A new Python package wants AI agents to talk in compact code, not English

What was released

As dev.to reported in a post by the project's author, TOAI-elegant-lang (also called Project Telepath) has reached version 5.4.1 and can now be installed from PyPI, the public library where Python developers share code. The package uses only Python's built-in tools, so it pulls in no outside software.

The release also includes a small server for MCP, the standard that lets AI assistants connect to outside tools. The post says this server offers seven tools and is meant to work with Claude Desktop, Cursor and VS Code Copilot. The project is also archived on Zenodo, a research repository, so academics can cite it.

The post says the protocol began as an experiment among 10 AI agents. It then went through two rounds of code and design review by the models DeepSeek and Qwen.

The idea: machines don't need prose

The pitch is simple. When one AI agent hands work to another, they usually exchange full English sentences or wordy structured data. Language models measure and charge for text in tokens, which are small chunks of words, so longer messages mean more computing and more electricity.

The author argues that this is a waste, since English exists for humans. TOAI-elegant-lang replaces it with a compact shorthand that agents would use among themselves. The post frames this as a way to save energy at the scale of data centers.

How the format is meant to work

According to the post, the core is a fixed vocabulary of 256 primitives, plus 256 macros that are swapped in and out as needed. The author picked 256 because it equals 2 to the 8th power, which fits an 8-bit instruction space. It also keeps each term to two characters in the way language models split text. The author says the whole translation layer fits within the prompt-caching limit of under 1,000 tokens that modern models allow, so translating needs no extra model calls.

Messages follow a single grammar that mixes context labels, key-value pairs and arrows carrying probability weights for cause and effect. The post says this can carry business data without a JSON parser. JSON is the common text format programs use to swap structured data.

To keep people in the loop, the project includes what it calls a Rosetta Stone layer with three scripts, which it says leaves everything auditable by humans. Even the post's sign-off is written in the format, so readers get a live sample.

The claims, and what is missing

The headline numbers are big. The author says the system was checked across more than 4,500 dialogue records from Google DeepMind, OpenAI, Anthropic, Meta, xAI, DeepSeek, Alibaba, Moonshot and Zhipu AI. It reports 0.0% out-of-vocabulary terms, compression of up to 22.8x, and a cut of about 90% in token use and inference energy.

Treat these with caution. They appear in a post promoting the release, written by the project's side. We have no independent test, no peer review, and no explanation of how the energy savings were measured. The source says 100% alignment across specifications, code and telemetry, but that too is a self-report.

There is a deeper question the post does not address. Models are trained mostly on human language, so it is unclear how well they follow a made-up shorthand in real work, or whether savings on tokens hold up once accuracy is counted. The post says the translation needs no extra model calls, but it does not show how well agents perform with the format in practice.

Why this is worth a glance anyway

Even if this particular project never catches on, the problem it targets is real. As companies wire many agents together, the cost of their chatter adds up, and teams are starting to ask what they get for their AI spending. A compact agent-to-agent format is one possible answer, and others may follow.

The source also makes a pointed claim: standards come from running code, not committees. That is a statement of belief, not a finding. Whether anyone beyond the author adopts the format will say more than any of the numbers in the post.