Disk Is the Contract: Inside Threlmark’s Local-First Architecture

📊 Full opportunity report: Disk Is the Contract: Inside Threlmark’s Local-First Architecture on ThorstenMeyerAI.com — validation score, market gap, and execution plan.

TL;DR

Threlmark employs a unique local-first architecture where JSON files on disk are the primary data source, enabling open, portable, and restartable project management. This approach challenges traditional server-based models, emphasizing simplicity and control.

Threlmark has adopted a novel local-first architecture in which the entire project data ecosystem is built around JSON files stored on disk, with no server or cloud dependency. This design choice makes the filesystem the single source of truth, allowing for open, portable, and restartable project management that challenges traditional server-based tools.

The core architectural decision in Threlmark is that there is no dedicated server or database; instead, all data resides in structured JSON files within a designated directory (defaulting to ~/.threlmark). The system’s layout includes a manifest (threlmark.json), a dependency graph (links.json), individual project folders with metadata, and one file per roadmap card in the items/ directory. This structure ensures that every artifact is inspectable, portable, interoperable, and resilient against crashes or restarts.

Threlmark’s approach hinges on two key patterns: atomic file writes and read-merge-write updates. Files are written atomically via temporary files and renaming, preventing corruption during crashes. Updates read the current file, merge changes, and write back atomically, allowing multiple tools to modify data concurrently without conflicts. The system’s self-healing board reconciles actual files with lane orderings, maintaining consistency without locks or shared memory.

Disk is the contract: inside Threlmark’s architecture — ThorstenMeyerAI.com
ThorstenMeyerAI.com
Threlmark · Technical Deep-Dive
Threlmark · architecture

Disk is the contract: inside a local-first roadmap hub

A Next.js app on top of plain JSON files — no database, no cloud, no accounts. The key decision: the on-disk layout IS the API. Everything else cascades from taking that seriously.

Next.js · TypeScript · JSON-on-disk · MIT · part 2 of the Threlmark series
01The core decision

There is no server-of-record — the files are the record

The UI and any external tool reach the same files through the same discipline. The data root defaults to ~/.threlmark — home-based, because it’s a shared hub every one of your apps points at.

~/.threlmark/ ├─ threlmark.json # manifest ├─ links.json # dependency graph ├─ projects// │ ├─ project.json # meta + wipLimits │ ├─ board.json # lane ordering │ ├─ items/.json # ONE card per file ← source of truth │ ├─ suggestions/ # the Inbox (drop-zone) │ ├─ handoffs/ # recorded agent handoffs │ ├─ reports/ # agent report drop-zone │ └─ ROADMAP.md # human-readable mirror ├─ shared/items/ # cards many projects ref └─ archive/ # archived, still readable

Inspectable

Every artifact is a file you can cat, diff, grep, commit.

Portable · no lock-in

Back up with cp, sync with Dropbox / git, migrate trivially.

Interoperable

Any tool in any language joins by reading / writing files.

Restartable

No in-memory state to lose — stateless over the files.

02Making files safe
Amazon

JSON file editor for project management

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Two disciplined patterns instead of a database

“Just use files” is easy to get wrong. These two patterns — ported from a battle-tested sibling app — are what make file-based state sound rather than reckless.

Pattern 1

Atomic writes

Write to a temp file in the same dir, then rename() over the target. Rename is atomic on one filesystem — a crash mid-write leaves the complete old file or the complete new one, never a half.

write .tmp-pid-rand fsync rename() over target
Pattern 2 · one file per item

The board heals itself

A single roadmap.json array races when two tools write at once. One file per card makes writes collision-free. Lane order lives in board.json and reconciles on read.

The payoff: an external tool never touches board.json. It writes an item file — the board fixes itself on Threlmark’s next read. Unknown keys are preserved, so the contract is forward-compatible.
03Derived, never stored
Real-World Android App Projects with Kotlin and Jetpack Compose: Build Production-Style Android Apps with Modern Architecture, API Integration, State Management, Local Data Storage, Practical Projects

Real-World Android App Projects with Kotlin and Jetpack Compose: Build Production-Style Android Apps with Modern Architecture, API Integration, State Management, Local Data Storage, Practical Projects

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

The numbers can’t drift from the files

Anything computable from item state is computed — so the displayed numbers can never disagree with the underlying JSON. Priority is the clearest example: it’s calculated on read, never persisted.

priority — computed on read

Impact weighted heaviest; effort the only axis that subtracts. Reused verbatim from the original tool, so imported cards rank identically.

priority = max(0, round(impact·3 + evidence·2 + fit·2effort·1.5))
a 5 / 5 / 5 / 4 card 29
work-item age
now − lane-entry time. Past threshold (dev 7d, ranked 21d, idea 60d) → stale.
cycle time
first DevelopmentDone. Derived from append-only transitions[].
throughput
items reaching Done per ISO week, 8-week window.
WIP
count per lane; over the cap shows 3 / 2 in red.
04The closed agent loop · press play
Amazon

disk-based JSON data storage tools

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

A handoff is a first-class flow event

The genuinely 2026-shaped part: most building is done by AI agents, so Threlmark closes the loop. Watch a card go from ranked to Done without anyone dragging it.

Handoff → report → self-move

The brief carries a reporting protocol. The agent reports through REST or the filesystem — and a done report moves the card itself.

Ranked
Add price-drop alertsscore 31 · ready
Development
Handed off 🤖
Done
▶ preferred — REST
POST /api/projects/:id/
items/:itemId/report

Direct call. Applied immediately.

▶ fallback — filesystem
drop reports/.json
→ ingested on read

Robust even if the server’s down at finish time.

🤖 claude done: price-drop alerts shipped · typecheck + lint + build passed — card moved to Done
05Portfolio score & deployment
Heavy Duty Job File Folders for Project Tracking Donkey Auto Products (100 Count) (Blue)

Heavy Duty Job File Folders for Project Tracking Donkey Auto Products (100 Count) (Blue)

HEAVY DUTY CONSTRUCTION: Made from robust 100 lb paper stock that withstands frequent handling, active job files, and…

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

A small formula, and an honest hosting caveat

Because items are globally addressable (/), the Portfolio ranks everything together by a status-weighted score — finishing beats starting, blockers get a boost.

Portfolio ranking — status-weighted

In-flight work floats to the top; bottlenecks cost the most, so blockers get nudged up.

score = priority · statusWeight (+ 0.1 · blockedCount · priority)
1.3
development
1.0
ranked
0.85
idea
0.15
done
Path 1

Static read-only demo

Seeded data, writes to localStorage. Try-before-you-clone.

Path 2

Personal Node instance

Password-gated, persistent backed-up THRELMARK_DATA_DIR.

Path 3

Multi-tenant SaaS

Add accounts + per-tenant isolation. A separate build.

The elegant part: the store interface src/lib/*/store.ts is the natural seam — the same boundary that keeps the local tool simple is the one you’d extend for multi-tenancy. The architecture doesn’t fight that future; it just doesn’t pay for it until you need it.
ThorstenMeyerAI.com
Threlmark · open source (MIT) · github.com/MeyerThorsten/threlmark · part 2 of a series · file layout, formula, weights & agent-loop channels are Threlmark’s actual mechanics.

Why Disk-Based Data Matters for Project Management

This architecture offers significant advantages in transparency, portability, and safety. Since data is stored in plain JSON files, users can easily back up, migrate, or integrate with other tools without vendor lock-in. The stateless nature of the process means data recovery is straightforward after crashes, and external tools can participate seamlessly by reading and writing files. This approach empowers users with full control over their project data, reducing reliance on centralized servers and proprietary databases.

The Evolution of Local-First and File-Based Systems

Traditional project management tools often rely on centralized servers or cloud services, which can introduce lock-in, data opacity, and dependency on network connectivity. Threlmark’s design echoes broader trends toward local-first, peer-to-peer, and open data systems, emphasizing user sovereignty and resilience. Its focus on JSON files as the contract aligns with practices in version control and configuration management, but applies them to project workflows, enabling a new level of interoperability and safety.

“The on-disk layout is the API. Everything cascades from that decision, making the system open, portable, and restartable.”

— Thorsten Meyer, creator of Threlmark

Outstanding Questions About Threlmark’s Scalability

It is not yet clear how Threlmark’s architecture performs at scale with very large project sets or in multi-user environments. The current focus is on single-user workflows and local-first principles, but the system’s behavior under concurrent multi-user editing or complex dependency graphs remains to be tested and documented.

Future Developments and Community Adoption

Threlmark’s developers plan to expand features around multi-user collaboration and integrate more external tools that leverage the file-based contract. Community feedback and real-world testing will shape the evolution of the system, potentially leading to broader adoption of local-first project management principles.

Key Questions

How does Threlmark handle concurrent updates?

It uses atomic file writes and read-merge-write patterns to prevent conflicts, enabling safe concurrent modifications without locks.

Can external tools participate in managing projects?

Yes, any tool that can read and write JSON files in the specified structure can participate, making the system highly interoperable.

What are the benefits of storing data on disk instead of in a cloud or database?

It provides full control, easy portability, resilience against crashes, and eliminates vendor lock-in, fostering open and transparent workflows.

Is Threlmark suitable for team collaboration?

Currently, it is optimized for single-user workflows, but future features may support multi-user collaboration with appropriate synchronization mechanisms.

How does this approach compare with traditional project management tools?

Unlike cloud-based or server-dependent tools, Threlmark’s file-based system offers simplicity, safety, and interoperability, but may require different workflows for team-based projects.

Source: ThorstenMeyerAI.com

You May Also Like
When One Agent Isn’t Enough: Claude Now Builds Its Own Team Of Agents On The Fly

When One Agent Isn’t Enough: Claude Now Builds Its Own Team Of Agents On The Fly

Claude now autonomously builds and manages its own team of agents on the fly, enhancing performance on complex tasks. This breakthrough is a significant step in AI orchestration.
Fable 5 Is Back. GPT-5.6 Is Next. And Anthropic Reportedly Already Has Something Stronger.

Fable 5 Is Back. GPT-5.6 Is Next. And Anthropic Reportedly Already Has Something Stronger.

Fable 5 is back after an 18-day blackout, GPT-5.6 is in preview awaiting approval, and rumors suggest Anthropic already has a more capable model.
Mac vs GPU Tower for Local LLMs: The Heat-and-Noise Tradeoff

Mac vs GPU Tower for Local LLMs: The Heat-and-Noise Tradeoff

Analyzing the heat and noise differences between Mac Silicon machines and GPU towers for local large language model inference, highlighting key tradeoffs.
Capital: The Lever Beneath the Levers

Capital: The Lever Beneath the Levers

Exploring how capital funding shapes AI development, with recent IPOs and circular investments revealing vulnerabilities in the industry’s financial structure.