Ellements

A Python toolkit for extreme experimentation with LLM systems.

Ellements is best understood as a continuously extended set of building blocks for LLM experimentation: model clients, prompt context, execution strategies, agent abstractions, benchmarking harnesses, and terminal UI primitives. The repository is organized into focused source roots, but ships as a single PyPI package, ellements, containing every public ellements.* subpackage.

[!WARNING] Do not treat this as a normal dependency. This repository exists to support my own projects, experiments, and tooling. Probably nobody other than me should depend on it directly.

[!NOTE] Almost all of it is created with AI assistance, and I myself understand only a fraction of it at any given time. This can produce fast-paced changes, as well as both good and bad surprises. Pull requests are unlikely to be accepted. This is published for transparency and my own reuse, not as a community-maintained library or a general-purpose package roadmap.

Why yet another LLM library?

True, there are many other similar libraries. This one, however, is mine. It exists so I can keep extending it, test ideas, change direction, delete abstractions, and rebuild APIs as I see fit: not by committee, not by roadmap, not by community consensus.

Design principles

The practical consequences are the following:

• One install, focused internals. pip install ellements installs the full namespace: ellements.core, ellements.execution, ellements.agents, ellements.benchmarking, ellements.cli, ellements.domain_specific, ellements.reporting, ellements.standard_tools, and ellements.fslm.
• Extension before stabilization. The point is to keep adding mechanisms as they become useful in my projects. Stability comes later, if it comes at all.
• Async-only public API. No hidden sync wrappers. Callers keep explicit control of concurrency.
• Explicit model selection. LLMClient(model=...) is required. There is no hidden default.
• Explicit local caching. Pass LocalCacheConfig(directory=...) to cache exact text/Responses calls through LiteLLM and image generation/edit responses through Ellements' content-addressed disk cache. URL-only image responses are not persisted because provider URLs expire. Caching is off when omitted.
• Structured observability. Every LLM call emits request, response, and error events. JsonlPromptLogger writes durable JSON-lines traces.
• Composable strategy layer. Single-call, reflection, self-consistency, tree-of-thought, and collaborative editing all conform to one Strategy protocol.
• Backend-agnostic agents. ellements.agents is an abstraction layer over external agent libraries. OpenAI Agents and Claude Agents are adapters; they are not the identity of the module.
• Finite-state linguistic machines. ellements.fslm keeps agentic workflows bounded by explicit state graphs while still allowing natural-language guards, invariants, actions, and outputs where they are useful. The L is deliberate: language is part of the control surface, not a decorative interface.

Install

pip install ellements

Optional extras install integration-specific dependencies. They do not split the package or change which modules ship in the wheel.

pip install "ellements[agents]"        # declared agent-adapter dependencies
pip install "ellements[benchmarking]"  # lm-eval integration
pip install "ellements[cli]"           # Textual-powered terminal UI helpers
pip install "ellements[fslm]>=0.2.0"   # finite-state linguistic machines
pip install "ellements[finance]"       # Yahoo Finance asset tools
pip install "ellements[web]"           # web search, crawl, and YouTube tools
pip install "ellements[reporting]"     # chart/report export helpers
pip install "ellements[all]"           # all declared optional integrations

In particular, the agents module is not OpenAI-specific. It defines shared builder, controller, runner, and event abstractions over external agent libraries. The OpenAI adapter has declared optional dependencies; the Claude adapter is included in the wheel, but its upstream SDK is still evolving, so install it separately when using that backend.

What ships

Quick start

import asyncio

from ellements.core import JsonlPromptLogger, LLMClient, LocalCacheConfig


async def main() -> None:
    client = LLMClient(
        model="openai/gpt-5.5",
        observers=[JsonlPromptLogger("./logs")],
        local_cache=LocalCacheConfig("./cache"),
    )

    answer = await client.complete("Explain attention in one paragraph.")
    print(answer)


asyncio.run(main())

Run a strategy

from ellements.execution import ReflectionConfig, ReflectionStrategy

strategy = ReflectionStrategy()

result = await strategy.execute(
    prompts={
        "generate": "Draft a haiku about distributed systems.",
        "critique": "Review this draft and return a CritiqueResult:\n\n{{response}}",
        "revise": (
            "Revise the draft using the issues below.\n\n"
            "Draft:\n{{response}}\n\nIssues:\n{{issues}}"
        ),
    },
    client=client,
    config=ReflectionConfig(max_rounds=3),
)

print(result.output)

Runtime strategy templates accept both Mustache placeholders such as {{response}} and PromptSpec-style placeholders such as @{response}.

Build an agent

OpenAI is shown here as one concrete adapter; the builder targets the backend-agnostic AgentBackend protocol.

from ellements.agents import AgentBuilder, OpenAIAgentsBackend

agent = (
    AgentBuilder("researcher", backend=OpenAIAgentsBackend())
    .with_model("gpt-4.1")
    .with_instructions("Research carefully, cite evidence, and be concise.")
    .with_tool("search", my_search_tool)
    .build()
)

Use finance and web tools

Finance and web tools expose canonical ToolRegistry surfaces, so LLM clients, agents, and PromptSpec examples can bind the same tools without local adapters.

pip install "ellements[finance,web]"
playwright install chromium  # required by crawl4ai-backed page crawling

from ellements.domain_specific.finance.yahoo_finance import finance_tools
from ellements.standard_tools.web.crawler import web_crawler_tools
from ellements.standard_tools.web.search import web_search_tools

tools = (
    finance_tools()
    .merge(web_search_tools())
    .merge(web_crawler_tools(max_content_tokens=4000))
)

print(sorted(tools))

Typical stock-research tools include search_asset, get_asset_quote, get_asset_profile, get_financial_metrics, search_web, search_news, and crawl_url.

Packaging model

The repo is modular at the filesystem level:

ellements-core/src/ellements/core
ellements-execution/src/ellements/execution
ellements-agents/src/ellements/agents
ellements-benchmarking/src/ellements/benchmarking
ellements-cli/src/ellements/cli
ellements-domain-specific/src/ellements/domain_specific
ellements-fslm/src/ellements/fslm
ellements-reporting/src/ellements/reporting
ellements-standard-tools/src/ellements/standard_tools

Those source roots are discovered into a single wheel, ellements-<version>-py3-none-any.whl. Users install one PyPI project and import the modules they need:

from ellements.core import LLMClient
from ellements.domain_specific.finance.yahoo_finance import finance_tools
from ellements.execution import TreeOfThoughtStrategy
from ellements.agents import AgentBuilder
from ellements.fslm import FSLMKernel
from ellements.standard_tools.web.search import web_search_tools

This gives the repo clean internal boundaries without creating multiple PyPI entries to publish, document, secure, and maintain.

Local development

pip install -e ".[dev,all]"
python -m pytest -q
python -m ruff check .
python -m mypy --strict ellements-core/src ellements-execution/src \
  ellements-agents/src ellements-benchmarking/src ellements-cli/src \
  ellements-domain-specific/src ellements-fslm/src ellements-reporting/src \
  ellements-standard-tools/src

License

MIT, with the project status and maintenance notice in LICENSE.