The Edge Agent

Small LLMs hallucinate. TEA fixes that with Prolog.

One binary. No cloud. Neurosymbolic AI that actually reasons.

TEA combines LLMs with symbolic reasoning (Prolog) to create AI agents that can prove their conclusions, not just generate plausible-sounding text. Perfect for small/local models (Llama, Mistral, Phi, Ollama) where symbolic reasoning compensates for limited model capacity.

30-Second Example

# LLM extracts facts, Prolog derives relationships via temporal reasoning
name: hero-family-reasoning

nodes:
  - name: extract
    uses: llm.call              # LLM extracts: mother(alice, bob). affair(alice, dave, 1980, 1990).
    with:
      model: "gemma3n:e4b"
      messages:
        - role: user
          content: "Extract family relationships as Prolog facts from: {{ state.text }}"
    output: llm_response

  - name: reason
    language: prolog            # Prolog derives: child_of_affair, half_sibling
    run: |
      child_of_affair(Child, Partner) :-
          mother(Mother, Child), birth_year(Child, Year),
          affair(Mother, Partner, Start, End), Year >= Start, Year =< End.

      half_sibling(X, Y) :-
          mother(M, X), mother(M, Y), X \= Y,
          \+ (father(F, X), father(F, Y)).

      state(facts, Facts), tea_load_code(Facts),
      findall(H, half_sibling(bob, H), Results),
      return(half_siblings, Results).

Run it:

tea run examples/prolog/neurosymbolic/hero-family-reasoning.yaml \
  --input '{"text": "Alice had two children: Bob and Carol. Alice had an affair with Dave from 1980 to 1990. Bob was born in 1985. Carol was born in 1975.", "person": "bob"}'
# Output: {"answer": "bob's half-siblings: Carol"}

What happens: LLM extracts facts → Prolog proves Bob is Carol's half-sibling (born during affair = different father).

Full runnable example: examples/prolog/neurosymbolic/hero-family-reasoning.yaml

Why TEA?

Challenge	TEA Solution
Small LLMs make reasoning errors	Prolog handles logic, math, and constraints while LLM handles language
LLMs hallucinate facts	Knowledge graphs with verifiable inference chains
Complex agent frameworks	Simple YAML syntax, learn in minutes
Need for external services	Single binary, zero dependencies, runs offline
Cloud vendor lock-in	Portable agents run on any platform
Building everything from scratch	20+ built-in actions for LLM, RAG, memory, and storage
No visibility into agent behavior	Built-in observability with distributed tracing

Quick Install

Python (Recommended) - Full features, 20+ built-in actions:

# Option 1: pip install (requires Python 3.10+)
pip install the-edge-agent
tea --version

# Option 2: AppImage (self-contained, includes Prolog)
VERSION=$(curl -s https://api.github.com/repos/fabceolin/the_edge_agent/releases/latest | grep -Po '"tag_name": "v\K[^"]+')
curl -L "https://github.com/fabceolin/the_edge_agent/releases/download/v${VERSION}/tea-python-${VERSION}-x86_64.AppImage" -o tea && chmod +x tea

Rust - Minimal footprint, embedded/offline:

VERSION=$(curl -s https://api.github.com/repos/fabceolin/the_edge_agent/releases/latest | grep -Po '"tag_name": "v\K[^"]+')
curl -L "https://github.com/fabceolin/the_edge_agent/releases/download/v${VERSION}/tea-${VERSION}-x86_64.AppImage" -o tea && chmod +x tea

See Installation Guide for all platforms, Docker images, and AppImage variants.

Offline LLM Support

Run LLM workflows without internet using bundled GGUF models:

# Download Python LLM-bundled AppImage (~2GB with Gemma 3 1B)
VERSION=$(curl -s https://api.github.com/repos/fabceolin/the_edge_agent/releases/latest | grep -Po '"tag_name": "v\K[^"]+')
curl -L "https://github.com/fabceolin/the_edge_agent/releases/download/v${VERSION}/tea-python-llm-gemma3-1b-${VERSION}-x86_64.AppImage" -o tea-llm.AppImage
chmod +x tea-llm.AppImage

# Run offline chat
./tea-llm.AppImage run examples/llm/local-chat.yaml \
  --input '{"question": "What is the meaning of life?"}'

See LLM-Bundled Distributions for all model variants and configurations.

vs Alternatives

Feature	TEA	LangGraph	AutoGen
Symbolic reasoning (Prolog)	Yes	No	No
Single binary	Yes	No (Python)	No (Python)
Offline operation	Yes	Limited	No
YAML-first	Yes	Code-first	Code-first
Neurosymbolic	Yes	No	No
Local LLM optimized	Yes	Partial	No
Edge/embedded ready	Yes	No	No

TEA Does More

TEA includes 20+ built-in actions. Full documentation in docs/capabilities/:

Capability	Description	Key Actions
Neurosymbolic	LLM + Prolog hybrid reasoning	language: prolog
LLM Integration	OpenAI, Azure, Ollama, 100+ providers	llm.call, llm.structured
RAG	Vector search and document retrieval	rag.search, rag.embed
Memory	Short-term, long-term, cloud-synced	memory.store, memory.recall
Web	Scraping with Firecrawl, ScrapeGraphAI	web.scrape, web.crawl
Observability	Distributed tracing and debugging	trace.span, trace.log

Documentation

Topic	Link
YAML Reference	docs/shared/YAML_REFERENCE.md
CLI Reference	docs/shared/cli-reference.md
Python Guide	docs/python/getting-started.md
Rust Guide	docs/rust/getting-started.md
Human-in-the-Loop	docs/guides/human-in-the-loop.md

Implementations

Implementation	Status	Best For
Python	Production-ready (Recommended)	Full features, 20+ built-in actions, neurosymbolic AI
Rust	Active development	Embedded, offline, resource-constrained environments
WASM	Prototype	Browser-based, client-side LLM, zero-server

Python is the reference implementation with the complete feature set. Rust provides a lighter-weight alternative for embedded scenarios with a subset of actions. All implementations share the same YAML syntax.

Browser LLM (WASM)

Run TEA workflows with local LLM inference entirely in the browser:

# Install from npm
npm install tea-wasm-llm

# Or download from releases
VERSION=$(curl -s https://api.github.com/repos/fabceolin/the_edge_agent/releases/latest | grep -Po '"tag_name": "v\K[^"]+')
curl -L "https://github.com/fabceolin/the_edge_agent/releases/download/v${VERSION}/tea-wasm-llm-${VERSION}.tar.gz" -o tea-wasm-llm.tar.gz

import { initTeaLlm, loadModel, executeLlmYaml } from 'tea-wasm-llm';

await initTeaLlm();
await loadModel('./models/phi-4-mini.gguf');

const result = await executeLlmYaml(yamlWorkflow, { question: "What is 2+2?" });

Features:

• Offline operation after initial model download
• IndexedDB model caching for fast subsequent loads
• Multi-threaded inference via SharedArrayBuffer
• Opik observability integration

Try it live: Interactive WASM Demo - run LLM inference directly in your browser (no server required).

See WASM LLM Deployment Guide for server configuration and examples.

Examples

See examples/ for ready-to-run agents:

• examples/prolog/neurosymbolic/ - LLM + Prolog reasoning
• examples/llm/ - Pure LLM workflows
• examples/llamaindex/ - RAG and document retrieval
• examples/web/ - Web scraping agents
• examples/academic/ - Academic writing workflows

Academic Writing Example

Generate research papers with human-in-the-loop review:

cd python
python -m the_edge_agent.cli run \
  ../examples/academic/kiroku-document-writer.yaml \
  --input @../examples/academic/sample-paper-spec.yaml

This workflow includes:

• AI-assisted title suggestion and research
• Topic sentence planning with manual review
• Draft generation with revision cycles
• Abstract and citation generation

See Migration Guide for migrating LangGraph workflows to TEA.

Contributing

We welcome contributions! Please open an issue or pull request on GitHub.

License

MIT License. See LICENSE for details.

Acknowledgements

TEA is inspired by LangGraph. We thank the LangGraph team for their innovative work in language model workflows.

---

If TEA helps your project, consider starring the repo!