SmithAgentic - Local Multi-Agent AI System

A fully local, open-source multi-agent AI system built on CrewAI + Ollama. No paid APIs. No cloud dependencies. Everything runs on your machine.

What it does

You give it a goal. A crew of specialized agents works through it - planning, researching, building, reviewing, validating, and documenting - entirely locally via Ollama.

All outputs are saved to outputs/ as markdown or code files.

Agent Team

Agent	Role	Crew	Model Tier
Orchestrator	Plans and decomposes goals	Default	7B
Researcher	Gathers information	Default	7B
Builder	Produces deliverables	Default	7B-Coder
Critic	Reviews output quality	Default	7B
QA Sentinel	Executes code, blocks on failures	Safety	8B
Security Reviewer	OWASP/vulnerability scanning	Safety	7B
Deployment Validator	Compile and deploy checks	Safety	7B
Documentation Writer	Generates structured docs	Ops	8B
Memory Manager	Consolidates run knowledge	Ops	8B
Observability Monitor	Run telemetry and audit	Ops	8B
PLC Planner	PLC program structure	PLC	8B
PLC Developer	Ladder logic / AOI code	PLC	8B
PLC Safety Reviewer	NFPA/IEC safety compliance	PLC	8B
UI Planner	React component design	React	8B
UI Builder	React/MUI component code	React	8B
UI Reviewer	React code quality review	React	8B
Vision Analyst	Queries Vision_Inspect API, parses defect results	Vision	7B
Vision Reporter	Synthesizes findings into structured reports	Vision	8B
Vision QA Validator	Health-checks pipeline, audits reports, detects anomalies	Vision	8B

Agents

Default crew agents

File	Role	Specialization
`agents/orchestrator.py`	Orchestrator	Decomposes any goal into a structured execution plan
`agents/researcher.py`	Researcher	General-purpose technical research and synthesis (shared across all crews)
`agents/builder.py`	Builder	Produces deliverables for any domain
`agents/critic.py`	Critic	Reviews deliverables against goal and plan

Safety crew agents

File	Role	Specialization
`agents/qa_agent.py`	QA Sentinel	Executes code artifacts, blocks on crashes or import failures
`agents/security_agent.py`	Security Reviewer	OWASP Top-10 audit, hardcoded credential detection, injection risk scanning
`agents/deploy_agent.py`	Deployment Validator	py_compile checks, import chain validation, deploy readiness verdict

Ops crew agents

File	Role	Specialization
`agents/docs_agent.py`	Documentation Writer	Generates structured README, API reference, and quickstart docs
`agents/memory_agent.py`	Memory Manager	Distills run outputs into tagged ChromaDB memory entries
`agents/observability_agent.py`	Observability Monitor	Reads all run outputs, reconstructs agent trace, produces telemetry report

PLC crew agents

File	Role	Specialization
`agents/plc_planner.py`	PLC Program Planner	ControlLogix architecture, program hierarchy, I/O tag lists, AOI/UDT planning; aware of `plc_generator/` codebase
`agents/researcher.py`	PLC Standards Researcher	Reuses general researcher with PLC-specific tool access (codebase reader + memory)
`agents/plc_developer.py`	PLC Developer	Ladder logic, structured text, function blocks, AOI/UDT authoring; reads and extends `plc_generator/` modules
`agents/plc_safety_reviewer.py`	PLC Safety Reviewer	NFPA 79, ISO 13849, IEC 62061, ISA-88 compliance; fault handling and interlock completeness

React crew agents

File	Role	Specialization
`agents/ui_planner.py`	React UI Planner	Component tree design, data flow, MUI v5 theming, industrial HMI patterns
`agents/researcher.py`	React Technical Researcher	Reuses general researcher with codebase reader access to study existing components
`agents/ui_builder.py`	React UI Developer	React 18 + MUI v5, dark industrial theme, complete drop-in components
`agents/ui_reviewer.py`	React UI Reviewer	Correctness, industrial UX, accessibility, performance, stack consistency

Vision crew agents

File	Role	Specialization
`agents/vision_analyst.py`	Vision Inspection Analyst	Queries Vision_Inspect `/inspections` API; parses defect type, severity, confidence, zone; produces `vision_findings.md`
`agents/vision_reporter.py`	Vision Inspection Reporter	Reads analyst findings + ChromaDB trends; writes `inspection_report.md` with executive summary, statistics, defect breakdown, trend comparison, and recommended actions
`agents/vision_qa.py`	Vision QA Validator	Health-checks `/health` endpoint; audits report completeness and numeric consistency; flags zero-defect anomalies vs. historical baseline; writes `vision_qa_report.md`

Features

Feature	Details
ChromaDB Memory	Agents store and query insights across sessions via `memory/chroma/`
Reflexion Loop	Default crew runs plan, research, build, and critique in sequence
Per-Agent Routing	Each agent uses its optimal model tier (configured in `agent_models:`)
Human-in-the-loop	Plan approval step before full crew runs (can skip with `--no-hitl`)
Code Executor	Agents can run Python snippets and capture output
QA Sentinel	Blocks pipeline if generated code crashes, fails imports, or has syntax errors
Security Reviewer	OWASP Top-10 audit of all generated code before delivery
Deployment Validator	py_compile checks on all Python artifacts before final output
Observability Monitor	Produces telemetry_report.md after every run for developer review
Git Tool	Agents can stage, commit, and push changes via GitPython
Codebase Reader	Agents can read any file in the repo before generating new code
PLC Crew	Domain-tuned agents for Rockwell Logix / ladder logic / AOI development
React Crew	Domain-tuned agents for industrial React/MUI HMI development
Vision Crew	Connects to Vision_Inspect backend; runs defect analysis, report generation, and pipeline QA
Safety Crew	Standalone QA, security, and deploy validation pipeline for any deliverable
Ops Crew	Standalone documentation, memory consolidation, and telemetry pipeline
Chain Flag	`--chain` runs safety then ops automatically after any primary crew finishes
Web UI	FastAPI + React CDN frontend with live WebSocket output streaming

Prerequisites

Python 3.10+
Ollama installed and running: https://ollama.com
At least one model pulled into Ollama:
```
ollama pull llama3.1:8b
```

Setup

# 1. Clone the repo
git clone https://github.com/AddisonTech/Smith_Agentic.git
cd Smith_Agentic

# 2. (Recommended) Create a virtual environment
python -m venv .venv
.venv\Scripts\activate          # Windows
# source .venv/bin/activate     # macOS/Linux

# 3. Install dependencies
pip install -r requirements.txt

# 4. Make sure Ollama is running
ollama serve

Running - CLI

# Default crew
python main.py --goal "Write a technical spec for a REST API rate limiter"

# PLC-specialized crew
python main.py --goal "Build a Rockwell Logix program for a conveyor E-stop system" --crew plc

# React-specialized crew
python main.py --goal "Build a React dashboard for machine OEE metrics" --crew react

# Vision inspection crew (requires Vision_Inspect running on port 8000)
python main.py --goal "Run a defect analysis on today's inspection batch" --crew vision

# Safety crew - validates deliverable in outputs/ for code quality, security, and deploy readiness
python main.py --goal "Validate the deliverable in outputs/" --crew safety

# Ops crew - generates docs, consolidates memory, and produces telemetry for the last run
python main.py --goal "Generate docs and telemetry for the last run" --crew ops

# Override model
python main.py --goal "..." --model qwen2.5:14b

# Run safety and ops crews automatically after the primary crew finishes
python main.py --goal "Write a Python rate limiter with tests" --chain

# Skip human-in-the-loop plan approval
python main.py --goal "..." --no-hitl

# Suppress verbose agent output
python main.py --goal "..." --no-verbose

Running - Web UI

python ui/server.py

Open http://localhost:8765 in your browser.

Features:

Goal input + crew/model selector
Live streaming agent output (WebSocket)
Download output files directly from the browser
Ollama online/offline indicator

Human-in-the-Loop Plan Approval

Before the full crew runs, the Orchestrator generates an execution plan and shows it to you:

============================================================
  HUMAN-IN-THE-LOOP: Plan Approval
============================================================
  Goal: Build a conveyor E-stop system
  Generating execution plan...

------------------------------------------------------------
  PROPOSED PLAN (iteration 1):
  1. Research NFPA 79 E-stop requirements...
  2. Define I/O tags and UDTs...
  ...

  [A]pprove and continue  /  type revision notes:

Press Enter or type a to approve and launch the full crew.
Type revision notes to have the Orchestrator regenerate the plan.
Use --no-hitl (CLI) or the Web UI (always runs without blocking) to skip.

ChromaDB Memory

Agents automatically store key findings and recall relevant past work:

MemoryStoreTool - saves insights with a topic tag + session ID
MemoryQueryTool - retrieves semantically similar past memories
Persists to memory/chroma/ across runs
Session ID tags let you distinguish different run histories

Memory is enabled by default. To disable:

# config/config.yaml
memory:
  enabled: false

Output Files

Default crew

File	Contents
`research.md`	Researcher's findings
`deliverable.md`	Builder's initial deliverable
`critique.md`	Reviewer's notes (APPROVED / NEEDS REVISION)
`deliverable_revised.md`	Builder's revised deliverable
`revision_summary.md`	What changed and why

Safety crew

File	Contents
`qa_report.md`	QA Sentinel verdict (SENTINEL_PASS / SENTINEL_BLOCK), exit code, errors
`security_report.md`	Security Reviewer findings by category and severity
`deploy_report.md`	Deployment Validator verdict (DEPLOY_READY / DEPLOY_BLOCKED) with numbered issues

Ops crew

File	Contents
`docs/deliverable_docs.md`	Documentation Writer output: overview, setup, usage, API reference, examples
`memory_summary.md`	Memory Manager report listing what was stored to ChromaDB
`telemetry_report.md`	Observability Monitor trace: agents run, files produced, verdicts, recommendations

Models

Each crew runs on its optimal model by default. Models are configured in config/config.yaml under crew_models - no code changes needed to swap them.

Per-crew defaults (set in crew_models:) and per-agent routing (set in agent_models:):

Crew	Default model	Why
default	`qwen2.5:7b` (plan/research) + `qwen2.5-coder:7b` (build)	Balanced reasoning + specialized coder for generation
safety	`llama3.1:8b`	Fast and reliable for code execution, audit, and validation tasks
ops	`llama3.1:8b`	Strong instruction following for documentation and memory consolidation
plc	`qwen2.5:7b`	Strong instruction following for safety-critical PLC work
react	`qwen2.5-coder:7b`	Purpose-built for code generation - React/JS/TS
vision	`qwen2.5:7b`	Solid reasoning for defect analysis and report generation

Pull the recommended models:

ollama pull qwen2.5:7b           # planning, research, review
ollama pull qwen2.5-coder:7b     # code generation
ollama pull llama3.1:8b          # fast specialist agents (QA, Docs, Memory, Observability)

Override for a single run:

python main.py --goal "..." --model qwen2.5:14b

Swap a crew's default permanently - edit config/config.yaml:

crew_models:
  default: qwen2.5:32b
  plc:     qwen2.5:14b
  react:   qwen2.5-coder:14b   # ← change this

Web UI - the model selector auto-switches to the configured default when you change crews. You can still select any installed model manually.

Fallback models (if 14B models aren't available):

ollama pull llama3.1:8b          # solid all-around fallback
ollama pull mistral:7b           # fast, instruction-tuned

Vision Crew

The Vision Crew integrates with the Vision_Inspect backend to run automated inspection analysis entirely within the agent pipeline.

Prerequisites

The Vision_Inspect FastAPI service must be running before starting the crew:

cd ../Vision_Inspect
uvicorn backend.main:app --host 0.0.0.0 --port 8000

What it does

Step	Agent	Output
1	Vision Analyst	Calls `/inspections` API, parses defect results, writes `vision_findings.md`
2	Vision Reporter	Reads findings + ChromaDB trends, writes `inspection_report.md`
3	Vision QA Validator	Health-checks `/health`, audits report, flags anomalies, writes `vision_qa_report.md`

Run

python main.py --goal "Analyze today's inspection batch and report defect trends" --crew vision

Output files

File	Contents
`vision_findings.md`	Raw inspection statistics: pass/fail counts, defect breakdown by type and zone
`inspection_report.md`	Full QA report: executive summary, statistics, defect analysis, trend comparison, recommended actions
`vision_qa_report.md`	Pipeline health verdict (VISION_QA_PASS / VISION_QA_BLOCK), report completeness check, anomaly flags

Config

Connection settings are in config/config.yaml:

vision_inspect:
  base_url: http://localhost:8000   # Vision_Inspect FastAPI service
  timeout:  30                      # seconds per API call

Configuration reference

config/config.yaml:

llm:
  base_url: http://localhost:11434
  temperature: 0.7
  timeout: 600                # 14B models need more time

crew_models:
  default: qwen2.5:7b        # per-crew model assignments
  plc:     qwen2.5:7b
  react:   qwen2.5-coder:7b
  vision:  qwen2.5:7b
  safety:  llama3.1:8b
  ops:     llama3.1:8b

agent_models:                # per-agent routing
  orchestrator:        qwen2.5:7b
  researcher:          qwen2.5:7b
  builder:             qwen2.5-coder:7b
  critic:              qwen2.5:7b
  qa_agent:            llama3.1:8b
  security_agent:      qwen2.5:7b
  deploy_agent:        qwen2.5:7b
  docs_agent:          llama3.1:8b
  memory_agent:        llama3.1:8b
  observability_agent: llama3.1:8b
  vision_analyst:      qwen2.5:7b
  vision_reporter:     llama3.1:8b
  vision_qa:           llama3.1:8b

llm_fallback:
  model: llama3.1:8b         # used if crew not in crew_models

crew:
  process: sequential        # sequential | hierarchical
  verbose: true
  max_rpm: 10
  output_dir: outputs
  hitl: true                 # human-in-the-loop plan approval

memory:
  persist_dir: memory/chroma
  collection: smith_agentic_memory
  enabled: true

Project structure

smith_agentic/
├── main.py                     # CLI entrypoint
├── requirements.txt
├── README.md
├── config/
│   ├── config.yaml             # model + crew settings
│   └── loader.py
├── agents/
│   ├── orchestrator.py         # default: goal decomposition
│   ├── researcher.py           # shared: general-purpose research (all crews)
│   ├── builder.py              # default: generic deliverable builder
│   ├── critic.py               # default: generic reviewer
│   ├── qa_agent.py             # safety: code execution and QA validation
│   ├── security_agent.py       # safety: OWASP Top-10 vulnerability audit
│   ├── deploy_agent.py         # safety: compile checks and deploy readiness
│   ├── docs_agent.py           # ops: structured markdown documentation
│   ├── memory_agent.py         # ops: ChromaDB knowledge consolidation
│   ├── observability_agent.py  # ops: run telemetry and audit trail
│   ├── plc_planner.py          # plc: ControlLogix architecture + plc_generator/ awareness
│   ├── plc_developer.py        # plc: ladder logic, ST, AOI/UDT, L5X authoring
│   ├── plc_safety_reviewer.py  # plc: NFPA 79, ISO 13849, fault/interlock review
│   ├── ui_planner.py           # react: component tree, data flow, MUI v5
│   ├── ui_builder.py           # react: React 18 + MUI v5, dark industrial theme
│   ├── ui_reviewer.py          # react: correctness, industrial UX, accessibility
│   ├── vision_analyst.py       # vision: queries Vision_Inspect API, parses defects
│   ├── vision_reporter.py      # vision: synthesizes findings + trends into reports
│   └── vision_qa.py            # vision: health-checks pipeline, audits reports
├── tasks/
│   ├── plan.py
│   ├── research.py
│   ├── build.py
│   ├── critique.py
│   ├── revise.py
│   ├── qa_task.py
│   ├── security_task.py
│   ├── deploy_task.py
│   ├── docs_task.py
│   ├── memory_task.py
│   ├── observability_task.py
│   └── vision_tasks.py
├── crews/
│   ├── default_crew.py         # 4-agent crew: plan, research, build, critique
│   ├── safety_crew.py          # QA, security, and deployment validation crew
│   ├── ops_crew.py             # documentation, memory, and observability crew
│   ├── plc_crew.py             # Rockwell Logix / ladder logic crew
│   ├── react_crew.py           # industrial React / MUI crew
│   ├── vision_crew.py          # Vision_Inspect integration crew
│   └── hitl.py                 # human-in-the-loop plan approval helper
├── tools/
│   ├── file_tools.py           # FileReadTool, FileWriteTool, FileListTool
│   ├── search_tool.py          # WebSearchTool (DuckDuckGo)
│   ├── web_fetch_tool.py       # WebFetchTool (raw URL fetch + HTML extraction)
│   ├── code_executor.py        # CodeExecutorTool (subprocess Python runner)
│   ├── git_tool.py             # GitStatusTool, GitStageTool, GitCommitTool, GitPushTool
│   ├── codebase_reader.py      # CodebaseReadTool, CodebaseListTool, CodebaseGlobTool
│   ├── target_repo_tools.py    # read/write/list/glob tools scoped to an external repo
│   ├── vision_inspect_tool.py  # VisionInspectAPITool + Vision_Inspect file I/O tools
│   └── project_file_tool.py    # ProjectFileReadTool (reads files listed in agent context)
├── memory/
│   ├── memory_store.py         # MemoryStoreTool, MemoryQueryTool (ChromaDB)
│   └── chroma/                 # ChromaDB persistent storage (auto-created)
├── ui/
│   ├── server.py               # FastAPI + WebSocket backend (port 8765)
│   └── index.html              # React CDN frontend
├── tests/
│   ├── conftest.py             # sys.path setup for all tests
│   ├── test_config.py          # config loader: crew models, agent models, target repo
│   ├── test_file_tools.py      # FileReadTool, FileWriteTool, FileListTool
│   ├── test_codebase_reader.py # CodebaseReadTool, CodebaseListTool, CodebaseGlobTool
│   ├── test_vision_inspect.py  # VisionInspectAPITool + file I/O tools
│   ├── test_memory.py          # MemoryStoreTool, MemoryQueryTool, create_memory_tools
│   └── test_server.py          # FastAPI endpoints: status, outputs, run, crew validation
└── outputs/                    # all agent outputs land here
    └── .gitkeep

Adding a new crew

Create crews/my_crew.py with a build_crew(goal, config) function returning a crewai.Crew.
Add the crew name to _CREW_BUILDERS in main.py.
Add the crew name to the builders dict in ui/server.py and to the api_crew_defaults return value.
Add a default model entry under crew_models: in config/config.yaml.
Run with: python main.py --goal "..." --crew my_crew

Adding a new agent

# agents/my_agent.py
from crewai import Agent, LLM

def create_my_agent(llm: LLM, tools: list, verbose: bool = True) -> Agent:
    return Agent(
        role="My Agent Role",
        goal="What this agent is trying to accomplish.",
        backstory="Domain expertise and constraints for this agent.",
        llm=llm,
        tools=tools,
        verbose=verbose,
    )

Create agents/my_agent.py with a create_my_agent(llm, tools, verbose) function.
Create the corresponding task in tasks/my_task.py.
Import and wire the agent into the relevant crew file (crews/default_crew.py or a new crew).
Optionally add a per-agent model entry under agent_models: in config/config.yaml.

Adding a new tool

# tools/my_tool.py
from crewai.tools import BaseTool
from pydantic import BaseModel, Field
from typing import Type

class _MyInput(BaseModel):
    param: str = Field(description="What this parameter does.")

class MyTool(BaseTool):
    name: str = "My Tool Name"
    description: str = "What this tool does and when to use it."
    args_schema: Type[BaseModel] = _MyInput

    def _run(self, param: str) -> str:
        return f"Result for: {param}"

Then pass an instance to any agent's tools=[] list in the crew file.

Troubleshooting

"Cannot connect to Ollama" Run ollama serve in a separate terminal. The default URL is http://localhost:11434.

ChromaDB install fails On some systems: pip install chromadb --no-build-isolation. If it still fails, set memory.enabled: false in config.yaml to disable memory entirely.

Web UI shows blank / won't load Make sure python ui/server.py is running and port 8765 is not blocked. The UI uses CDN React - you need an internet connection on first load.

Runs are slow

Use a smaller model (llama3.1:8b, mistral:7b)
Increase timeout in config.yaml if hitting timeouts
Set verbose: false to reduce console I/O

Agent loops or repeats itself Lower max_iter in the agent file. Use --no-hitl to skip plan approval if it loops there.

DuckDuckGo search fails DDG rate-limits aggressively. The tool fails gracefully and the agent continues without search results.

Navigation Menu

Repository navigation

About

Repository files navigation

SmithAgentic - Local Multi-Agent AI System

What it does

Agent Team

Agents

Default crew agents

Safety crew agents

Ops crew agents

PLC crew agents

React crew agents

Vision crew agents

Features

Prerequisites

Setup

Running - CLI

Running - Web UI

Human-in-the-Loop Plan Approval

ChromaDB Memory

Output Files

Models

Vision Crew

Prerequisites

What it does

Run

Output files

Config

Configuration reference

Project structure

Adding a new crew

Adding a new agent

Adding a new tool

Troubleshooting

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
agents		agents
config		config
crews		crews
memory		memory
outputs		outputs
scripts		scripts
tasks		tasks
tests		tests
tools		tools
ui		ui
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
launch_smith_agentic.bat		launch_smith_agentic.bat
main.py		main.py
openapi.json		openapi.json
pytest.ini		pytest.ini
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

SmithAgentic - Local Multi-Agent AI System

What it does

Agent Team

Agents

Default crew agents

Safety crew agents

Ops crew agents

PLC crew agents

React crew agents

Vision crew agents

Features

Prerequisites

Setup

Running - CLI

Running - Web UI

Human-in-the-Loop Plan Approval

ChromaDB Memory

Output Files

Models

Vision Crew

Prerequisites

What it does

Run

Output files

Config

Configuration reference

Project structure

Adding a new crew

Adding a new agent

Adding a new tool

Troubleshooting

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages