v1.1.2 · Now available

The AI you can trust
with long-term work.

A self-hosted, local-first AI assistant with persistent memory. It remembers your projects across sessions and shows its work — every routing decision is logged and replayable — all on your own hardware.

Start free — self-hosted View on GitHub →

Memory

Remembers across sessions

Provenance

Every decision can be replayed

Private

Nothing leaves your machine

Who It's For

Built for people who need control.

Not another chatbot. A private AI platform for developers and engineers who want expert help without cloud dependency.

Solo Developers

Debug code and build features faster
Keep project memory across sessions
Run locally with no subscription

Homelab & IT Engineers

Network troubleshooting and DNS
Linux administration and scripting
Infrastructure planning and review

AI Builders

Local-first agentic platform
Persistent memory and planning
Full observability into AI decisions

Privacy-Conscious Teams

No cloud lock-in or data exposure
Full audit trail of every decision
Self-hosted on your own infrastructure

Dashboard

See how the system thinks.

Every decision is logged, visualized, and replayable. Not a black box — a glass box.

UCI Dashboard

UNIFIED COGNITIVE INDEX · h_UCI 80.8 ↑ +5.7

Reliability · Intelligence · Adaptation · Productivity

Intelligence Metrics

Plan Graph

EXECUTION DAG

parse

→

route

→

reason

↓

verify

→

respond

Plan Graph

Memory Browser

592 MEMORIES · FAISS BACKEND

python_dev

ai_ml

it_network

knowledge

Memory Browser

Decision Replay

THEN vs NOW

THEN

NOW

agent: same · confidence: improved · reflect: reduced

Decision Replay

Why AMAGRA

Not a chatbot. A team.

Generic AI gives generic answers. Specialist agents give expert answers — every time, locally, privately.

Complete Privacy

Runs entirely on your hardware. Conversations, code, and context never leave your machine. No telemetry, no external logging, no training on your data.

Deterministic Routing

A geometric scoring engine classifies intent in milliseconds — not an LLM call. Result: 98% routing accuracy at sub-second latency, every time.

Persistent Memory

Semantic memory across sessions using a FAISS vector backend with 52× LRU cache speedup. The system remembers your projects, lessons, and context.

Full Observability

Every routing decision is logged with confidence scores, domain signals, and regret metrics. You can audit exactly why the system chose what it chose.

Consumer Hardware

Runs on an RTX 2050 with 4GB VRAM. No A100, no cloud GPU, no inference bill. phi4-mini via Ollama delivers specialist-quality answers on a gaming laptop.

Self-Learning

Thumbs-up/down feedback adjusts agent weights through a learning kernel. The system adapts to your usage patterns over time and gets better the more you use it.

Cognitive OS

More than chat.

Most tools stop at the response. Amagra lets you see inside the process — replay decisions, inspect memory, trace every step.

Plan Graph

Visualize execution as a dependency graph. See which steps ran in parallel, which blocked, and why.

Decision Replay

Re-run any past query and compare how the system would decide now vs then. Measure improvement over time.

Event Log

A live stream of every cognitive event — routing, reflection, memory access, contradiction detection.

Memory Browser

Browse, search, prune, and consolidate semantic memories. See quality scores and usage frequency for every entry.

Risk Observatory

Track reflection triggers, critic gate acceptance rates, and routing conflict rates in real time.

Step Verification

Each goal step is scored before delivery. Low-confidence responses trigger automatic retry with a higher reflection level.

Specialist Agents

Each expert knows its domain.

The routing engine reads every query and dispatches to the right specialist in under one second — automatically.

Python Dev

Python, FastAPI, asyncio, pandas. Complete, runnable scripts — not code fragments. Handles edge cases and imports.

.NET Dev

C#, ASP.NET Core, Blazor, Entity Framework. Full working code for the .NET ecosystem including DI, async, and xUnit tests.

Web Dev

React, TypeScript, Next.js, Node.js, Tailwind. Full-stack web from hooks to APIs. Handles build tooling and async state.

IT & Networking

Wi-Fi, DNS, SSH, VPN, Linux networking. Runs real diagnostic commands and gives specific, actionable fixes.

AI & ML

PyTorch, TensorFlow, LLMs, embeddings, fine-tuning. Explains concepts clearly and implements training pipelines.

DevOps

Docker, Kubernetes, GitHub Actions, Bash, systemd. Dockerfiles, Compose configs, CI/CD pipelines, and deployment scripts.

Data Analyst

pandas, SQL, NumPy, matplotlib. Data cleaning, EDA, and statistical summaries with working visualization code.

Writer

READMEs, API docs, blog posts, commit messages. Technical writing that developers actually want to read. Edits and proofreads.

Knowledge

Explains any concept clearly. Saves lessons to memory so you never lose what you learn. Corrects wrong analogies directly.

Terse

One-line answers. The command, the syntax, the flag. No explanation unless you ask. Fastest path to the answer you need.

How It Works

From query to expert answer in seconds.

The routing pipeline is deterministic and auditable at every step.

Signal Extraction

Your query is scored against domain-specific keyword registries. Each token votes for one or more specialist domains — no LLM required.

Confidence Routing

If one domain scores above threshold with no competitor, it routes deterministically. Ambiguous queries use phi4-mini as tiebreaker.

Specialist Response

The specialist receives your query with domain context and relevant memory injected. Response time: 2–8 seconds depending on complexity.

Memory + Learning

The response is stored as a semantic memory. Your feedback adjusts agent weights through the learning kernel — improving future answers.

First 5 minutes

Run Docker

docker-compose up

Open Dashboard

localhost:3000

Ask a Question

Any domain

Watch Routing

See the decision

Explore Memory

Plans & replay

        # Install & run in 3 commands

        git clone https://github.com/d4shm1r/amagra

        cd agentic-ai && cp .env.example .env

        docker-compose up

        # → Dashboard: http://localhost:3000

        # → API:       http://localhost:8000

How We Compare

Built different.

Cloud AI tools are powerful but opaque, dependent, and expensive at scale. Amagra is the alternative for engineers who want control.

Feature	AMAGRA	ChatGPT	Claude	Copilot
Runs locally (no cloud)	✓	✗	✗	✗
Persistent memory	✓	Limited	Limited	✗
Decision replay	✓	✗	✗	✗
Plan graph (DAG)	✓	✗	✗	✗
Event log & observability	✓	✗	✗	✗
Self-hosted	✓	✗	✗	✗
Routing audit trail	✓	✗	✗	✗
Learns from feedback	✓	Via RLHF	Via RLHF	✗
No subscription required	✓ Self-hosted	✗	✗	✗

Built in Public

Every phase, in public.

An open architecture and a full, documented API. Nothing is hidden — you can see exactly how it works.

Development phases, in the open

Specialists on call

100+

Documented API endpoints

<1ms

To recall what it knows

100%

On your hardware

MIT

Open source — yours to run

Pricing

Start free. Pay when it saves you time.

The self-hosted version is free forever. Pro adds managed hosting with a frontier model backend.

Self-Hosted

Run on your own hardware with your own models. Full access, no subscription required, no limits.

All specialist agents
Local phi4-mini (4GB VRAM)
Unlimited queries
Full memory + learning system
Mission Control dashboard
Plan graph, replay, event log
Managed hosting
Claude / GPT-4o backend

Get the Code

FAQ

Do I need a GPU?

A 4GB VRAM GPU is recommended for phi4-mini. The system also runs on CPU with reduced throughput. Any modern NVIDIA RTX card works out of the box.

What models does it support?

Any model served by Ollama — phi4-mini by default. Swap to Llama 3, Mistral, Gemma, or any GGUF-compatible model by changing a single config line.

Does data ever leave my machine?

No. The self-hosted version runs entirely locally — inference, memory, and routing all happen on your hardware. Zero telemetry.

How is routing handled?

A signal-first QuerySignal pipeline classifies each query using phrase-weighted keyword scoring across a 21-node skill graph, hitting 98% accuracy without any external API calls.

What is h_UCI?

The Hierarchical Unified Cognitive Index — a composite score across Reliability, Intelligence, Adaptation, and Productivity. Current value: 80.8/100. It updates live as the system processes queries.

Is there a managed cloud option?

Yes — the Developer and Team plans (see Pricing) provide managed hosting with UCI dashboard persistence, no GPU required. Contact us to get on the early access list.

The AI you can trust with long-term work.

Built for people who need control.

See how the system thinks.

Not a chatbot. A team.

More than chat.

Each expert knows its domain.

From query to expert answer in seconds.

First 5 minutes

Built different.

Every phase, in public.

Start free. Pay when it saves you time.

FAQ

Your AI. Your hardware. Your data.

The AI you can trust
with long-term work.