100% Open Source · MIT License · clacky-ai/openclacky

The Most Token-Efficient Open-Source AI Agent

Claude Code-class capability. Lower cost than any Agent in its class.

30 days since launch · 7-day network-wide cache hit rate 94.9% · 498,877 tasks completed
Backed by  ·  MiraclePlus · ZhenFund · Yingdong Capital · Variable Capital · Sequoia China · Hillhouse Ventures
Live Stats · Real Data

Don't take our word for it. Watch the network.

Not marketing copy. These numbers come from anonymous telemetry across all OpenClacky users, refreshed hourly.

Network-wide cache hit rate
Last 30 days
94.9% 7-day hit rate

What does this mean? About 85% off your input bill. Cached tokens are billed at ~10% of full price (as low as 1% on DeepSeek & MiMo) — the higher the hit rate, the more you save.

95% excellent 90% good 60–70% typical peer agents
Tokens processed yesterday
8.76B
Total tasks completed
498,877
All figures from real anonymous telemetry
Auto-refreshed every hour
What users say

Saves tokens · Works with any model · Easy to install

Real feedback from real users, unedited.

Under 100 MB install vs Hermes at 48 GB. The WebUI is fantastic. After using this, I realized multi-agent is unnecessary. Our entire cross-border e-commerce team is on it now.

Same user
Same user
3 weeks later, full team switched

Compared to Hermes and OpenClaw, OpenClacky's tool calling is rock solid. Works great with domestic models.

Ops Lead, Major Listed Company
Ops Lead, Major Listed Company

All quotes are genuine user feedback, published publicly or with permission.

Why we save tokens

Four things done right.
Tokens follow.

Not by cutting features — by compounding correct choices at every layer.

Ultra-High Cache Hit Rate

Session never restarts. Double-window cache markers. Insert-then-Compress reuses cache even during history compression — system prompt never mutates.

Measured cache hits approach 100%

Minimal Toolset, Maximal Reach

Just 16 core tools. Capabilities extend via the invoke_skill meta-tool. Tool count isn't competitive advantage — task completion is.

16
OpenClacky
40+
Claude Code
23
OpenClaw
52
Hermes

Idle-time compression. Don't pay while you're away.

Step away for a meeting or coffee — the agent quietly compresses long context in the background. By the time you're back, the history loaded into the model is already down to its essence. The tokens normally burned between turns? Gone.

Between-turn idle gaps can account for 30%+ of total session cost

Pick your model. Pick your route.

Any OpenAI-compatible API, plug and play. Official direct, aggregated routing, compatible relays — the choice is 100% yours, and so is the floor on your cost.

Claude GPT DeepSeek Kimi Minimax OpenRouter
BYOK · Bring Your Own Key

Pick your model. Set your cost.

Bring your own key — completely free. Any OpenAI-compatible API works out of the box. Or use OpenClacky Keys for the hands-off path: direct upstream routing, 99% cache hits, official pricing.

Claude
Anthropic
GPT
OpenAI
DeepSeek
Flagship CN
Kimi
Moonshot
Minimax
CN
OpenRouter
Aggregator

Not locked to one vendor. Main task on Claude, subagents auto-route to DeepSeek — another big chunk of tokens saved.

One-click install

Double-click to install.
No setup required.

Download, double-click, start using — environment, dependencies, Skills all ready. Desktop installer recommended; CLI also available for power users.

Recommended · Desktop Installer

Double-click to install — full environment configured automatically

Or install via command line

One line in your terminal — for developers who prefer the CLI

$
Apple Silicon & Intel supported
$
Windows 10 2004+ or Windows 11
$
Ubuntu, Debian, CentOS and more
Skill Power

Skills are the soul of any modern agent

Summon them, build them, grow them, sell them — OpenClacky treats Skills as native capability, not plug-ins.

/ Summon any Skill

Type / to browse, fuzzy-search, and invoke instantly. Hundreds of Skills one keystroke away — no menus, no memorizing names.

Build Skills by chatting

Tell the agent what you want. It drafts the SKILL.md, breaks the steps down, runs validation. No code required.

Skills self-evolve

After each run, the AI reviews context and outcome, then updates the Skill itself. Next time you invoke it, it performs more reliably.

Open formats, ready to monetize

Compatible with Skill formats from any agent ecosystem. Polished Skills can be packaged, encrypted, licensed, and priced on your terms.

OpenClacky Skill Claude Skills Markdown Pack Custom format
Advanced · Creator Program

Once you master it,
package your workflow and sell it

Professionals are already building their own vertical AI experts on OpenClacky — encrypted distribution, license management, your own pricing.

Legal · Law Firms
QingClaw

10 years of legal expertise,
now running on every lawyer's desktop

"I just put my 10 years of expertise into every colleague's computer."

Life Planning · High Net Worth
JoyClaw

AI life planning assistant
for every major decision

"Wealth gets divided. Life wisdom in the system stays yours."

Healthcare · Primary Care
MediClaw

Junior doctors access
senior expertise instantly

"The veteran retired, but the expertise stayed."

Trust & Backing

Open-source code. Verifiable results.

All code is public, all data is public. One month in — every number can be checked.

30 days
Since launch
94.9%
7-day cache hit rate
498,877
Tasks completed
970
GitHub Stars
Investors
MiraclePlus
ZhenFund
Yingdong Capital
Variable Capital
Sequoia China
Hillhouse Ventures

Make every token earn its keep

Open, free, BYOK. Download OpenClacky and take back control of your AI budget.

Free Download ⭐ Star on GitHub 970