Local AI agents in minutes
Build, run, and ship — fully local.
Introducing Forge by l6e, now in alpha:
- Start < 1 minute, ship < 1 hour: a few lines of Python and custom prompts.
- Run while you sleep: fully local agents with no surprise cloud bills.
- Save money: prove your MVP free on‑device; upgrade to OpenAI or clusters later.
Auto Models picks and installs the best OSS model for your hardware. Ships with Qdrant memory, Postgres storage, and a built‑in dev UI.
A.I. Phone Home
$ pip install l6e-forge l6e-forge-cli$ forge create agent assistant --template=assistant$ forge pkg build agents/assistant -o dist📦 Packaged dist/assistant-0.1.0.l6e (ready to ship)$ forge pkg install dist/assistant-0.1.0.l6e -w /workspaces/prod$ forge up🚀 Agent running on your stack✨ Your AI runs where you need it.
Highlights
A focused toolkit for building private, portable AI agents that start local and scale when you need.
Auto Models (alpha)
One command picks and installs the right OSS model for your hardware. Apple Silicon optimized. Ollama today; more soon.
Agents on Rails
Opinionated defaults and tooling to build fast without lock‑in. Start local, keep interfaces stable as you grow.
Portable Agents
Ship a single portable package and run it anywhere—dev to prod.
Local-first
Run on laptops or your VPC with zero required cloud services; private data by default.
Personal-scale by design
Perfect for single-user assistants. Runs great on Apple Silicon; scale up only when you need to.
Optional Docker stack
Spin up API, monitoring, chat UI, and relational + vector storage with one command. Swap components for your infra when ready.
Your stack, your data
Agents run entirely on your infrastructure—developer machines or internal cloud. Data never leaves unless you say so.
Fast Dev Loop
Hot reload, instant testing, and a focused CLI to move fast.
Ready-made Templates
Start from proven assistants, coding, and research patterns—customize in minutes.
Pluggable Runners
Use LM Studio or Ollama out of the box. llama.cpp support is coming soon.
Interfaces over rewrites
Adapt existing agents or frameworks (e.g., LangChain) without rebuilding your codebase.
MVP to cloud path
Ship a single-user MVP quickly, then move to multi-user cloud with the same core contracts.
Vector memory out of the box
Use Qdrant for fast, scalable embedding search with zero setup.
Conversations in Postgres
Relational storage for conversations and base messages using Postgres.
Built-in Agent UI
Develop and interact with agents via an out-of-the-box chat and tools UI.
Agents on Rails
Rails for agents: opinionated defaults to ship MVPs fast. Bring up a full local stack via Docker—API, monitor, chat UI, and pluggable stores (Qdrant/LanceDB, Postgres)—no cloud required. Base stack is single-user; scale by deploying backing services and swapping adapters. Unlike classic Rails, Forge is adapters-first and ecosystem-friendly: conventions speed you up, interfaces keep you from lock-in.
Full local stack
Run API, monitor, and chat UI with one command.
Portable packages
Ship a single .l6e file and run it anywhere.
BYO stores
Use Qdrant/LanceDB for vectors and Postgres for state.
Opinions without lock-in
Adapters-first: pick runners, memory, and DBs; swap providers as you grow.
# Agents on Rails$ pip install l6e-forge l6e-forge-cli$ forge create agent my-agent --template=assistant$ forge up# Local stack: API (8000), Monitor (8321), Chat UI# Single-user by default (not horizontally scalable)# Scale later: deploy Qdrant/Postgres and switch adapters# Adapters-first: swap providers without rewrites
Auto Models
Ship once; run anywhere. Forge detects the system and installs an optimized open-source model automatically—so developers aren't forced to handpick models and their users don't sweat the details. Tailored for personal-scale agents and Apple Silicon; currently in alpha with Ollama support.
For developers
Build independently of hardware and model constraints. Package your agent, and let Auto Models select the best fit at install or run time.
For users
Run the agent; Forge detects your system and pulls the right models automatically—no setup.
Current support
Alpha, Apple Silicon optimized. Ollama today; LM Studio discovery supported. More providers coming soon.
# Auto Models (alpha)# Inspect and sanity-check your setup$ forge models doctor# Pick models interactively and write config$ forge models bootstrap --agent agents/test-demo --provider ollama# Non-interactive: accept defaults and skip the picker$ forge models bootstrap --agent agents/test-demo --provider ollama --yes# Run the agent (first run pulls missing models automatically)$ forge up
- OS, CPU cores, RAM, GPU/VRAM
- Providers up (Ollama, LM Studio)
- Internet availability
- Merge installed models + curated catalog
- Include provider tags and alternates
- Apple Silicon friendly defaults
- Use installed sizes when available
- Else catalog or param based estimate
- Compare to VRAM or system RAM
- Prefer installed and available
- Fit first; apply quality hint
- Respect provider order
- Resolve tags; pull sensible alternates
- Write chat + embedding to config.toml
- First run auto pulls if missing
- OS, CPU cores, RAM, GPU/VRAM
- Providers up (Ollama, LM Studio)
- Internet availability
- Merge installed models + curated catalog
- Include provider tags and alternates
- Apple Silicon friendly defaults
- Use installed sizes when available
- Else catalog or param based estimate
- Compare to VRAM or system RAM
- Prefer installed and available
- Fit first; apply quality hint
- Respect provider order
- Resolve tags; pull sensible alternates
- Write chat + embedding to config.toml
- First run auto pulls if missing
Pluggable Runners & Interop
Ollama
Pull and run open LLMs locally. Multi-model workflows supported.
LM Studio
Local model runner with desktop UX. Use alongside l6e/forge for zero-cloud dev.
llama.cpp (coming soon)
Lightweight, CPU/GPU-friendly runner. Bring-your-own binaries and configs.
Your stack
Interfaces over rewrites. Adapt existing agents or frameworks (LangChain, custom services) into your cloud when you need it.
Deploy Your First AI Agent
l6e/forge is "Agents on Rails": ship an MVP fast with opinionated defaults, then scale on your own infrastructure. Start local on Apple Silicon; move to your VPC or cloud by swapping the same interfaces.
# Install l6e Forge$ pip install l6e-forge l6e-forge-cli# Create and customize your agent$ forge create agent my-agent --template=assistant$ # Edit agents/my-agent/agent.py with your logic# Optional: Auto-pick local models (alpha; Apple Silicon optimized; Ollama only)$ forge models doctor$ forge models bootstrap --agent agents/my-agent --provider ollama# Test locally$ forge chat my-agent# Package for distribution$ forge pkg build agents/my-agent -o dist📦 Created dist/my-agent-0.1.0.l6e (portable agent)# Install into another workspace (e.g., your VPC or another machine)$ forge pkg install dist/my-agent-0.1.0.l6e -w /path/to/workspace# Bring up the optional Docker stack (API, Monitor, Chat UI, Memory)$ forge up🚀 Stack running locallyAPI: http://localhost:8000Monitor: http://localhost:8321Chat UI: served by API (root)Memory: in-memory by default (swap when scaling)