laniameda.gallery

A personal AI creatorship vault for production media and external references, organized so people and agents can find, reuse and preserve source provenance.

Objects and agent access

• Media assets: images/videos with sources, descriptions, tags and linked prompts.
• Skills: reusable techniques/recipes with text and optional prompt/media steps.
• Worlds: story universes represented by root collections; collections group media and Skills and support one child level. Exact rendering styles are tags.
• Stories/Scripts: private text ideas, scripts and style locks with revision history.
• X bookmarks: original post text and linked saved media.
• YouTube references: separate video/channel research with copied stills and themes. Motion/Cinematography are tagged asset/Skill views.

Start agents at /llms.txt, the canonical skills/laniameda-gallery/SKILL.md and authenticated Gallery MCP. Without a connector use the dependency-free Bun client. The deployed manifest reports version/fingerprints. Owner-only world/maintenance policy is private agentInstructions data; public copies are fetch bootstraps. Do not publish private IDs or credentials. Legacy pillar/workflow/folder/pack names are compatibility details, not additional user-facing organization.

Stack

• Next.js (App Router) + TypeScript
• Convex — realtime DB, file storage, ingest actions, vector search
• Cloudflare R2 — image/video object storage for new uploads
• Telegram auth — login + ingest delivery channel
• Bun — package manager and runtime

Self-hosting

You'll need:

• A Convex deployment (convex.dev)
• A Telegram bot (@BotFather) for login
• (Optional) A second Telegram bot for "Saved" notifications after ingest
• A Cloudflare R2 bucket for media uploads (required for the extension drop desk)
• (Optional) Gemini API key for semantic search embeddings

Setup

bun install
cp .env.example .env.local

Fill in .env.local. Minimum to boot:

NEXT_PUBLIC_CONVEX_URL=https://<your-deployment>.convex.cloud
CONVEX_URL=https://<your-deployment>.convex.cloud
SESSION_SECRET=<at least 32 chars>

For Telegram login:

TELEGRAM_LOGIN_BOT_TOKEN=...
NEXT_PUBLIC_TELEGRAM_BOT_USERNAME=<bot username without @>

To restrict admin actions (delete, curate-to-public) to yourself, set:

KB_OWNER_USER_ID=<your numeric Telegram ID>
CURATION_ADMIN_USER_IDS=<your numeric Telegram ID>,telegram:<your numeric Telegram ID>
NEXT_PUBLIC_CURATION_ADMIN_USER_IDS=<your numeric Telegram ID>,telegram:<your numeric Telegram ID>
CURATION_ADMIN_SECRET=<any long random string>

Run Convex dev (in one terminal):

bun run convex:dev

Run the app (in another):

bun run dev

App listens on http://localhost:3317 by default.

Browser extension

Set EXTENSION_API_TOKEN and EXTENSION_OWNER_USER_ID in the app environment, then load extension/ with Load unpacked from chrome://extensions (Chrome 114 or newer). Click the toolbar icon to open the persistent upload desk, enter the gallery origin and the same extension token under Workspace, and save the settings. You can then choose a collection and drop images, videos, or an entire folder into the panel; supported media inside folders is queued recursively before upload.

The extension is loaded unpacked from this repo's extension/ folder. After a pull, open chrome://extensions and press the reload arrow on Save to Gallery; there is no build step.

On midjourney.com the extension adds three things for Dear Annete work:

• Moodboard dock. Under the Imagine prompt box, one chip per gallery moodboard of the chosen world (GET /api/extension/moodboards?world=…). A click sets that board's --profile <code> in the prompt and keeps the taste profiles; clicking the active chip clears it. Hovering a chip shows six full-size pictures of the board. Codes are read from the gallery on every refresh (45 s, tab focus, the ↻ button) and never cached, so a board whose code changed after a Midjourney sync shows the new code at once. Only the world choice is stored.
• Hotkeys on a selected Midjourney image (Mac, Ctrl elsewhere): ⌘⇧D adds it to the Dear Annete world, ⌘⇧L as a location, ⌘⇧P as a scene, ⌘⇧M adds it and opens a comment box. The targets live in extension/save-targets.js; comments post to /api/extension/comments.
• Prompt packs. A bar of short phrases from the world's lexicon (style openers and the 35mm film block, architecture by city level, exterior and interior rooms, light, camera, wardrobe by class, world markers, the fantasy elements). A click inserts the phrase at the cursor, or before or after your own words (the mode switch on the bar); a second click removes it. Flags such as --profile and --ar are never touched. The phrases are in extension/prompt-packs-data.js and carry no real place, river, artist or brand names; the gallery world (open_world dear-annete, docs/midjourney.md) is their source.

See agent-docs/ENV_MATRIX.md for the complete env reference and agent-docs/AUTH.md for the full Telegram login setup.

Permissions model

The gallery has two user tiers:

• Anyone who logs in can save assets to their own scope, browse the public gallery, and view their own saves.
• Configured admins (listed in CURATION_ADMIN_USER_IDS) can additionally delete assets and mark assets as public/featured. Non-admins see no delete affordance and the server rejects delete requests they shouldn't be making.

This is enforced both server-side (Convex mutations require CURATION_ADMIN_SECRET

• admin-allowlist match) and in the UI (delete controls only render for admins).

Key commands

bun run dev          # Start Next.js
bun run lint         # Lint
bun test             # Run the complete backend, UI-helper, and extension suite
bun run typecheck    # Type check
bun run convex:dev   # Canonical cloud backend; serves real gallery data
bunx convex codegen  # Regenerate Convex types after schema changes

Convex schema

See convex/schema.ts for the data model. Key tables:

• prompts — prompt text, type, domain, tags, owner scope
• assets, assetFolders, assetTags — media with source/prompt and multiple memberships
• designInspirations — design references with platform/workflow type metadata
• workflows, workflowFolders — internal storage for reusable Skills
• assetPacks — internal media grouping, distinct from collections/Skills
• stories, storyRevisions, galleryPresets — private text and curated views
• bookmarks, videoRefs — X post text and separate YouTube research
• agentInstructions — private owner skill resources with version/hash guards
• tags — tag system with categories (model_name, style, content_type, etc.)
• folders — optional user-defined folder organization
• semanticDocuments — vector index for semantic search
• users, runs, generationLineage, ingest_failures — supporting tables

Project structure

app/             Next.js App Router pages + API routes
components/      React components (v8/ is the current dashboard)
convex/          Backend: schema, queries, mutations, actions
lib/             Shared helpers (auth, identity, gallery filters, etc.)
skills/          Repo-local agent skills (ingest + query)
agent-docs/      Architecture docs, env matrix, auth setup, observations
extension/       Chrome extension entry for one-click design saves
tests/           Bun tests for backend + API routes

Documentation

The agent-docs/ directory has the full set:

• PROGRESS.md — what's been built and recent changes
• OBSERVATIONS.md — known quirks and lessons learned
• BACKEND_CONVEX_SETUP.md — Convex setup walkthrough
• AUTH.md — Telegram auth setup end to end
• ENV_MATRIX.md — every env var, where it's read, and what for
• SEMANTIC_SEARCH.md — Gemini embedding + vector index setup
• DESIGN.md — UI design system
• MCP_AGENT_ACCESS.md — agent access and schema discovery

License

MIT. See LICENSE.