TokenSlim

High-performance Rust token compression engine for LLM inputs.
Plugin-based · 50–95% token savings · AI-export diagnostics · CLI / Server / IDE / SDK

What is TokenSlim · Why · Features · Installation · Usage · Plugins · Integrations · License

English · 简体中文 · 日本語 · 한국어 · Español · Français · Deutsch · العربية

What is TokenSlim?

TokenSlim is a high-performance, plugin-based text compression engine written in Rust. Its core mission is to dramatically reduce the token cost of LLM inputs and to make it possible to fit long, noisy real-world logs (build pipelines, CI runs, web access logs, database traces, cloud logs, VCS output, stack traces, etc.) into LLM context windows — without losing the diagnostic signals the model needs.

On highly structured, repetitive inputs (compiler logs, build output, CI logs, access logs, etc.), TokenSlim typically delivers 50%–90% reduction while preserving 100% of the original information. In its AI Export mode, designed specifically for LLM consumption, the reduction reaches 90%–95% with context-aware denoising that keeps the error/warning context the model needs to reason about.

Beyond compression, TokenSlim ships with environment-diagnostic tooling (workspace, encoding, rule, env commands) that auto-detects OS, shell, code page, Python/Node/JDK encoding configuration, flags mojibake risk, and emits actionable fixes. Combined with a subprocess decoding fallback chain (UTF-8 first, codepage candidates next), it stays reliable across mixed-language environments.

See It in Action

Real-world daily usage — tokenslim gain

This is what tokenslim gain looks like after months of daily use on git commands:

$ tokenslim gain

TokenSlim Cumulative Savings Report
====================================

Usage Statistics:
  Total runs:          7,244
  Input tokens:        13.2M
  Output tokens:       9.4M
  Tokens saved:        3.9M
  Overall compression: 29.3%

Estimated Savings:
  Tokens saved:        3,883,551 tokens
       claude-4.8:     $19.42 USD ($5.00/1M)
       gpt-5.5:        $19.42 USD ($5.00/1M)
       gemini-3.1-pro: $7.77 USD  ($2.00/1M)

💡 tokenslim gain tracks every compression you run and shows cumulative savings. The numbers above are from a single developer's daily workflow — your team's savings multiply from here.

Compression varies by input type

Not all inputs compress equally — and that's expected. Highly repetitive, structured logs compress much more than information-dense content like git diffs:

The overall range is 20–95% depending on how repetitive and structured your input is. Use tokenslim gain to track your real savings over time. Before — git status (22 lines, ~680 characters):

$ git status
On branch master
Changes to be committed:
  (use "git restore --staged <file>..." to unstage)
        modified:   .gitignore
        modified:   src/core/dictionary_engine/test.rs
        modified:   src/plugins/cloud_log_plugin/test.rs

Changes not staged for commit:
  (use "git add <file>..." to update what will be committed)
  (use "git restore <file>..." to discard changes in working directory)
        modified:   Cargo.toml
        modified:   resources/messages.zh-CN.json
        modified:   src/bin/tokenslim-server.rs
        modified:   src/core/plugin_config_loader/mod.rs

Untracked files:
  (use "git add <file>..." to include in what will be committed)
        tests/server_webui_e2e.rs
        webui/

After — tokenslim git status (8 lines, ~280 characters — same information, zero loss):

git status
BR:master
M .gitignore
M src/core/dictionary_engine/test.rs
M src/plugins/cloud_log_plugin/test.rs
M Cargo.toml
M resources/messages.zh-CN.json
M src/bin/tokenslim-server.rs
M src/core/plugin_config_loader/mod.rs
? tests/server_webui_e2e.rs
? webui/

Every developer runs git status dozens of times a day. TokenSlim strips the boilerplate hints, unifies the status markers, and delivers the same information in ~60% fewer tokens — and this adds up across thousands of LLM interactions.

Why TokenSlim?

1. Real money saved

LLM API cost is dominated by input token count. TokenSlim cuts that by 50–95%:

• Lower API bills — 50–95% fewer input tokens.
• Context-aware AI Export (--ai-export) — strips routine lines, keeps the error/warning window the model actually needs; reduces hallucination on noisy inputs.
• Longer effective context — same context window, more real signal.
• Faster prefill — shorter inputs usually mean faster model prefill and lower TTFT.

2. Industrial-grade performance

• Zero-copy pipeline — built on Rust Cow<'a, str>, parallel block processing with rayon, and Bump arena allocation. Processes 100 MB of industrial-grade log in ~250 ms, ~400 MB/s throughput.
• Deterministic global reordering — a streaming build-target tracker fixes the out-of-order interleaving produced by make -jN / Ninja. Two identical parallel builds always produce the same error stack order.
• Sidecar mode — high-throughput REST API server, embeddable into IDE / CI / Agent workflows with zero startup overhead.

3. Data-driven extraction

• Radix-trie path extraction — TokenSlim does not slice line-by-line. After scanning 100 MB of input, it builds a project-wide radix trie in memory and only emits directory dictionaries ($D) on hot branches (weight > 10), eliminating fragmentary tokens.
• Semantic markers — environment-aware substitutions for Android, iOS, GCC, MSVC, and linkers.
• Full build ecosystem detection — C/C++, Rust, Go, Java, Android, iOS/Xcode, MSVC, Swift, and major linkers, with context-aware folding and error deduplication.

Features

• Three runtimes
  • CLI — scriptable batch processing
  • Server — long-lived REST API for full ecosystem integration
  • SDKs — Java, Python (PyO3), Node.js
• Plugin ecosystem (60+ plugins covering the most common LLM-input sources)
  • Mobile — android_gradle, xcode_log
  • General dev — gcc_log, java_stack, python_traceback, dotnet, rust_go, maven, gradle, node_error, nodejs, php_ruby, unity_unreal
  • Structured data — json, yaml, xml_html, ndjson, protobuf
  • Build artifacts — artifact_summary (SARIF / JUnit XML), with semantic preservation of test status, SARIF level/rule/location/tool
  • Cloud & ops — cloud_log (AWS / GCP / Azure / Alibaba / OCI / Tencent / Huawei / Cloudflare), web_log (Nginx / Apache / ingress / Envoy / CloudFront / IIS / ALB / Cloudflare), db_log (PostgreSQL / MySQL / MongoDB / Redis), syslog
  • CI/CD — ci_log (GitHub Actions / GitLab CI / Jenkins / Azure Pipelines / CircleCI / Buildkite / local act / TeamCity / Travis CI)
  • VCS — unified vcs_plugin for git / svn / hg / p4 / cvs / bzr / fossil / darcs, plus git_diff, smart_code (AST-level), smart_path
• Environment diagnostics — workspace, encoding, rule, env subcommands detect mojibake risk and emit fix recipes.
• AI-native output modes
  • --ai-export — context-aware denoising, keeps error/warning windows
  • --ai-signal — lossy but high-signal, preserves the most decision-relevant fields
• Plugin introspection — tokenslim explain-plugin and tokenslim run --explain-route explain route selection, fallbacks, confidence, alternatives, and replay misclassifications for audit.

Installation

One-liner install (any platform — recommended)

# Project-local
npm install tokenslim

# Or globally so `tokenslim` / `tokenslim-server` are on PATH
npm install -g tokenslim

tokenslim ships 6 platform-specific optionalDependencies (@tokenslim/cli-binary-linux-x64-gnu, …-linux-arm64-gnu, …-darwin-x64, …-darwin-arm64, …-windows-x64, …-windows-arm64). npm/pnpm/yarn automatically installs the one matching your OS + CPU, pulling the tokenslim + tokenslim-server binaries and 60+ plugin configs into node_modules/. A small Node wrapper at bin/tokenslim.js then forwards each call to the real binary.

If network access is unavailable and the optional package fails to install, the postinstall script transparently falls back to downloading from GitHub Releases. If that also fails, the install still succeeds — only the CLI commands become unavailable; the JS SDK keeps working as a REST client.

From source (Rust toolchain ≥ 1.75)

git clone https://github.com/nuoyazhizhou/tokenslim.git
cd tokenslim
cargo build --release

The binaries land at ./target/release/tokenslim and ./target/release/tokenslim-server (or *.exe on Windows).

Prebuilt binaries (no Node)

Download both binaries from the Releases page.

Configuration (optional)

All runtime configuration goes through environment variables. Copy .env.example to .env and fill in your local values. .env is git-ignored by default; only the example template is tracked.

Most users only need RUST_LOG=info (or debug for verbose tracing). The LLM-audit related variables (OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODEL) are only required if you run scripts/audit_*.py --llm-audit — without them, audits degrade to lint-only mode.

Editor / IDE integrations

• VS Code — see vscode-extension/
• Chrome — see chrome-extension/
• JetBrains — see jetbrains-plugin/

SDKs

• Node.js / TypeScript — npm i tokenslim (source: packages/sdk-nodejs/). Ships the tokenslim + tokenslim-server binaries as optionalDependencies — see Installation above.
• Python — see sdk/python/tokenslim_sdk.py (single-file client)
• Java 11+ — see sdk/java/TokenSlimClient.java

📖 5-minute Quickstart · Full SDK usage guide · User guide

Usage

CLI

# Compress a build log
tokenslim -i build.log -o output.json --reorder

# AI-friendly denoised diagnostic report
tokenslim decompress -i output.json -o ai_report.txt --ai-export

# High-signal lossy mode (keeps error window + key metadata)
tokenslim decompress -i output.json -o ai_signal.txt --ai-signal

# Static rule validation (single file)
tokenslim --verify-rule tests/fixtures/static_rule/sample_rule.toml \
  --verify-fixture tests/fixtures/static_rule/sample_fixture.log \
  --verify-expected tests/fixtures/static_rule/sample_expected.txt

# Static rule validation (batch, directory mode)
tokenslim --verify-rule tests/fixtures/static_rule/sample_rule.toml \
  --verify-fixture tests/fixtures/static_rule \
  --verify-expected tests/fixtures/static_rule

# Project bootstrap & shell hooks
tokenslim init
tokenslim workspace
tokenslim --dry-run workspace --inject
tokenslim workspace --inject
tokenslim hooks install
tokenslim hooks status
tokenslim hooks uninstall

Server (Sidecar)

tokenslim-server
# Listens on 127.0.0.1:<port>, see /health, /compress, /decompress

Docker

# Official image (multi-arch: linux/amd64 + linux/arm64)
docker run -d -p 10086:10086 ghcr.io/nuoyazhizhou/tokenslim:latest

# With API Key authentication
docker run -d -p 10086:10086 -e TOKENSLIM_API_KEY=my-secret ghcr.io/nuoyazhizhou/tokenslim:latest

# With JWT authentication
docker run -d -p 10086:10086 \
  -e TOKENSLIM_AUTH_MODE=jwt \
  -e TOKENSLIM_JWT_SECRET=my-secret \
  -e TOKENSLIM_API_KEY=my-key \
  ghcr.io/nuoyazhizhou/tokenslim:latest

Web UI

The sidecar ships a built-in single-page UI for interactive compression and live log tailing. All frontend static assets are compiled directly into the binary executable. Whether installed via npm or pip, it runs out-of-the-box from any directory with zero configuration.

Run

# Run from any directory (serves the embedded Web UI automatically)
tokenslim-server

# Frontend dev mode (serves from a physical directory for hot-reloading)
TOKENSLIM_WEBUI_DIR=./webui tokenslim-server

# Pick a port and bind address
TOKENSLIM_PORT=10086 TOKENSLIM_HOST=127.0.0.1 tokenslim-server

# Disable auth while poking around locally (default: off when env var unset)
# TOKENSLIM_API_KEY=changeme tokenslim-server

Open http://127.0.0.1:10086/ in a browser. The same /compress, /decompress, /plugins and /metrics endpoints that the CLI uses are exposed under the JSON API — the UI is just a thin client on top.

Environment variables

Features

• Drop a file onto the left pane, or paste a log dump.
• Switch between JSON, side-by-side diff and AI export views in the right pane.
• SSE 流式 checkbox streams /compress progress as Server-Sent Events so very large inputs do not block the UI.
• The history sidebar keeps the last few compressions in localStorage; the plugin-hit list shows which families matched the input.

A E2E test (tests/server_webui_e2e.rs) covers the static asset loading and the /compress round-trip; run it with cargo test --test server_webui_e2e.

JWT Authentication

TokenSlim Server supports three authentication modes:

# Get a JWT token
curl -X POST http://127.0.0.1:10086/auth/token \
  -H "Authorization: Bearer YOUR_API_KEY"
# {"token":"eyJ...","expires_in":3600,"token_type":"Bearer"}

# Refresh before expiry
curl -X POST http://127.0.0.1:10086/auth/refresh \
  -H "Authorization: Bearer YOUR_CURRENT_JWT"

WebSocket Bidirectional Compression

The /ws/compress endpoint provides a persistent bidirectional channel:

• Binary frames → raw data → compressed → Binary frame response
• Text frames → JSON control commands:
  • {"action":"flush"} — immediately compress and clear buffer
  • {"action":"reset"} — clear buffer and reset session
  • {"plugin":"<name>"} — switch compression plugin

Plugin Configuration Management

tokenslim config plugin status                       # View all plugin states
tokenslim config plugin disable gcc_log_plugin       # Disable a plugin
tokenslim config plugin enable gcc_log_plugin        # Enable a plugin
tokenslim config plugin list-params gcc_log_plugin   # List configurable params
tokenslim config plugin set gcc_log_plugin convert_timestamps false
tokenslim config plugin reset                        # Reset all plugin config

SDK

# Python
from tokenslim import compress, decompress
compressed = compress(open("build.log").read())
print(decompress(compressed, mode="ai-export"))

// Node.js
const { compress, decompress } = require("tokenslim");
const compressed = compress(fs.readFileSync("build.log", "utf8"));
console.log(decompress(compressed, { mode: "ai-export" }));

// Java
TokenSlimClient client = new TokenSlimClient("http://127.0.0.1:8080");
String compressed = client.compress(logText);
String report = client.decompress(compressed, "ai-export");

Log Reordering

Parallel build tools (make -jN, ninja, Bazel, MSBuild, …) interleave logs from multiple targets in an order that is non-deterministic and that breaks every diff / cache / regression comparison. TokenSlim ships a deterministic global reorderer that streams through the log, tracks the active build target, and emits lines in a stable target-grouped order.

# Built-in: the --reorder flag forces the reorderer and falls back to serial mode.
tokenslim -i build.log -o output.json --reorder

# Standalone tool: for pure log-to-log diff (Jenkins / CI envs) without the full pipeline.
cargo build --release --bin log_reorder
./target/release/log_reorder -i messy_build.log -o sorted_build.log --deterministic -n -p
#   --deterministic  : group lines by module / build target
#   -n  (--normalize) : sort out-of-order flags, redact addresses & random hashes
#   -p  (--shorten-paths) : collapse /home/userA/workspace/... to last 3 segments

The same engine is exposed via POST /compress (request field reorder: true), the WebUI checkbox "Enable reorder", and the Python / Node SDKs.

Plugins

TokenSlim ships with 60+ plugins covering the inputs that dominate real LLM traffic. Each plugin is data-driven (JSON / TOML config under config/plugins/) and dispatch is route-based, so adding a new source format is a config-only change in most cases.

Browse the full registry at config/plugins/, or run:

tokenslim plugins list
tokenslim explain-plugin --explain-command "cargo build"

Integrations

MCP Server (AI Agent integration)

TokenSlim ships a built-in MCP (Model Context Protocol) server that lets any MCP-compatible AI agent — Claude Code, Cursor, Windsurf, Qoder, OpenCode, and more — call compression tools directly through the standard protocol.

cd mcp-server && npm install && npm run build

Then add to your agent's MCP config (example for Cursor .cursor/mcp.json):

{
  "mcpServers": {
    "tokenslim": {
      "command": "node",
      "args": ["/path/to/mcp-server/dist/index.js"]
    }
  }
}

📖 Full setup guide, tool reference, and agent config examples: mcp-server/README.md

Architecture

TokenSlim follows a layered pipeline:

1. Route dispatcher — selects plugin(s) by command / content signature.
2. Plugin chain — each plugin owns extraction, folding, semantic substitution.
3. Compression core — radix-trie path extraction, dictionary layering, global dedup.
4. Rehydration — round-trip-safe so the original input is fully recoverable from the compressed form.
5. AI Export / Signal — context-aware post-processing for LLM consumption.

See docs/development/ARCHITECTURE.md for the full design.

🛡️ AI Agent Governance & Anti-Drift Sandbox

The hardest problem in autonomous AI code generation is "preventing the Coding Agent from writing code and writing its own self-congratulatory mock tests (garbage in, garbage out)", and "preventing subsequent refactoring from introducing silent target drift (behavior regressions)."

In a complex ecosystem with over 105k+ LOC of core source code, 60+ plugins, and 1000+ physical test cases, TokenSlim remains robust not by manual debugging, but through an automated, closed-loop Quality Sandbox that tames AI code-gen behavior:

1. Intent Extraction & Code Documentation Injection (extract_plugin_design.py): Scans parser source code, leverages LLMs to extract core design contracts (design_intent/keep_signals), and automatically injects them back into mod.rs as module-level //! doc comments! This forces future AI and human coders to respect the source of truth design boundaries.
2. Multi-Language Automatic Translation Sync (translate_messages_fields.py): Double-checks translation files for Chinese and English, automatically aligning missing entries via LLM translation to avoid key dropouts.
3. Physical Case Quality Audit (audit_sample_case_quality.py): The Micro-Judge. Validates the realism and alignment of physical raw log cases against the extracted design contract, filtering out fake/synthesized AI cases. It scans sample folders and automatically outputs code lines to append to showcase.rs if new physical logs are unregistered.
4. Compression Fidelity Audit (audit_case_metrics.py): The Meso-Judge. Enforces showcase.rs registration, samples/ physical files, and target/ generated reports alignment. It verifies G1-G4 deterministic gates (ensuring critical errors and command anchors are never lost) and uses LLMs to cross-check compression against the design contract.
5. State Freeze & Regression Prevention: Once audited, the output is locked with a SHA256 hash. If future AI changes break the expected output, the CI/CD pipeline instantly rejects and blocks the release, preventing silent drift.
6. Global Health Governance (audit_all_case_metrics.py): The Macro-Judge. Orchestrates parallel audits across all 60+ plugins in CI/CD, compiling a global health matrix (audit_health.md) to finalize quality blockings.

Contributing

Contributions are welcome. Please open an issue first to discuss larger changes; small fixes and new plugin configs can go straight to a PR.

# Run tests
cargo test

# Run with a sample
tokenslim -i samples/web_log_plugin/case_001_access.log -o out.json --reorder

License

MIT