[A] Atlas Studio

Adaptive prompt system for structured thinking and decision-making.

Status: v1.5.0 — public repo hygiene and prompt-pack maturity

👉 Live demo: https://mcamner.github.io/atlas-one/

---

Atlas Studio is a local-first adaptive prompt system that turns raw intent into structured thinking workflows.

Instead of writing prompts ad-hoc, you:

• define intent
• select a reasoning mode
• generate a structured prompt
• copy or hand off cleanly to ChatGPT

---

Visual flow

Input
  ↓
Route selection (auto-classify mode)
  ↓
Workflow (analysis / architecture / strategy)
  ↓
Structured prompt pipeline
  ↓
ChatGPT execution   OR   mq-agent execution (plan / review / audit / signal)

---

UI overview

The interface runs entirely in the browser — no install required.

┌─────────────────────────────────────────────────────────┐
│ [A] Atlas Studio                              v1.5.0    │
│ ─────────────────────────────────────────────────────── │
│  SIDEBAR              │  MAIN PANEL                     │
│  ─────────────────    │  ─────────────────────────────  │
│  Type: Architecture   │  Goal input                     │
│  Search modes…        │  > Design a secure remote       │
│                       │    access architecture          │
│  Library              │                                 │
│  > Atlas One Core     │  Prompt type + reasoning route  │
│                       │  Requirements → Options → Rec.  │
│  Modes (13)           │                                 │
│  > Analyze            │  Final prompt output            │
│  > Architect          │  ChatGPT-ready structured text  │
│  > Debug              │                                 │
│  > …                  │  [ Generate ] [ Copy ] [ → ]    │
│                       │                                 │
│  Packs                │                                 │
│  > mq-mcp Safety      │                                 │
│  > Release Readiness  │                                 │
│  > Systems Thinking   │                                 │
│  > …                  │                                 │
└─────────────────────────────────────────────────────────┘

Keyboard shortcuts: Ctrl+Enter generate · / focus command · s focus search · Esc clear

---

Screenshot

Interface preview from the GitHub Pages build:

Try it

👉 https://mcamner.github.io/atlas-one/

No install. Runs in your browser — no Java required.

---

Why it matters

Most AI usage today is unstructured, inconsistent and hard to repeat. Atlas Studio adds structure to thinking, consistency to execution and repeatability to workflows — a system for thinking and execution, not another prompt box.

---

What it does

Atlas Studio turns intent into structured AI workflows:

• Adaptive prompt modes — switch between analysis, architecture, research, strategy, decision, problem solving, execution and writing
• Quick action controls — pick /atlas, /research, /write, /strategy and other modes without rewriting the prompt
• System reasoning preview — see the selected problem type, route, pipeline and rationale before you generate
• Structured prompt library — reusable prompts loaded from web/prompts.json, with source prompts kept in docs/prompts/
• Local-first execution — runs entirely on 127.0.0.1, no external dependencies
• ChatGPT handoff — copy the final prompt and open ChatGPT in one action
• mq-agent execution — run the routed goal directly via mq-agent (plan, review, audit, signal) when the Java backend is running; results appear inline, the panel takes a repo path, and unmapped modes show as mode → plan (fallback)

---

Example flow

User input:
"Design a secure remote access architecture"

↓

Prompt type selected:
Architecture

Route:
Requirements and Constraints → Design Options → Review and Recommendation

↓

Generated output:
ChatGPT-ready structured prompt

---

Architecture

Atlas One is the prompt and interaction layer in the local mq ecosystem. It packages reasoning modes, prompt packs and handoff patterns. It does not own runtime validation or review decisions.

Current role split:

flowchart TD
    U([User goal]) --> AO

    subgraph AO["Atlas One — think / prompt / route"]
      R[Mode routing + prompt packs] --> HO[Handoff text]
    end

    HO -->|copy| EXT[ChatGPT / external LLM<br/>execute externally]
    HO -->|run| AG[mq-agent<br/>execute / orchestrate]

    AG --> MC[mq-mcp<br/>review / risk / validation / memory contracts]
    AG -. optional provider .-> OL[Ollama<br/>local model, never a decision authority]
    MC --> OB[(mqobsidian<br/>persist / source of truth)]
    R -. Save to brain .-> OB

Loading

• Atlas One — prompts, mode routing, interaction patterns and handoff text
• mq-agent — orchestration and command execution flow
• mq-mcp — review, risk, validation, safety classes and memory contracts
• mqobsidian — durable memory: persists decisions and curated context (Save to brain)
• Ollama — optional local model provider, never a decision authority

atlas-one/
├── docs/prompts/     # Core prompt, modes, and seed prompt packs
├── docs/             # GitHub Pages and prompt library docs
├── src/              # Java server (API + static hosting)
├── web/              # UI (routing, prompts, visualization)
├── archive/legacy/   # Retired prototype source (not built)
├── build_and_run.sh  # Local dev runner
├── package_mac_app.sh
└── run_mac.command

Backend

• Java HTTP server on 127.0.0.1:8766
• Serves UI + API endpoints:
  • /api/prompts
  • /api/health
  • /api/execute — routes goal → mq-agent CLI command
  • /api/core/runs — runs a task through Atlas Core (see below)

Frontend

• Vanilla JS application
• Handles:
  • routing logic
  • prompt generation
  • UI state
  • local storage

Atlas Core run

The Atlas Core run panel sends the goal to Atlas Core and shows what Core reports. Atlas One does not grade anything: every value in the panel comes from the Core CLI, and a value Core did not report is shown as not reported.

Endpoint	Core command
POST /api/core/runs {task, repo_path}	atlas create, then atlas run --run-id ID --event-log LOG --json [--repo-path P]
GET /api/core/runs/{id}/status	atlas status ID --json (atlas-status.v1)
GET /api/core/runs/{id}/events	atlas events ID (atlas-event.v1)
GET /api/core/runs/{id}/inspect	atlas inspect ID --json (atlas-inspect.v1)
GET /api/core/runs/{id}/result	the run's stdout (atlas-run.v1), exit code and stderr
POST /api/core/runs/{id}/cancel	atlas cancel ID --json

POST requires Content-Type: application/json (otherwise 415) and, when the request carries an Origin, it must be Atlas One's own localhost origin (otherwise 403): a page on another origin cannot start or cancel a run. The body is decoded as strict JSON and the task is passed to Core exactly as typed.

Tests, after javac -d out src/*.java tests/*.java:

Command	Checks	CI
java -cp out CoreRunHandlerTest	the 415/403 refusals and the strict JSON decoding, against a fake atlas	yes
java -cp out CoreContractTest	every endpoint relays Core's document and exit code unchanged, including a run that did not pass and a run with no document, against a mock Core that prints tests/fixtures/core	yes
java -cp out CoreSmoke	a run on a one-file repository through a real Core: each answer equals what the atlas CLI prints, Core read README.md with its real SHA-256, and the schema tags match the fixtures	no, needs Core (ATLAS_BIN or atlas on PATH)

The fixtures were printed by Atlas Core 1.0.0 (7142524) for a run on a repository holding one README.md. When Core's output changes, capture them again with atlas run … --json, atlas status, atlas inspect and atlas events and rerun CoreSmoke.

One run is one event log: ~/.atlas-one/runs/<run_id>.jsonl (ATLAS_ONE_RUNS_DIR overrides the directory). atlas must be on PATH, or set ATLAS_BIN.

The panel shows the run id, state, status, stop reason, iterations, budget usage, requires_user_approval, uncertainties, each source's path and SHA-256 from inspect, and the last evaluation's citation checks. The route preview and editor are a prompt preview: nothing there has been executed.

---

Prompt library

Atlas One v0.2.0 adds a file-based prompt library:

docs/prompts/
├── atlas-one.md
├── modes/
└── packs/

Core docs:

• Prompt Library
• Modes
• Routing
• Examples

Canonical modes:

Analyze
Decide
Architect
Research
Explain
Teach
Edit
Coach
Plan
Debug
Review
Summarize
Create

---

Run locally (Java backend)

Requires Java 17+. Runs the full server version at http://127.0.0.1:8766.

./build_and_run.sh

Or as a macOS app bundle:

./package_mac_app.sh

---

Design principles

• Local-first — no cloud dependency
• Structured thinking over raw prompting
• Repeatability over improvisation
• Separation of intent and execution
• Composable workflows

---

Roadmap

Current: v1.5.0 shipped. Next: v2.0.0 — Atlas as a full local prompt/workflow studio (direction; not yet scoped). See ROADMAP.md for the full release map.

---

Use cases

• Architecture design
• Technical decision-making
• Problem analysis
• Structured research
• Workflow standardization
• Prompt engineering at scale

---

Status

Stable prompt operating system with mq ecosystem prompt packs and a documented execution boundary. Current public alignment target: v1.5.0.

---

Security

Atlas Studio runs locally. No data is sent anywhere unless you configure an external AI provider.

Do not commit API keys or credentials. Use environment variables or ignored local files for sensitive values.

---

License

MIT

---

Author

Mattias Camner IT Architect — building practical systems where infrastructure, automation, and usability work together

---

Final note

Atlas Studio brings structure, repeatability and system thinking to how we use AI.

macOS helper scripts

The repository supports helper scripts for macOS packaging and running. To keep these portable:

• Prefer setting MQ_ROOT environment variable to your repo root:

  export MQ_ROOT="$(pwd)"

  Or rely on the scripts' autoprobe (they try MQ_ROOT, git rev-parse, then a script-relative fallback).

• Create Claude local settings from the example:

  cp .claude/settings.local.example.json .claude/settings.local.json
  # Edit .claude/settings.local.json and replace <MQ_ROOT> with your path (or set MQ_ROOT env var)

• Packaging requires jpackage (JDK 16+ or a JDK distribution with jpackage). If jpackage is missing, package_mac_app.sh will fail with a clear message.