Artifice
Machine

Building tools, protocols,
and systems for thinking with machines.

Section 01 Current Work

What is being
made now

Small by design. Each project here is active for a reason.

Every tool here is built and tested on local models: Ollama, vLLM. All running on-premise.

0242026Hot

Pencil Sync

Tool · design-sync · 2026–

Bidirectional sync between pencil.dev designs and frontend code. Edit a component in the canvas, the code updates. Change the code, the design follows.

  • design-sync
  • tooling
  • claude-code
0232026Hot

Vidistiller

Tool · extraction · 2026–

Self-hosted video intelligence platform. Transcribes, detects slides, summarizes with LLMs, and exports to Obsidian-ready Markdown.

  • extraction
  • video
  • transcription
0272026Hot

Token Diet

Tool · optimization · 2026–

Unified installer and dashboard for RTK, tilth, Serena, and ICM. Slashes AI session token costs by 40–90% without changing how you work.

  • tokens
  • optimization
  • tooling
  • mcp
0252026Hot

Superharness

Protocol · multi-agent · 2026–

Multi-agent task coordination for Claude Code, Codex CLI, Gemini CLI, and OpenCode. A shared contract layer that keeps AI coding assistants from stepping on each other.

  • multi-agent
  • coordination
  • agents
0262025Hot

Obsidian Semantic MCP

Tool · memory · 2025–

Turns an Obsidian vault into a queryable brain for Claude. Semantic search across every note using local embeddings, no cloud services required.

  • mcp
  • memory
  • obsidian
  • embeddings
0282026In Forge

Tessera MCP

Tool · memory · 2026–

Local MCP server for semantic search over non-markdown Obsidian attachments. Covers PDFs, images, audio, CSV, and code: everything markdown cannot see. Sibling to Obsidian Semantic MCP.

  • mcp
  • obsidian
  • embeddings
  • rust
0222026In Forge

Morpheme

Tool · observability · 2026–

A live, in-browser dashboard for AI agent session state. One dashboard, multiple data sources. Connects to Superharness natively.

  • observability
  • agents
  • monitoring
0212026In Forge

Synod

Protocol · multi-agent · 2026–

A coordination protocol for clusters of AI agents that deliberate, decide, and learn together.

  • multi-agent
  • protocol
  • deliberation
Section 02 Writing

What is being
learned

Essays, field notes, failures, and the occasional longer argument.

Written from direct practice. Local models, local memory, local infrastructure. No cloud-first assumptions.

May 2026Why Local-First MattersWhy local is the source of truth and the network is a deliberate export, not a continuous mirror.local-first3minApr 2026Why the Workshop Stays SmallOn focus, scope, and the deliberate refusal to scale before the work is understood.principles2minMar 2026On Building the HarnessNotes on why the infrastructure around AI agents matters as much as the models themselves.ai-agents2min

Operator note Artifice Machine

Maxime Roy builds Artifice Machine. The work is published here.

One developer. A suite of AI instruments. Not a product studio.

Artifice MachineRead more
Section 03 Method

How the work
is approached

Operating principles

These principles were not designed. They are not rules. They came from the failures.

  • Build the harness, not just the model
  • Sign the work
  • Small enough to understand, useful enough to ship
  • Record the failures. They are part of the work.
  • Tools should outlast the moment that produced them
  • The interface is part of the instrument
  • Run local. Models, memory, and tools stay on the machine
Full archive: findings & notesProjects 8  •  Writing 3