MageTech AI Forge
MageTech AI Forge
Brand: MageTech Solutions
SKU: MTS-AIFORGE-001
A local-first, open-source AI application and agent platform. It compounds documents, knowledge bases and AI workflows to accelerate agencies and businesses — with RAG, autonomous agents, MCP tooling, multi-tenant workspaces and observability on a self-hosted stack.
Choose Your Plan
Flexible pricing for every business size. Start free, scale as you grow.
CommunityFor developers, students, experimentation and POCs. Free forever.
₹0
/mo
Get Started
|
Most Popular
DeveloperFor individual developers and AI builders. Everything in Community plus advanced RAG, LangGraph workflows and API access.
₹2,499
/mo
Get Started
or ₹24,990/yr
Save 17%
|
TeamFor teams building internal AI applications together. Everything in Developer plus RBAC, shared knowledge bases and audit logs.
₹8,499
/mo
Get Started
or ₹84,990/yr
Save 17%
|
BusinessFor companies building internal AI applications and business agents at scale.
₹20,999
/mo
Get Started
or ₹209,990/yr
Save 17%
|
EnterpriseFor larger organizations requiring private AI infrastructure and full control. Custom pricing.
Custom
Get Started
Tailored for your requirements
|
|---|---|---|---|---|
|
1
Products
|
5
Products
|
20
Products
|
Unlimited
Products
|
Unlimited
Products
|
|
1
Users
|
3
Users
|
10
Users
|
25
Users
|
Unlimited
Users
|
|
0
WhatsApp Messages/Day
|
0
WhatsApp Messages/Day
|
0
WhatsApp Messages/Day
|
0
WhatsApp Messages/Day
|
0
WhatsApp Messages/Day
|
|
Unlimited
AI Queries/Day
|
Unlimited
AI Queries/Day
|
Unlimited
AI Queries/Day
|
Unlimited
AI Queries/Day
|
Unlimited
AI Queries/Day
|
|
1 user · 1 project
Local LLM / Ollama + Knowledge Base
Basic RAG · PDF / DOCX / TXT / MD upload
Basic AI Chat · basic embeddings
3 AI agents · 3 MCP tools
Basic conversation memory
Self-hosted Docker deployment · Community support
MageTech AI Forge branding
|
Everything in Community
5 projects · 1–3 users
Unlimited knowledge bases
25 agents · 25 MCP tools
Advanced RAG · agent memory
LangGraph workflows · agent execution history
Agent testing · API access · AI Gateway
Document source citations · Basic AgentOps
Custom prompts · Email support
|
Everything in Developer
10 users · 20 projects · unlimited agents & MCP tools
Advanced AgentOps · team collaboration
Role-based access (Admin / Manager / Developer / Viewer)
Shared knowledge bases · advanced memory
Agent execution monitoring · audit logs
MCP permissions / allowlists · API + webhooks
Multiple AI providers · advanced RAG
Priority support
|
Everything in Team
25 users · unlimited projects, KBs, agents, MCP tools
Advanced AgentOps · advanced RBAC · audit trail
SSO · advanced security controls
Private AI Gateway · multiple LLM providers
Custom MCP & business-system integrations
Usage analytics · advanced monitoring
Custom retention policies · deployment assistance
Priority support
|
Unlimited users & projects
Private / on-premise / Kubernetes deployment
Dedicated AI infrastructure · private local LLM environment
Custom MCP servers & custom integrations
SSO / SAML · advanced RBAC · audit logs
Security reviews · data-retention policies
Dedicated support · SLA · architecture consultation
Custom onboarding · custom contract / invoicing
|
Your Local-First AI Application & Agent Platform
MageTech AI Forge is a full-stack AI application and agent platform that combines retrieval-augmented generation (RAG), autonomous agents and the Model Context Protocol (MCP) into one product surface — deliberately local-first, so customers keep data on hardware they control.
100%
Local-First · No Mandatory API Keys800
Char Chunking / 160 Overlap for RAG384-dim
Embeddings · Cosine Retrieval1
docker compose up to DeployExecutive Summary
MageTech AI Forge compounds documents, knowledge bases and AI workflows to accelerate agencies and businesses. The default runtime is a fully self-hosted stack with Ollama serving open-weight models, so data never leaves the deployment boundary. The architecture is provider-agnostic: managed providers (OpenAI, Google Gemini, Anthropic, Azure OpenAI) can be added later without redesigning the core.
Core Value Proposition: Run the entire AI stack locally by default — no mandatory third-party API keys — while exposing every capability through a documented, versioned HTTP API backed by security, rate limiting, audit logging and RBAC-ready tenancy from day one.
Product Pillars
Knowledge & RAG
Upload documents, parse and chunk them, embed and store them with cosine search, then answer questions grounded in the knowledge base with citations.
Agents
Persistent agent runs that can call MCP tools, remember context and stream step-by-step execution. LangGraph orchestration is the planned evolution.
MCP
A first-class Model Context Protocol layer: connect, enable/disable and invoke external tools through a standardized JSON-RPC surface with strict security gates.
Workspaces
Multi-tenant workspaces isolate knowledge, agents, conversations and tools with column-level tenancy on every business entity.
Observability
Structured JSON logs, token accounting, agent-run traces and usage analytics inform cost, quality and debugging decisions.
Extensible
Webhook events, a future MCP SDK, ServiceFlow/CRM integrations and OAuth2 2-legged flows keep the platform open.
System Context
Design Principles
| Principle | What it means here |
|---|---|
| Local-first | Default runtime is self-hosted with open-weight models; no external dependency is mandatory. |
| Open source | Transparent, auditable stack; no black-box components in the critical path; OpenAPI schema at /docs. |
| Provider-agnostic AI | LLMClient gateway isolates provider SDKs behind one protocol-agnostic interface. |
| Secure by default | Fail closed: argon2id hashes, short-lived JWTs, refresh cookies, rate limits, production guards, no tool can shell out. |
| Tenant isolation | Every business entity belongs to a workspace; queries are always scoped by tenant (X-Workspace-Id). |
| Async-first backend | Async SQLAlchemy + asyncpg, async Redis, SSE streaming via StreamingResponse. |
| Reliability & observability | Every request is traceable; structlog JSON logging, AppError envelope, audit log, route executions, health endpoints. |
| Evidenced roadmap | Everything labelled planned/future has written design intent (LangGraph, LlamaIndex, MCP SDK, ai-gateway, pgvector, RLS). |
Technology Stack
Frontend
| Next.js | 16.3.5 (App Router) |
| React | 19.2.8 |
| TypeScript | 5.9 |
| Tailwind CSS | 4.3.3 |
| TanStack Query | 5.103.2 |
| react-hook-form + Zod | 7.88 / 4.6.5 |
Backend
| Python | 3.12 |
| FastAPI | 0.141 |
| SQLAlchemy 2.x | 2.0.54 + asyncpg |
| Alembic | 1.20 |
| Argon2-cffi / PyJWT | 25.1 / 2.14 |
| redis / slowapi | 6.4 / 0.1.10 |
AI, Data & Infrastructure
| Ollama | qwen2.5:7b + nomic-embed-text |
| PostgreSQL | 16 (pgvector-ready) |
| pgvector | Extension provisioned |
| Redis | 7-alpine |
| LangGraph / LlamaIndex | Planned |
| OpenAI / Gemini / Anthropic | Future providers |
AI Architecture — LLMClient Gateway
The AI tier is built around a single provider gateway (services/llm_client) so the rest of the platform never talks to a specific model vendor. Today the gateway speaks to Ollama; tomorrow it can route to OpenAI, Gemini or Anthropic with configuration only.
Provider selection
LLM_PROVIDER=auto probes the local Ollama tag endpoint with a 30-second availability cache. If no provider is reachable the gateway degrades to a deterministic mock provider so tests and demos stay stable.
Token accounting
Streaming responses report eval_count from Ollama; when omitted, tokens are estimated with len(text)//4. This feeds usage analytics and per-model cost views.
Knowledge & Retrieval (RAG)
Ingestion pipeline
Upload (multipart PDF/TXT/MD/CSV, capped at 2,000,000 chars) → parse → chunk (size 800 / overlap 160) → embed (nomic-embed-text, 384-dim) → store in PostgreSQL. Async reprocess supported.
Retrieval & answering
Query embedding → top-k cosine retrieval (workspace-scoped) → grounded prompt with source citations. Answers link back to source documents.
Vector storage evolution
Now: JSON-array brute-force cosine (LIVE). Near: native vector(384) column + pgvector ANN index (PLANNED). Later: hybrid BM25 + re-ranks (FUTURE).
Agents & LangGraph
Agents in the MVP are persistent, tool-capable runtimes: create an agent, attach a system prompt and model, then run it with input. Runs produce a stream of steps (tool calls and text) recorded for audit and replay.
Run lifecycle
pending → running → completed | failed. The executor generates a response, invokes enabled MCP tools when requested, and feeds results back until the loop terminates.
Loop bound
Agent tool loops are bounded by a maximum iteration count enforced in the executor to prevent infinite tool-call cycles and runaway cost.
MCP Integration
| Tool | Description | Safety posture |
|---|---|---|
| memory_store | Persist a key/value memory entry. | Workspace-scoped writes, audit logged. |
| memory_recall | Read back stored memories. | Workspace-scoped reads only. |
| get_current_time | Return the current date/time. | Read-only, no side effects. |
| calculator | Evaluate arithmetic expressions. | AST whitelist (arithmetic/numeric ops only). No shell, no OS, no dynamic import. |
MCP stdio transport is rejected by design — it would allow the platform to spawn arbitrary OS processes. Remote calls are enable-gated and audit-logged.
Data & Multi-Tenancy
Live schema
users · workspaces · user_workspace_roles · audit_logs · knowledge_bases · documents · document_chunks · agents · agent_runs · conversations · messages · mcp_servers · mcp_tools · models
Planned: roles, permissions, projects, agent_versions, agent_tools, agent_execution_steps, memories, token_blacklist, subscriptions, usage_limits.
Tenant model
A workspace is the tenant boundary. Every business table carries workspace_id; the API validates X-Workspace-Id against membership and falls back to the user's first workspace. IDOR tests assert cross-workspace leakage is impossible. PostgreSQL RLS is planned as defense-in-depth.
API Design & Security
Conventions
- Base URL
/api/v1; OpenAPI 3.1 at/docs - Bearer access JWT + rotating HttpOnly refresh cookie
- Uniform error envelope via AppError
- SSE streaming:
meta → delta* → donethendata: [DONE] - Routers: auth, chat, agents, knowledge, mcp, models, usage, executions, dashboard
Security measures
- Argon2id password hashing with strict policy
- Dual JWT: 15-min access + 7-day rotating refresh; separate secrets
- slowapi rate limits: auth 10/min, default 200/min (Redis + in-memory fallback)
- Production guard refuses default/weak secrets in prod
- No raw SQL; Pydantic v2 validation; MCP stdio rejected; calculator AST whitelist
- Structured audit log for auth, workspace, MCP and document events
Observability, Performance & Safety
Observability
structlog JSON logging with request_id join keys · usage analytics (?days=7..90, per-model) · executions dashboard · health endpoint (PG/Redis/Ollama) · immutable audit log.
Performance
Async I/O (asyncpg, async Redis), connection pooling (10/20, pool_pre_ping), 30s provider availability cache, SSE token streaming (TTFB = first token), bounded ingestion caps.
AI safety
No arbitrary execution · grounded answers with citations · tenant isolation (IDOR-tested) · loop bounds · audit trail. Local-first is itself a data-governance win: knowledge and prompts never leave the deployment boundary.
Roadmap
| Milestone | Content | Status |
|---|---|---|
| MVP-1 | Auth (JWT + argon2id), workspaces, RAG pipeline (upload→embed→chat), dashboard, streaming chat. | LIVE |
| MVP-2 | Agents + runs + executions, MCP server/tool management, usage analytics, multi-model registry. | LIVE |
| MVP-3 | RBAC enforcement, settings system, refresh-token blacklist, demo mode polish, E2E expansion. | PLANNED |
| MVP-4 | Native pgvector search, document reprocess UX, webhook delivery, OAuth2 2-legged tokens. | PLANNED |
| Phase 4 | LangGraph agent graphs (state machine, checkpoints, memory), LlamaIndex orchestration. | PLANNED |
| Phase 5 | MCP SDK: expose platform tools to external agents; mcp-sdk-python companion. | PLANNED |
| Phase 7 | Managed providers (OpenAI/Gemini/Anthropic) behind LLMClient, ai-gateway service, hosted offering. | FUTURE |
Compounding Documents, Knowledge & AI Workflows
Local-first · Open source · One docker compose up · RAG + Agents + MCP + Observability
| Product Type | Local-First AI Application & Agent Platform |
| Version | 1.0.0 |
| Status | SCALED MVP |
| License | Open Source (Local-First) |
| Backend | FastAPI 0.141 · Python 3.12 · Uvicorn 0.53 |
| Frontend | Next.js 16.3.5 (App Router) · React 19.2.8 · TypeScript 5.9 |
| Styling | Tailwind CSS 4.3.3 + shadcn/ui pattern |
| Server State | TanStack Query 5.103.2 · react-hook-form + Zod 4.6.5 |
| ORM | SQLAlchemy 2.0.54 + asyncpg 0.31 |
| Database | PostgreSQL 16 (pgvector-ready) · Alembic 1.20 migrations |
| Vector Search | JSON-array cosine (LIVE) → native pgvector ANN (PLANNED) |
| Cache / Queue | Redis 7-alpine · Celery 5.6.3 (opt-in worker) |
| LLM Serving | Ollama — qwen2.5:7b (generation) + nomic-embed-text (384-dim embeddings) |
| Managed Providers | OpenAI / Gemini / Anthropic / Azure OpenAI (FUTURE) |
| Agent Orchestration | MVP linear runtime (LIVE) → LangGraph state graphs (PLANNED) |
| Indexing | MVP pypdf/parser pipeline (LIVE) → LlamaIndex orchestration (PLANNED) |
| Auth | Argon2id + dual JWT (HS256, 15-min access / 7-day rotating refresh cookie) |
| Rate Limiting | slowapi — auth 10/min · default 200/min (Redis + in-memory fallback) |
| API | REST /api/v1 · OpenAPI 3.1 at /docs · SSE streaming |
| Tenancy | Column-level workspace_id + X-Workspace-Id · RLS (FUTURE) |
| Observability | structlog JSON logs · audit_logs · usage analytics · health endpoints |
| Chunking | CHUNK_SIZE 800 · CHUNK_OVERLAP 160 · MAX_DOCUMENT_CHARS 2,000,000 |
| MCP Transport | SSE JSON-RPC (enabled) · stdio rejected for safety |
| Deployment | docker compose up — nginx, frontend (3000), backend (8000), worker, postgres (:5433), redis (:6380), ollama (:11434) |
| Security | No arbitrary execution · calculator AST whitelist · SSRF posture on MCP · production guards |
| Streaming Protocol | SSE events: meta → delta* → done → data: [DONE] |
| Plans | Community (Free) · Developer ₹2,499/mo · Team ₹8,499/mo · Business ₹20,999/mo · Enterprise (Custom) |
| Company | MageTech Solutions |
AI Forge — Demo Document
View the complete product documentation below.
Related Products
You May Also Like
MTS AI Commerce Assistant
MTS AI WhatsApp CRM
₹4,999.00 /mo after trial
MTS Laravel RBAC Package
Need Help Choosing?
Our team is ready to help you find the perfect solution for your business.
Get a Free Quote