              # test_thai, test_mcp, test_hook, ...
├── README.md
└── STACK.md            # full stack + techniques

Tech stack & techniques

Python 3.11 · MCP · PyThaiNLP newmm · tiktoken · Ollama (optional) · SQLite · hatchling · pytest.

The technique behind each mode — TF-IDF scoring, embedding centrality, cascade pipeline, content routing, and graceful degradation — is documented in STACK.md.

Testing

.\.venv\Scripts\python.exe -m pytest tests/ -q

The suite covers Thai NLP, the MCP tool contracts, every mode (including Ollama-fallback paths), the roundtrip-verify guarantee, and the auto-compression hook (incl. cross-process retrieval).

Quality is measured separately, since it needs a model: bench\eval_qa.py (see Does Claude still get it right?). It runs on the local claude CLI by default, or --backend api with an Anthropic API key; --dry-run and --limit keep a check cheap; a full run is ~170 Claude calls.

Releasing

1. Bump version in pyproject.toml and server.json (both places in it).

2. Push a matching tag: git tag v0.2.0 && git push origin v0.2.0. The publish workflow tests, builds and uploads to PyPI through trusted publishing (no token).

3. List the release in the MCP registry: mcp-publisher login github, then mcp-publisher publish (reads server.json; the mcp-name comment at the top of this README proves the PyPI package belongs to it).

The plugin needs none of this: it installs from the repo and updates with each commit.

Roadmap

• Grow the QA eval past 8 documents (long tool outputs, logs, mixed code) before changing defaults such as auto → semantic on the strength of it.

• Move to mcp 2.x (pinned <2 for now: 2.x dropped Server.list_tools).

• Optional Typhoon 2.1 abstractive (blocked on an Ollama template fix).

• Per-content adaptive target_ratio.

• Published PyPI package (sarup) and MCP registry listing (io.github.PHUICMT/sarup).

• Installable as a Claude Code plugin.

License

MIT