Hologram

What this is

Hologram watches a glTF asset pipeline and streams what's happening to a local dashboard in real time — including the tool calls your AI coding agent is making right now. It also exposes the pipeline to agents through a small MCP server, so Claude (or any MCP client) can enumerate, introspect, render, and health-check your exported .glb assets. The Claude Code plugin wraps all of it in guided skills (/hologram:start, inspect, check, status, create-skill) so you can drive the pipeline in plain language.

It stays deliberately read-only and non-destructive: it observes, introspects, validates, and previews your pipeline — but it never modifies your assets, and the MCP server imports none of your project code. No framework, no build step, no database — a stdlib HTTP server, a JSONL event log, and pure-Python glTF parsing.

Why I built it

I build games with my friend, and the bulk of my asset work runs through Blender into glTF — characters, props, weapons, the lot. At some point my AI coding agent became a real part of that pipeline: it writes the Blender scripts, runs the exports, and rearranges the .glb files I ship. That was a huge speed-up, right up until I realised I had no real idea what it was doing. Assets changed, exports appeared, and I'd be scrolling back through a terminal trying to reconstruct which step touched which file.

Hologram closes that gap. It tails a single event log and shows the agent's live activity — edits, shell commands, exports — right next to the assets those actions produce, in one local dashboard. Then it hands the agent that same pipeline back as a few read-only MCP tools, so we end up looking at the same picture instead of talking past each other. It started life as pipeline-specific glue buried inside one game's repo; this is the clean-room version — none of that project's code and none of its assumptions, just the pattern pulled out and made generic.

Features

• Live activity feed. A Server-Sent-Events dashboard that tails an append-only event log. Sessions, shell commands, file edits, MCP calls, and slash-command invocations stream in as they happen — failures and in-flight work included.
• Asset visualizer. Browse exported GLBs grouped by category, click any one to introspect its glTF structure, and see an in-browser preview of the model.
• Agent vision. render_asset renders a GLB to an image through your live Blender, so the agent can see an export — not just count its nodes. Non-destructive (a throwaway scene, your scene restored) and degrades to a clear error when Blender isn't running.
• Read-only checks + regression diffing. hologram check runs assertions you author in .hologram/checks.py over each asset (naming, root count, whatever you like) and fingerprints every asset, so the dashboard can answer “what changed since the last check.” Checks can't modify anything and never run inside the MCP server.
• Guided skills. The plugin bundles five Claude Code skills as a natural-language front door: /hologram:start, inspect, check, status, and create-skill.
• Pure-Python glTF introspection. Nodes, hierarchy, animations, materials, skins, and meshes, parsed with pygltflib. No Blender required.
• Generic by config. A single hologram.toml describes your paths and an optional category taxonomy. Flat projects need no categories at all.

The MCP surface

The server imports no project code — it only reads files and your config. An agent gets five tools — four read-only, plus a non-destructive render:

• list_assets — enumerate exported GLBs, grouped by category.
• inspect_asset — parse one GLB into the flat Asset struct (nodes, animations, materials, skins…).
• render_asset — render a GLB to a PNG via the live Blender. Non-destructive; returns a clear {error, hint} if Blender isn't reachable.
• tail_events — read recent pipeline activity from the shared event log.
• pipeline_status — “what's wrong right now”: recent failures, the last check summary, and recent asset diffs, from one read of the log.

How it works

No framework, no build step, no database — a stdlib ThreadingHTTPServer with SSE and vanilla JS, an append-only JSONL log the dashboard tails by watching byte size, and a bpy-free Asset API parsed with pygltflib. Blender is driven over a socket, never imported, so import purity holds.

your pipeline ──┐
Claude Code  ───┼──> .hologram/events.jsonl ──> dashboard (SSE live tail)
MCP server   ──┘                                    │
                                                    └─> /api/inspect ─> pygltflib
live Blender ──(socket :9876)── render_asset ─> PNG

How it's delivered

Four pieces, each install-free. The plugin bundles a stdlib-only activity hook (logging sessions, shell commands, edits, and MCP calls), the five MCP tools, and the five guided skills — wired in one step. The dashboard is the one piece you launch yourself, by design: a Claude Code plugin contributes hooks, commands, and MCP servers, not long-running web servers.

# the dashboard — point it at any project with exported GLBs
uvx --from hologram-gltf hologram dashboard

# the Claude Code plugin — live feed + MCP tools + skills
/plugin marketplace add akaieuan/Hologram
/plugin install hologram

uvx downloads the code (and a matching Python 3.10+) the first time and caches it, so there's no release to download and no environment to maintain. Prefer a classic install? pip install hologram-gltf (or, from a clone, pip install -e .) still gives you a plain hologram command.

Status & roadmap

Shipped so far: the live dashboard + MCP surface (v0.1), failures + GLB previews + hologram watch (v0.2), the read-only checks engine + asset visualizer (v0.3), agent vision + regression diffing + pipeline_status (v0.4), and guided skills as the plugin's front door (v0.5). The current release on PyPI is v0.6.0. Still on the table, built on the stable Asset API: an offline/headless render path (no running Blender), render thumbnails on disk, asset/scene templates, and export orchestration. MIT licensed.