Token compression
Reduces tokens by up to 70%. Lower costs, higher speed.
Available now
MCP
Python 3.11+
Synthelion
Compress Protect Optimize
The universal token compressor and prompt manager for AI agents. Less tokens. Lower costs. Faster responses. More value.
Compresses prompts before they reach any model, protects personal data and governs your agents. Works with Claude Code, OpenAI, LangChain, Cursor and more.
Up to
70%
token savings
50+languages
46MCP tools
Why use many tokens when few do the trick?
70%fewer tokens
50+supported languages
46MCP tools
0ML models required
Highlights
Everything you need to spend less and protect more.
Multi-language
Supports 50+ languages. No configuration needed.
Zero ML
Works without AI models. Fully offline and deterministic.
Content routing
Optimizes every content type: code, JSON, logs, HTML and more.
Adaptive compression
Adjusts the level to the content. Always the right balance.
PrivacyGuard
Detects and masks sensitive data. Your privacy, first.
EnterpriseGuard
Blocks secrets and critical content. Protects your business.
Agent guardrails
Rules for each agent type. More control, less risk.
AI Compliance
Manages regulatory controls. Always compliant.
MCP native
46 MCP tools. Compatible with all agents.
Multi-agent
Works with many agents in a single solution.
Local proxy
Protects and optimizes calls. Any agent, any provider.
Provider failover
Switches provider automatically. Maximum uptime.
Circuit breaker
Handles errors and rate limits. Keeps you online.
Web dashboard
Full control in real time. Monitor, analyze, act.
Security & sustainability
No data sent online. Less resource consumption.
More securityfor your data
Lower costsfor your agents
Full compatibilitywith your ecosystem
Sustainablefor the future
Synthelion. The AI power, in your hands.
Why Synthelion
Every token costs. Remove the ones that do nothing.
Synthelion removes words that carry no meaning and reduces inflected words to their base form: the model gets the same information, without the grammatical packaging.
Deterministic compression
No ML model by default, no network calls. JSON, HTML, diffs, logs, code and prose each get their own strategy.
PrivacyGuard
51 detection rules across 33 countries with real checksum validation: IBANs, cards, tax IDs. Masking with recoverable placeholders.
EnterpriseGuard
An outbound data-loss firewall: controls what leaves your machine toward providers.
Agent guardrails
A per-agent-type policy engine, plus prompt-injection screening before untrusted text reaches the context.
Compliance engine
GDPR, AI Act, NIS2, PCI-DSS and NIST 800-53 flags attached automatically, with a localized AI transparency notice.
Local proxy and dashboard
Works even where MCP and hooks don't: a local reverse proxy with multi-provider failover, a dashboard and cluster deployment.
Before
I would like to know if it is possible to receive information about cheap restaurants in Rome, please.
20 tokens
After
like know possible receive information about cheap restaurant Rome
9 tokens · -55%
Dashboard
Everything under control, locally.
In preparation
Synthelion 1.2.6: the graphical interface is coming.
In version 1.2.6 we are preparing a desktop application with a graphical interface for Windows, Linux and macOS: configure compression, privacy and providers without touching the terminal.
WindowsComing with 1.2.6
LinuxComing with 1.2.6
macOSComing with 1.2.6
Installation
Ready in under a minute.
Requirements: Python 3.11+ (from python.org, tick "Add to PATH"). Pick your system.
One command (recommended)
Download the script: it installs Synthelion, detects Python, configures MCP for Claude Code and sets up the auto-compression hook.
Invoke-WebRequest https://raw.githubusercontent.com/francescopaolopassaro/synthelion/main/install_claude.ps1 -OutFile install_claude.ps1
powershell -ExecutionPolicy Bypass -File install_claude.ps1Or manual
pip install synthelion
synthelion compress --text "Hello world, how are you today?" --jsonIf `synthelion` is not recognised, close and reopen the terminal (PATH refresh).
One command
curl -fsSL https://raw.githubusercontent.com/francescopaolopassaro/synthelion/main/install_claude.sh | bashOr manual
pip3 install synthelion
synthelion compress --text "Hello world, how are you today?" --jsonRecommended with uv: no PATH issues.
pip install uv
uvx synthelion-mcpOne command
curl -fsSL https://raw.githubusercontent.com/francescopaolopassaro/synthelion/main/install_claude.sh | bashOr in a virtualenv
python3 -m venv ~/.venvs/synthelion
source ~/.venvs/synthelion/bin/activate
pip install synthelion
synthelion compress --text "Hello world, how are you today?" --json
echo 'export PATH="$HOME/.venvs/synthelion/bin:$PATH"' >> ~/.bashrcZero install (all platforms)
uv runs Synthelion in an isolated environment, always at the latest version.
pip install uv
uvx synthelion-mcp-
1
Register Synthelion in Claude Code
synthelion install # ~/.claude.json (globale) synthelion install --local # solo progetto -
2
Restart Claude Code
Close and reopen the window (or run `claude` again). Synthelion is now an MCP tool.
-
3
Verify
Type in Claude Code: “Use Synthelion to compress this…”. Claude will call the tool and return the compressed version.
Update
pip install --upgrade synthelion
synthelion --version
Compared
Synthelion versus other compression tools.
A capability comparison, not a performance benchmark: we do not publish head-to-head token-savings numbers because we have not run other projects' code in a controlled, reproducible setup. The table is based on reading each project's public source.
| Synthelion | headroom | kompact | tokenless | squeez | bu-ketao | |
|---|---|---|---|---|---|---|
| Zero ML models, zero network calls | No (torch/transformers/SigLIP, fine-tuned classifiers) | Optional ONNX embedding | n/a (prompt rules, no code) | |||
| NLP compression, 50+ languages | — (ML feature extraction instead) | Yes (TF-IDF) | – | – | n/a | |
| Content-type auto-routing (JSON/HTML/diff/log/code) | Yes (shape-based) | – | n/a | |||
| Tool-schema to compact signature | – | Yes (TF-IDF selection) | – | n/a | ||
| Credential-shape detection before persisting | – | – | – | Yes (origin) | n/a | |
| Terminal ANSI/noise cleanup + success collapse | – | – | – | Yes (origin) | n/a | |
| Masking old tool output + retrieval (Artifact Index) | – | Yes (origin) | – | – | n/a | |
| Diff-on-repeat for identical tool calls | – | – | – | – | n/a | |
| Adaptive compression scaling by content size | – | Yes (origin) | – | – | n/a | |
| JSON chain-depth / dot-path collapsing | – | – | Yes (origin) | – | n/a | |
| Advisory command-rewrite (never executes) | – | – | Executes it (rtk wrapper) | – | n/a | |
| Local multi-page web dashboard | Yes (simpler) | – | – | n/a | ||
| Cluster / master-slave deployment | – | – | – | – | n/a | |
| MCP protocol (Claude Code, etc.) | Yes, 46 tools | Yes (CLI plugins: Claude/Codex/Gemini) | No (HTTP proxy instead) | Yes (hooks) | – | n/a |
| Vision/image token optimization | No (text-only) | Yes (tile-aligned resize + trained router) | – | – | – | n/a |
| Provider cache-breakpoint-aware read staleness | Yes (origin) | – | – | – | n/a | |
| Response-style compression (output-side, CJK-aware) | – | – | – | – | Yes (origin) | |
| PII detection + masking (33 countries, GDPR/AI Act compliance) | – | – | – | – | – | |
| Prompt-injection guard (jailbreak/instruction-override screening) | – | – | – | – | – | |
| Local reverse proxy (any agent, any provider) | Yes (origin) | Yes (proxy instead of MCP) | – | – | n/a | |
| Proxy: automatic failover across backup providers (up to 10) | Not found in public docs | – | – | – | n/a | |
| Proxy: circuit breaker on repeated rate-limit/5xx | Not found in public docs | – | – | – | n/a | |
| Proxy: WAF/firewall protecting the proxy surface itself | – | – | – | – | n/a | |
| Proxy: enforced block-on-risk (request never reaches the provider) | – | – | – | – | n/a | |
| Proxy request log is metadata-only: no prompt/response ever stored | No (stores content when --log-messages is on) | – | – | – | n/a |
headroom. The broadest-scope alternative we looked at: proxy architecture, persistent memory, vision-token optimization and ML-based response-length prediction. That breadth comes at the cost of the zero-ML guarantee Synthelion makes: several of its decisions depend on downloaded, trained models rather than deterministic heuristics.
Not in the table: tokensave, a Rust code-intelligence tool that solves a different problem (structural understanding of a codebase) and is not like-for-like with context compression. If you have measured a real head-to-head, open an issue with your methodology and we will link it.
Compatibility
Works with any agent.
Claude CodeOpenAILangChainCursorOpenCodeAiderCodex CLIMCP
Open tooling, models on Hugging Face.
Code, issues and releases on GitHub. SynthelionML and PrivacyGuardML models on Hugging Face.