pykel 1.8.0

A production-grade agentic AI framework for Python: a unified multi-provider model gateway with built-in observability, cost governance, memory, retrieval (RAG), multi-agent orchestration, and deterministic testing.

kel — The Open-Source Agentic AI Framework for Python

kel is a production-grade agentic AI framework for Python: a unified multi-provider model gateway with built-in observability, cost governance, memory, retrieval (RAG), multi-agent orchestration, self-healing, and deterministic testing — engineered to close the gaps that LLM orchestration frameworks commonly leave open: opaque execution, no native cost control, no reproducible testing, and context/loop failures.

If you're evaluating AI agent frameworks, production-ready agent orchestration libraries, or a Python framework for building autonomous AI agents with real observability and cost governance — this is built for exactly that.

GitHub · PyPI · Design & Architecture · Usage Guide · Issues

Install with pip install pykel — the PyPI distribution is named pykel (the kel name was already registered), but the import stays import kel and the CLI command stays kel.

---

Why kel

Common pain point in LLM orchestration frameworks	How kel solves it
No first-class tracing — bolt on a third-party observability product or go without	Every call traced by default, self-hostable via Grafana/OTel
Hidden token/cost consumption, surprise bills	Budget objects threaded through every call — hard caps, not suggestions
Deeply nested abstractions, painful debugging	Flat, composable classes — no multi-layer wrapper hierarchies to step through
Agents loop forever, no stuck-loop detection	Built-in stuck-loop + step-budget guardrails
No deterministic testing story	Record/replay testing — real API calls once, deterministic CI forever
Multi-agent state gets lost between agents	Shared context bus — downstream agents see what upstream agents decided
Real disclosed CVEs (deserialization, path traversal, SQLi)	Hardened by design — restricted unpickling, parameterized queries, Trivy-scanned every push
Forces credentials everywhere	Credentials optional by design — works natively on EC2/EKS/IAM roles, IRSA, self-hosted anonymous access
Single shared agent instance means every caller shares one conversation	Per-session Agent factories in the FastAPI/WebSocket adapters — one isolated conversation per session, no extra plumbing

---

Feature Highlights

• Model Gateway — one interface across Anthropic, OpenAI, Cohere, Gemini, Mistral, sync + real async
• Observability — every span traced automatically → console, Grafana, or the built-in live dashboard
• Budget & Rate Limiting — token/cost/tool-call caps + RPM/TPM throttling, composable
• Caching — in-memory or SQLite response caching, never double-charges budget on a hit
• Memory — working / episodic / semantic / procedural, layered like a real cognitive architecture
• Retrieval (RAG) — Qdrant, Pinecone, Weaviate, Chroma, pgvector, hybrid search, recursive splitting, PDF loading
• Multi-Agent Orchestration — sequential, supervisor, parallel, and swarm patterns, streaming included
• Brain — fast rule/embedding routing with LLM fallback, parallel-to-finish scheduling
• Self-Healing — diagnosis-driven retries with a non-negotiable idempotency guardrail
• Built-in Tools — web search (7 providers), URL fetch, Python/shell exec, SQL query, MCP servers
• Media Generation — image, video, text-to-speech, and lipsync via a generic gateway (fal.ai, Replicate), open for other vendors
• Testing — record/replay + LLM-graded evaluation, no live API key needed in CI
• Live Monitoring Dashboard — zero-dependency, real-time metrics + logs in your browser
• Production Deploy Adapters — stdlib HTTP, real WebSocket, and FastAPI (with per-session Agent isolation) — see USAGE.md §14
• DevSecOps Built In — Trivy vulnerability/secret scanning, pip-audit, and Bandit SAST on every push

---

Install

pip install pykel               # from PyPI — import kel, run `kel`, same as always
pip install "pykel[anthropic]"  # + Anthropic
pip install "pykel[openai]"     # + OpenAI
pip install "pykel[gemini]"     # + Google Gemini
pip install "pykel[mistral]"    # + Mistral
pip install "pykel[qdrant]"     # + Qdrant vector store
pip install "pykel[fastapi]"    # + FastAPI production adapter
pip install "pykel[fal]"        # + fal.ai image/video/audio/lipsync generation
pip install "pykel[replicate]"  # + Replicate image/video/audio generation
pip install "pykel[all]"        # everything

Working from a clone instead: pip install -e ".[dev]" (or -e ".[all]" for every extra).

Quickstart

from kel import get_model, Message

model = get_model("anthropic:claude-sonnet-5", api_key="...")
response = model.generate([Message.user("Hello!")])
print(response.text)

Swapping providers is a one-line change — same interface, zero rewrites:

model = get_model("openai:gpt-5.2", api_key="...")
model = get_model("gemini:gemini-2.5-flash")      # falls back to env vars / ADC
model = get_model("mistral:mistral-large-latest")  # falls back to MISTRAL_API_KEY

A minimal tool-calling agent, deployed as a production FastAPI app with one Agent per session (see USAGE.md §14 for the full walkthrough):

from kel.agents import Agent
from kel.sdk import create_fastapi_app

def make_agent() -> Agent:
    return Agent("assistant", get_model("anthropic:claude-sonnet-5"))

app = create_fastapi_app(make_agent)   # pass a factory, not an instance
# uvicorn app:app — POST /invoke {"input": "...", "session_id": "user-42"}

---

Status

Every subsystem in DESIGN.md has a working, tested implementation — model gateway, observability, budget, context/loop, memory, retrieval, specs, runtime graph, multi-agent orchestration, brain, self-healing, testing, storage, SDK/CLI, monitoring dashboard, and realtime orchestration (interfaces only, by design — see DESIGN.md §7). Known gaps are listed honestly at the bottom of USAGE.md — no overclaiming.

Security

Every push and pull request runs a full DevSecOps pipeline: Trivy filesystem + secret scanning, pip-audit for known CVEs in dependencies, and Bandit static analysis over the codebase. Results surface in the repo's Security tab. See security.yml. To report a vulnerability, see SECURITY.md.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Adding a new model provider, vector store, or tool follows the same lazy-import adapter pattern throughout the codebase — see any file under src/kel/models/providers/ for the template.

License

MIT © kvenkatprasad