🧬 AgentX-Kit

A provider-agnostic agentic framework + interactive project scaffolder for LangChain & CrewAI.

Pick your LLM provider (OpenAI, Azure, OpenRouter, Anthropic, Gemini, Vertex AI, Bedrock, Groq, Ollama, HuggingFace, Cohere, Mistral), choose your framework, agents, RAG, memory, MCP tools and skills — and AgentX-Kit generates a ready-to-run project in its own uv virtual environment.

pip install "agentx-kit[all]"
agentx new                 # interactive wizard → scaffolds a uv project

The PyPI distribution is agentx-kit; the import name and CLI are agentx (pip install agentx-kit → import agentx / agentx --help).

🚀 60-second walkthrough

# 1. Install
pip install "agentx-kit[all]"

# 2. See what you can target
agentx providers                      # 9 LLM providers + the env vars each needs

# 3. Scaffold a complete project from one line (no keys needed to generate)
agentx new --yes --name my-bot \
  --provider openai \
  --prompt "You are a support agent that answers from our docs."

# 4. Run it
cd my-bot && cp .env.example .env      # add your API key
uv sync && uv run my-bot

# 5. Tune prompts live (tokens, cost, quality, optimize) — optional UI
pip install "agentx-kit[dashboard]" && agentx dashboard

# 6. Use it from Claude / Copilot / Codex
claude mcp add agentx-kit -- agentx mcp

Prefer guided? Just run agentx new (interactive wizard) or agentx new --enterprise for the full production stack (tracing, guardrails, FastAPI, Docker, CI, evals, caching).

🧭 Command cheat-sheet

✨ Highlights

• 12 LLM providers + a curated model catalog (used by the wizard & dashboard).
• Structured project layout (new default): generated LangGraph projects are organised into nodes/ (one module per agent), state/, schemas/, prompts/, utils/ (llm.py, tools.py, rag.py, retriever.py, embeddings.py) and libs/ — a real project you can grow, not one big file.
• Multi-agent orchestration: choose supervisor (LLM router), sequential (pipeline), or parallel (fan-out + merge) when you have 2+ agents.
• Sub-agents / swarm (--subagents): attach delegate agents — each with its own tools / MCP / web search — to your agents via the agent-as-tool pattern.
• Voice I/O (--voice): local-first Speech-to-Text + Text-to-Speech (faster-whisper / edge-tts, OpenAI cloud fallback) — agentx.voice.
• Claw (--claw): a multi-channel content assistant (LLM intent router + a generic /claw/webhook) you can point WhatsApp / Telegram / Slack / email at.
• Streamlit UI (--streamlit): a chat front-end, with mic input & spoken replies when voice is enabled.
• Small/local-model resilient: models that emit tool calls as JSON text (e.g. llama3.2) are handled transparently — no more raw-JSON replies.
• RAG that actually chunks + embeds: LangChain splitter, FAISS or Chroma, 8 embedding providers (HuggingFace local needs no key), document loaders for PDF / Excel / CSV / Word / Markdown, and incremental re-index via a manifest.
• Autonomous & research agents (agentx agent run / agentx agent research) — sandboxed file tools, web search, citations.
• Deep agents (agentx agent deep, or agent_mode="deep" in the wizard) — a todo-list planning tool, sandboxed filesystem tools, sub-agent delegation (agent-as-tool), and an optional critic/reflection revision loop, the same primitives behind LangChain's deepagents and Claude Code's own harness.
• Flow — code as a DAG (agentx flow [path]) — a static AST call graph for a file or a whole project (no execution), or --live to run a file and see real call counts + timing via a @trace decorator. Export ascii/mermaid/ json/dot, or --ui for an interactive, colored 2D/3D graph viewer.
• Domain-aware seeding: name a project legal-assistant (or pass --domain) and it gets an expert system prompt + a seed knowledge base + RAG on.
• Dashboard v2: pick any provider/model, enter your API key in the UI, run tests with caching, LLM-judge relevance evals, and prompt history.
• agentx graph to see the agent flow; a VS Code extension in integrations/vscode-agentx/.
• Production-hardened: request timeouts, rate limiter, structured JSON logs, input guardrails, health/readiness probes, and quiet third-party logging.

▶️ Try the demos (no API keys needed)

bash examples/demo_local.sh            # verify local setup end-to-end
python examples/demo_mcp.py            # test the Claude/Copilot MCP path (real handshake)
python examples/mcp_toolkit_client.py  # web search / TTS / knowledge / DB tools over MCP

See examples/ for details.

📦 Installation

From PyPI (recommended)

pip install agentx-kit                 # core: CLI + scaffolder + base abstractions
pip install "agentx-kit[all]"          # everything

Each LLM provider is an optional extra so you only pull the SDKs you use:

pip install "agentx-kit[openai,langgraph]"        # OpenAI + LangGraph
pip install "agentx-kit[bedrock,crewai,rag,mcp]"  # Bedrock + CrewAI + RAG + MCP

Using uv

uv pip install "agentx-kit[all]"

From GitHub (latest, unreleased)

pip install "agentx-kit[all] @ git+https://github.com/muhammadyahiya/agentx-kit.git"

From a local clone (development)

git clone https://github.com/muhammadyahiya/agentx-kit.git
cd agentx-kit
uv venv && uv pip install -e ".[all,dev]"   # or: pip install -e ".[all,dev]"
pytest -q

Requires Python 3.10–3.13 and (for the scaffolder's .venv creation) uv.

Verify

agentx version
agentx providers     # lists every provider + the env vars it needs

Why

• One factory, every provider. get_chat_model("bedrock", ...) or get_chat_model("openrouter", ...) — same call, lazy imports, install only the extras you use.
• Two frameworks. LangChain/LangGraph and CrewAI from the same building blocks.
• Batteries included. RAG, short/long-term memory, MCP tools, and a skills registry — each optional and gracefully degrading.
• Scaffolder, not a black box. The generated project is readable, idiomatic code you own, pre-wired to your selections, in a fresh .venv.

Use as a library

from agentx import get_chat_model, list_providers

llm = get_chat_model("openai", "gpt-4o-mini")
print(llm.invoke("Say hi in 3 words").content)

for spec in list_providers():
    print(spec.id, "→", spec.label)

CrewAI:

from agentx import get_crewai_llm
llm = get_crewai_llm("openrouter", "anthropic/claude-3.5-sonnet")

Scaffold a project

agentx new                         # fully interactive
agentx new --name my-bot --yes     # accept sensible defaults
agentx providers                   # list providers + required env vars

The wizard asks, one option at a time:

1. Project name & target directory
2. Framework — LangGraph or CrewAI
3. LLM provider and model
4. Number of agents (and their roles)
5. RAG module? (vector store)
6. Memory? (short-term / long-term / both)
7. MCP tools? — and if so, which built-in ones your own MCP server exposes (web search, text-to-speech, knowledge research, database — see below)
8. Skills integration?
9. Prompt style (defaults or scaffolded custom prompts)
10. Create .venv and uv sync now?

It then renders the project, writes a feature-aware pyproject.toml + .env.example, and runs uv venv to create .venv.

Prompts: add at creation, or any time after

Prompts are not baked into code — every generated project keeps them in a prompts.json that agents.py loads dynamically. Add an entry and the project runs it on next start, no code changes.

# at creation
agentx new --yes -n chatops --prompt "You are a senior DevOps engineer. Be terse."

# after creation (run inside the project)
agentx prompt list
agentx prompt set assistant --text "You are an SRE. Prioritise reliability."
agentx prompt add reviewer --role "Code Reviewer" --goal "Review diffs" \
    --text "You review code for bugs and security."
agentx prompt remove reviewer

prompts.json:

{
  "with_rag": false,
  "agents": {
    "assistant": {"role": "...", "goal": "...", "system_prompt": "You are ..."}
  }
}

A blank system_prompt is auto-derived from the agent's role + goal. You can also just open prompts.json in an editor — the CLI is a convenience, not a gate.

📊 Prompt dashboard (observability + optimization)

A Streamlit workbench to understand and refine how your prompts talk to the LLM — launch it any time:

pip install "agentx-kit[dashboard]"
agentx dashboard                 # opens http://localhost:8501
agentx prompt set assistant -d   # edit a prompt AND open the dashboard

It gives you, live as you edit:

• Token count, context-window utilization gauge, and cost estimate (tiktoken-accurate).
• Quality score (0–100) with a checklist (role / goal / output-format / examples / constraints / specificity) and concrete suggestions + limit warnings.
• ✨ One-click LLM optimization — refines the prompt while preserving intent, shows a diff + rationale + token delta, and can apply the result straight back to prompts.json.
• ▶️ Test run — send the prompt to the model and see the response with tokens in/out, latency, and cost.
• 📈 Usage trends — tokens, cost, and latency over time, logged locally to .agentx/insights.jsonl.

Run it inside a generated AgentX project and it reads/writes that project's prompts.json; run it anywhere else for a free-form prompt scratchpad.

🔌 Use as a connector (Claude / Copilot / Codex)

AgentX-Kit ships an MCP server, so any MCP-capable assistant can scaffold a complete project from a single prompt with your problem statement.

pip install "agentx-kit[connector]"
agentx mcp --print-config        # prints the client config below

Add it to your client (then restart it):

// Claude Desktop / Codex / Copilot — under "mcpServers"
{ "mcpServers": { "agentx-kit": { "command": "agentx", "args": ["mcp"] } } }

# Claude Code one-liner
claude mcp add agentx-kit -- agentx mcp

Now just ask, in plain language:

“Build a customer-support agent that answers from our product docs and serves a REST API.”

The assistant calls AgentX-Kit's tools and you get a complete, runnable project:

• recommend_project(problem_statement) — suggests framework, provider, agent count, and features.
• create_agent_project(problem_statement, …) — generates the project (infers RAG/serve/memory/etc. from the statement, or take explicit overrides / enterprise=true) and returns the file tree + key file contents + run steps.
• list_providers, analyze_prompt, optimize_prompt — provider list + prompt insights.

So from one sentence the assistant produces a pre-wired project (prompts already seeded from your use case), ready to uv sync && uv run.

🛠️ MCP tool templates (web search · TTS · knowledge research · database)

AgentX-Kit ships ready-made MCP server tools, importable directly — no generated project required:

pip install "agentx-kit[connector,voice]"

from agentx.tools.mcp_server import build_mcp_server

mcp = build_mcp_server(
    name="my-tools",
    tools=["web_search", "tts", "knowledge_research", "database"],  # pick any subset
    knowledge_root="./knowledge",   # scanned by knowledge_research (md/txt/pdf/docx/csv/xlsx)
    db_path="./data.db",            # queried (read-only) by database
)
mcp.run()   # stdio MCP server — connect from Claude, a LangChain agent, or your own client

Try it: python examples/mcp_toolkit_server.py + python examples/mcp_toolkit_client.py.

Or generate a project with these baked in — pick "Integrate MCP tools?" in the wizard (or agentx new --yes --mcp --mcp-tools web_search,database) and the project gets its own src/<pkg>/mcp/server.py + a mcp/client_demo.py sample script, already registered in mcp_servers.json so the agent(s) can call these tools too:

uv run my-bot-mcp-server                        # run your generated MCP server
uv run python -m my_bot.mcp.client_demo          # sample client

🧠 Deep agents (planning · filesystem · sub-agents · reflection)

AgentX-Kit ships the same primitives behind LangChain's deepagents and Claude Code's own coding harness — usable directly as a library, via the CLI, or baked into a generated project.

from agentx.agents import DeepAgent, SubAgentSpec, ReflectionConfig

agent = DeepAgent.create(
    goal="Audit this repo's error handling and write a report.",
    provider="openai",
    workspace="./workspace",
    subagents=[
        SubAgentSpec(name="reviewer", description="Reviews code for bugs.",
                     prompt="You are a meticulous code reviewer."),
    ],
    reflection=ReflectionConfig(enabled=True, max_revisions=2),
)
result = agent.run()
print(result.summary)

From the CLI:

agentx agent deep "Audit this repo's error handling and write a report." --reflection

Or generate a project with a deep agent baked in — pick "Deep" as the agent mode in the wizard (or agentx new --yes --agent-mode deep) and the generated nodes/agent.py uses make_deep_agent_node(...) instead of the default chat node, with planning/filesystem/reflection wired per your choices.

🕸️ Flow — see your code as a DAG

Most Python devs understand a project by reading code or a static import graph. agentx flow builds an actual function-call DAG instead — either by parsing the file (no execution) or by running it and recording what really happened. Point it at a directory (or run it with no path at all) and it builds a whole-project graph — packages, modules, classes, and functions — instead of just one file.

agentx flow app.py                        # static call graph — no execution, works on any file
agentx flow app.py --entry train_model    # only the subgraph reachable from one function
agentx flow app.py -f mermaid             # paste into a .md file / VS Code / GitHub
agentx flow app.py -f dot > flow.dot && dot -Tsvg flow.dot -o flow.svg
agentx flow                               # whole project (cwd): modules, classes, functions

For the actual execution graph — real call counts and per-call timing — decorate functions with @trace and run your code normally, or let the CLI run it for you with --live (single file only):

from agentx.flow import trace

@trace
def clean_data(): ...

@trace
def train(): ...

train()   # each call is recorded — see agentx.flow.get_current_flow()

agentx flow app.py --live   # runs app.py, then renders the REAL execution graph

Interactive 2D/3D viewer

--ui skips the text renderers and opens a self-contained, interactive DAG viewer in your browser — no server, no CDN, works fully offline from one HTML file:

agentx flow --ui                 # whole project, opens the interactive viewer
agentx flow app.py --ui          # one file
agentx flow --ui --no-open -o flow.html   # write it without launching a browser

Nodes are colored by kind (function / class / module / external); a Modules → Classes → Full detail control collapses large projects down to a coarse module-to-module graph by default; click a node for its full source and file:line, click two nodes to highlight the call path between them, search by name, and toggle a secondary experimental 3D view (layered by call depth). Dark/light follows your system theme with a manual override.

Type-checking, schemas & live execution (opt-in)

Every node's side panel always shows its declared type-hinted signature and full source, plus a fields table for classes that look like Pydantic BaseModels — all pure ast, no execution, no new dependency. Two more capabilities are opt-in:

agentx flow --ui --typecheck        # attach ruff + ty diagnostics to nodes (red badge + list)
agentx flow app.py --serve          # click Run in the browser, watch it execute live

• --typecheck runs ruff (lint) and ty (Astral's type checker) as subprocesses and maps their diagnostics onto the nearest node — an inline red border marks nodes with errors, and the side panel lists them (each entry tagged with the tool that flagged it). Requires pip install "agentx-kit[typecheck]".
• --serve (single file only) starts a small local server — click Run in the viewer to execute the file as a subprocess, with stdout/ stderr and per-function call/return events streamed live into a log pane and pulsed onto the graph as they happen; Stop ends it. A command box in the same log pane doubles as a minimal terminal — type any command (e.g. streamlit run app.py) and it runs through your OS's own shell, exactly like typing it in a real terminal (quoting, &&, pipes all work; a mistyped or missing command just prints its own "not found" error like a real shell would, it can't crash the server). The one exception: if it's exactly python <file>.py and that file is part of a package, it's routed through the same package-aware path the Run button uses (so relative imports inside it resolve — see below), gaining trace events too. Binds to 127.0.0.1 only and every action requires a random per-session token embedded in the page, but clicking Run/typing a command does execute real code on your machine — only point it at code you trust. Requires pip install "agentx-kit[server]".
• --cdn references the 2D/3D graph libraries via CDN <script src> tags instead of inlining them (~2MB smaller file) — off by default, since the point of --ui is a single file that still works from a plain file:// URL with no network access.
• Large or accidental directories are guarded with --max-files (default 20000) — agentx flow errors out instead of silently scanning a huge/unrelated tree with no feedback.

Both --live and --serve run the target file the same package-aware way regardless of where you invoke agentx flow from: if the file sits inside a package (its directory has an __init__.py), it's run as that module (like python -m pkg.module) rather than as a bare script, so from .sibling import x-style relative imports inside it resolve correctly instead of failing with "attempted relative import with no known parent package".

Try it: python examples/flow_demo.py.

🧩 Editor & assistant integrations

The same connector powers ready-made integrations (see integrations/):

• VS Code extension (integrations/vscode) — commands for New Agent Project, Open Prompt Dashboard, Add Prompt, Cache Stats, and Register MCP Server for Copilot (writes .vscode/mcp.json). Build with vsce package.
• GitHub Copilot (agent mode) — add the MCP server via .vscode/mcp.json:

  { "servers": { "agentx-kit": { "command": "agentx", "args": ["mcp"] } } }
  

  (the VS Code command above writes this for you), then ask Copilot to build an agent.
• Claude Code plugin (integrations/claude-plugin):

  /plugin marketplace add muhammadyahiya/agentx-kit
  /plugin install agentx-kit@agentx-kit
  /agentx-kit:new-agent a support agent that answers from our docs and serves an API
  

• Claude Desktop / Codex — add the connector config from agentx mcp --print-config.

💾 Response caching (cost & latency saver)

Caching is the top 2026 token-optimization lever. Turn on a global LLM response cache and every provider call is served from a local store on repeat — no code changes:

from agentx import enable_caching, cache_stats
enable_caching()                 # all get_chat_model(...) calls are cached
...
print(cache_stats())             # {'hit_rate': 0.6, 'tokens_saved': 12000, 'est_usd_saved': 0.024, ...}

agentx cache stats               # hit rate + estimated tokens/$ saved
agentx cache clear

Generated projects can enable it automatically (it's part of --enterprise), and the dashboard's Trends tab shows live hit-rate and $ saved. TTL-capable, SQLite-backed at .agentx/llm_cache.sqlite.

🏢 Enterprise pack

Generate a production-shaped project with one flag — informed by a survey of CrewAI/LangGraph/create-llama/AgentStack/agno/pydantic-ai (see RESEARCH.md):

agentx new --yes -n my-bot --enterprise        # everything below
# or pick individually:
agentx new --yes -n my-bot --observability --guardrails --serve --docker --ci --evals

What --enterprise adds to the generated project:

• Observability — OpenTelemetry GenAI tracing + optional Langfuse (observability.py), opt-out via AGENTX_TELEMETRY=false.
• Guardrails — input/output validation + PII redaction (guardrails.py).
• FastAPI server — server.py with /health, /chat, and SSE /chat/stream.
• Docker — Dockerfile + docker-compose.yml (+ .dockerignore).
• CI — .github/workflows/ci.yml (lint + compile + tests, optional eval gate).
• Evals — evals/ LLM-as-judge harness runnable locally and in CI.
• Typed config — config.py via pydantic-settings (12-factor).
• Manifest — agentx.json declaring framework, provider, features (à la langgraph.json).

These are also usable as a library in any project:

from agentx import (
    setup_tracing, get_callbacks,          # observability
    build_resilient_chat,                  # retries + provider fallbacks
    UsageLimits, UsageTracker,             # token/cost budgets
    apply_guards, structured_model,        # guardrails + typed outputs
)
setup_tracing("my-service")
llm = build_resilient_chat("openai", "gpt-4o-mini", fallbacks=[("anthropic", "claude-3-5-sonnet-latest")])

Installation extras

See DESIGN.md for the architecture and RESEARCH.md for the competitive analysis behind these features.

License

MIT