Systems-design-first coding harness

Design the system.
Then let it build itself.

Point Lucid Train at your repo and it draws the architecture, one act at a time. Then hand that diagram to the coding agent as a specification and review the diffs that come back. It runs as a native desktop app, or as one 6 MB Rust binary in your terminal.

Download for macOS.dmg · Apple Silicon & Intelor for Windows
$curl -fsSL https://lucidtrain.com/install.sh | sh
prefer the CLI? →
DETERMINISTIC LAYOUT · DESIGN.MD IS THE SPEC · YOUR MODELS
LucidTRAIN
CodeDesignAuditDashboardsConnectorsDiagramShorts
↗ Open in ▾📁 opensre
Design

Diagram

Auto-align ✓Story+ LibrariesExport actsSave to READMEsImplement in code →Chat ✕

Draw architecture and flows, or ask the agent to diagram your codebase. Versions save as you go.

Diagrams
5 acts
repo root
23 nodes
repo root
16 nodes
gateway
flow
tools/investigation
Versions (3)
label (optional)Save version
v3just now
icons resolved12m
v11h
A five act storyboard of the opensre system drawn by Lucid Train: Context, Containers, Data flow, Data and Failure modes, each a full architecture diagram with a narrator character explaining it
20%
click to explore · drag to pan
Diagramopensre
Lucid agent
walk me through this system
Surveyed the repo. Five acts: Context, Containers, Data flow, Data, then Failure modes. Drawing each one with a narrator beside it, then I will screenshot every act and check my own alignment.
• emit_diagram · act 3 of 5 · 13 nodes · 16 edges
• auto-align · pass 2 · 0 overlaps
✓ saved docs/opensre-story/act-3-data-flow.svg
Manual ⇧⇥12K ctx · ↑4K ↓2K
Enter sends. Shift+Enter adds a line. Shift+Tab changes mode. ↑↓ history.
Trial · 7d leftAArnab

Four real diagrams of the opensre repository, drawn by the agent and exported straight out of the app. Pick one from the rail, drag to pan, scroll to zoom.

Systems design

The diagram is the specification.

Ask for a diagram and the agent surveys the repo, then narrates the system one act at a time. Each act is a real architecture drawing with a narrator beside it. You never place a box: the agent emits components and the edges between them, and the app runs an ELK layout to position and route everything, so a diagram cannot come out overlapping.

ACT 1

Context

The whole system in five to eight boxes, and the edges between them.

ACT 2

Containers

Every service, banded by tier: client, application, data.

ACT 3

Data flow

One request end to end. The path stays lit and everything else dims.

ACT 4

Data

The stores, and what actually gets written where.

ACT 5

Failure

The single points of failure, marked as problems.

The handoff

Implement in code →

One click turns the canvas into a markdown spec, every component, every relationship and the narrator's notes, then starts a real turn in the Code tab with it. Your DESIGN.md rides along as the authoritative specification, and a 17 KB one fits whole. The README goes with it as orientation. The agent implements against the drawing rather than against a paragraph you had to write twice.

the spec it sends
# Implement this architecture: checkout
## Components · ## Relationships · ## Notes
## Design document (DESIGN.md, authoritative)
## Project README (context, not the task)
It checks its own work

After drawing, the app screenshots each act and hands the image back to the agent to judge, a vision pass over its own output. It fixes what is wrong and re-emits the act in place rather than piling a correction up beside the original.

It lands in the repo

Save to READMEs writes docs/<name>.svg and merges an ## Architecture section into the README beside it, idempotently, so re-running updates that section instead of stacking another copy. The design review your repo never had now lives in it.

The desktop app

One app. Every surface.

Chat with the coding agent, sketch architecture, audit dependencies, watch your Grafana and wire up connectors. Each tab drives the same agent core. Click through it.

↗ Open in ▾📁 lucid-train
workspacetalindrew×vion×+
Chats
new session
agent · lucid-train
21/7/2026, 1:26 PM
agent · lucid-train
19/7/2026, 3:56 AM
agent · lucid-train
16/7/2026, 12:26 AM
Commands ▸
Type / in the message box to search commands.
CODEpersonal_project/lucid-train
agent ▾⊙ morestart
The login test is flaky. Find the race and fix it.
Enter sends. Shift+Enter adds a line. ↑↓ history. Use /model, /plan…
Trial · 7d leftAArnab Bhattacharya ▲
↑ click the tabs · Code · Audit · Dashboards · Connectors · Diagram
Agent orchestration

One request. The right model for every job.

Lucid routes each turn through a small orchestrator. A fast model drafts the plan, a strong coder writes the diff, a vision model reads your diagrams, and task personas (dep-audit, grafana-ops, diagram) seed each tab with the right prompt and tools.

Per-role model routing. Assign a different model to Plan, Edit and Vision: cheap where it can be, frontier where it counts.
Concurrent projects. Every open workspace runs its own agent sidecar, so several repos work at once. ⌘K to jump.
Plan → approve → apply. Propose-only until you say go; then the same agent flips to auto and applies its plan.
“Audit my dependencies, then draw the architecture.”
LUCID ORCHESTRATOR
↓   ↓   ↓   ↓
Plan role
qwen3:4b · fast
Edit role
kimi-k2.5 ● running
Vision role
gpt-oss · reads diagrams
Persona
dep-audit ● running
3 projects3 sidecars · concurrentMCP tools attached
Built for real work

Everything you'd expect, and the parts you didn't.

🗂

Multi-project workspaces

Open several repos at once. Each gets its own agent sidecar and chat history, all running concurrently.

±

Reviewable diffs & approvals

Every edit lands as a green/red diff. Commands are risk-tiered: safe runs, caution & dangerous prompt, catastrophic is refused.

⌘K

Command palette

Jump between tabs and projects, toggle the theme, or open the workspace in your editor, all from ⌘K.

🔌

MCP & skills

Connect any Model Context Protocol server, and install SKILL.md packs from skills.sh without leaving the chat.

🎙

Voice & vision

Push-to-talk your prompts with on-device transcription, and attach images for vision-capable models.

Light, dark & your editor

One warm editorial identity in both schemes. Detects VS Code, Cursor and Zed, then opens the workspace in one click.

>_ The CLI

Prefer the terminal?
So do we.

The same Terminal-Bench-grade harness, as a single 6 MB Rust binary with a fast, keyboard-first TUI. Live plans, thinking stream, diffs, token accounting. No IDE, no browser.

$curl -fsSL https://lucidtrain.com/install.sh | sh
/agent full tools
/plan propose-only
/research cited, no shell
~/code/lucid-train · lucid-train
$ lucid-train
╭────────────────────────────────╮
>_ Lucid Train (v0.2.0) │
│ model: kimi-k2.5
│ provider: openrouter
│ directory: ~/code/lucid-train
╰────────────────────────────────╯
map repo · done
locate flaky test · done
thinking… the refresh isn't awaited
• Edited src/auth.rs
cargo test · 24 passed 1.8s
tokens: 12.4k in · 3.1k out
Models · open models first

Your models. Your machine.

Any OpenAI-compatible endpoint works. Run fully local and free, or bring your own key. Your prompts go straight to the provider you pick, never through us.

Local · free

Ollama, zero-setup

/download installs Ollama and pulls the best model for your RAM. 100% offline.

gpt-oss:20b
qwen3-coder:30b
deepseek-r1
OpenRouter · BYO key

15+ models

Bring your key. Stored locally, sent only to the provider.

GPT · Gemini · Grok
DeepSeek V4 · Qwen
GLM · MiniMax
Subscriptions · no key

Already paying?

login claude or login codex to use the plan you already have, no API key.

Claude Pro / Max
ChatGPT Plus / Pro
via the official CLIs
Benchmarks

The numbers, with their sources.

Every score below is a published third-party or vendor result, labeled with the benchmark it came from. No composite index of our own invention.

$lucid-train models --bench
DeepSeek V4 Pro
Lucid's default · cheapest near-frontier
LiveCodeBench
93.5%
#1 of any model
SWE-bench Verified
80.6%
GPQA Diamond
90.1%
reasoning
GPQA Diamond · graduate-level science90.1%
MMLU-Pro · expert knowledge87.5%
coding · agentic
LiveCodeBench · competitive programming93.5%
SWE-bench Verified · real bug fixes80.6%
Terminal-Bench 2.0 · agentic terminal use59.1%
SWE-bench Pro · hard real-world SWE55.4%

The value pick: tops LiveCodeBench, 1M-token context for whole-repo work, and output runs ~30× cheaper than GPT-5.5.

context 1Min $0.43/1Mout $0.87/1Msource: Artificial Analysis, OpenRouter & vendor reports, Jun 2026

Source: Artificial Analysis, OpenRouter & vendor reports, Jun 2026

Price vs the frontier

Near-frontier coding, 7×–34× cheaper.

The same everyday coding work, side by side, sorted by what you actually pay per million output tokens.

ModelSignal$ / 1M output
included in a plan your own API key
DeepSeek V4 ProCLOUD
Cloud · DeepSeek
80.6%
SWE-bench Verified
$0.87
GLM 5.2CLOUD
Cloud · Z.ai
68.8
AA Coding Index
$3.00
Kimi K2.7 CodeCLOUD
Cloud · Moonshot
> GPT-5.4
SWE-bench Pro
$3.50
Kimi K3CLOUD
Cloud · Moonshot · open-weight
agentic
repos · tools · 1M ctx
$15.0
Claude Opus 4.8
Anthropic · API key
88.6%
SWE-bench Verified
$25.0
GPT-5.5
OpenAI · API key
frontier
coding / reasoning
$30.0

Sources: Artificial Analysis, OpenRouter & vendor reports, Jun 2026. You pay your provider directly.

Pricing

The CLI is free. The desktop app is $3.

One licence for the desktop app. Everything runs on your machine, against local models or your own API keys, so there is no usage to meter and nothing to top up.

Terminal
CLI
Free · forever
  • The full agent, in one Rust binary
  • Local models, or your own API keys
  • Open source · no account needed
$curl -fsSL https://lucidtrain.com/install.sh | sh

or: pip install lucid-train

Desktop app
2 months free
Lucid Train
$2.50/mo
billed $30 a year
  • Everything in the CLI, in a native app
  • Design, Diagram, Audit and Dashboards
  • Local models, or your own API keys
  • 7-day free trial · no card up front
Start your 7-day trial

Every model call goes straight from your machine to whichever provider you configure. We never see your code, and there is no usage bill to read.

In the news

Why the terminal won.

Third-party writing about the models Lucid Train runs and the shift to agentic coding. Context, not endorsements. Every link goes to its original source.

Artificial Analysis@ArtificialAnlys
benchmarks
DeepSeek is back — V4 Pro tops the open-weight charts

The independent benchmark lab's read on the open-weight frontier: DeepSeek V4 Pro and Flash reset the price/intelligence curve.

read →
Anthropic
frontier
Introducing Claude Opus 4.8 — 88.6% on SWE-bench Verified

The closed frontier at $25/1M output. Lucid's open models land within striking distance for everyday coding, at a fraction of the price.

read →
Moonshot AI
agentic
Kimi K2.7 Code — open agentic intelligence

Coding-first, built for long tool-using runs and MCP. Moonshot reports it edging GPT-5.4 and Claude Opus 4.6 on SWE-bench Pro.

read →
Latent Space@swyx
dev tools
The terminal agent is eating the IDE

swyx & the AI engineering crowd on why coding moved back to the terminal, exactly the surface Lucid Train is built for.

read →
OpenRouter
rankings
GLM 5.2 — highest open-weight Artificial Analysis intelligence

Live token rankings and benchmarks for the open coding models. Lucid routes to GLM 5.2, DeepSeek V4 Pro and Kimi K2.7, no key needed.

read →
Artificial Analysis@ArtificialAnlys
benchmarks
DeepSeek is back — V4 Pro tops the open-weight charts

The independent benchmark lab's read on the open-weight frontier: DeepSeek V4 Pro and Flash reset the price/intelligence curve.

read →
Anthropic
frontier
Introducing Claude Opus 4.8 — 88.6% on SWE-bench Verified

The closed frontier at $25/1M output. Lucid's open models land within striking distance for everyday coding, at a fraction of the price.

read →
Moonshot AI
agentic
Kimi K2.7 Code — open agentic intelligence

Coding-first, built for long tool-using runs and MCP. Moonshot reports it edging GPT-5.4 and Claude Opus 4.6 on SWE-bench Pro.

read →
Latent Space@swyx
dev tools
The terminal agent is eating the IDE

swyx & the AI engineering crowd on why coding moved back to the terminal, exactly the surface Lucid Train is built for.

read →
OpenRouter
rankings
GLM 5.2 — highest open-weight Artificial Analysis intelligence

Live token rankings and benchmarks for the open coding models. Lucid routes to GLM 5.2, DeepSeek V4 Pro and Kimi K2.7, no key needed.

read →
CoreWeave
inference
Kimi K2 is the best open-source model — and the fastest to serve

Why open-weight coding models are now production-grade: frontier quality, served fast and cheap at scale.

read →
codersera
showdown
Kimi K2.7 vs DeepSeek V4 — the open coding showdown (2026)

How the top open coding models stack up, the same ones Lucid Train routes to.

read →
codingfleet
head-to-head
GLM 5.2 vs DeepSeek V4 Pro — SWE-bench leader vs the algorithm king

GLM 5.2 takes the open-weight SWE-bench crown; DeepSeek V4 Pro leads LiveCodeBench. Lucid serves both.

read →
Hacker News
show hn
Open coding agents are topping Terminal-Bench

Terminal-native agents are having a moment, with open harnesses on open models. Lucid Train is built for exactly this.

read →
CoreWeave
inference
Kimi K2 is the best open-source model — and the fastest to serve

Why open-weight coding models are now production-grade: frontier quality, served fast and cheap at scale.

read →
codersera
showdown
Kimi K2.7 vs DeepSeek V4 — the open coding showdown (2026)

How the top open coding models stack up, the same ones Lucid Train routes to.

read →
codingfleet
head-to-head
GLM 5.2 vs DeepSeek V4 Pro — SWE-bench leader vs the algorithm king

GLM 5.2 takes the open-weight SWE-bench crown; DeepSeek V4 Pro leads LiveCodeBench. Lucid serves both.

read →
Hacker News
show hn
Open coding agents are topping Terminal-Bench

Terminal-native agents are having a moment, with open harnesses on open models. Lucid Train is built for exactly this.

read →
Install

Get the desktop app, or the binary.

One native app for macOS and Windows, or one 6 MB Rust binary for your terminal.

Download for macOS
v0.1.33 · .dmg · Apple Silicon & Intel
Signed & notarized by Apple · opens without a Gatekeeper warning
sha256 698955a0187bc039f4c74f5d05a9f10fbb27376a45788a0a241d1326b73873c7
Download for Windows instead
CLI$curl -fsSL https://lucidtrain.com/install.sh | sh
FAQ

Questions, answered.

How much does it cost?

The CLI is free. The desktop app is $3/month, or $30/year, which works out at $2.50/month, so two months free. There is a 7-day trial. Model usage is separate and never billed by us: run local models at no cost, or bring your own API key and pay the provider directly.

Desktop app or CLI: what's the difference?

Same agent core. The desktop app adds Code, Diagram, Audit, Dashboards and Connectors in one window with a command palette and themes; the CLI is a single Rust binary with a keyboard-first TUI for the terminal. Use either, or both.

Can it run fully offline?

Yes, 100% offline with local Ollama models. Your code never leaves the machine, and there's no API key or telemetry. Credentials live in the OS keychain.

What is agent orchestration?

Lucid routes each turn through an orchestrator that can assign different models to different roles (plan, edit, vision) and seed each tab with a task persona. Multiple projects run their own agent sidecars concurrently.

Which models does it support?

Any OpenAI-compatible endpoint: local Ollama, OpenRouter (GPT, Gemini, Grok, DeepSeek, Qwen, GLM, Kimi…), or hosted cloud models. Context windows are tracked and the conversation compacts before overflow.