TITLES MCP use case: The Cabinet of Minor Gods

Summary

The Cabinet of Minor Gods is a sixty-work synthetic folk-religion project developed through a hybrid TITLES MCP + local-production workflow. TITLES supplied model discovery, artist-model imagery, image-to-video transformations, multi-provider speech, execution state, outputs, and provenance. Local tools turned those outputs into a structured collection, two interactive websites, ten films, a complete evidence film, deterministic sound libraries, and a deployable static package.

The production demonstrates an important distinction:

Generating a media asset is one operation. Authoring a world in which sixty assets relate, sound, move, accumulate, and become navigable is another.

TITLES project record

Session ID: f2739b27-bd62-4a36-ad50-cccb6569c290
Session name: The Cabinet of Minor Gods — Launch Pantheon
Studio: https://www.titles.xyz/create/f2739b27-bd62-4a36-ad50-cccb6569c290
Publication state: not published

The final session canvas contains:

Node/output counts and model selectors below were confirmed against the live TITLES session during packaging, supplementing the repository's local provenance records.

Operator	Nodes	Successful outputs
txt2ImgNode	62	124
img2VidNode	62	60
txt2SpeechNode	119	119
Total	243	303

The two image-to-video nodes without outputs were moderated Lost Glove attempts. They were automatically refunded. A third, more restrained ambient-material prompt succeeded.

Model usage

Image generation

Model	Creator/type	Model ID	Adapter ID	Nodes	Outputs	Role
Cut From Elsewhere	@juujuumama / artist model	ffbe887a-4f38-415b-b45d-2ba4e51ad919	1a782b61-3561-411e-8a0d-e2d4170d826e	60	120	Final two-candidate portrait generation for all sixty gods
Collage Around Me	@sudutalien / artist model	7b5efa71-29a1-4976-8e73-f51784027af3	ab441a44-cf30-4105-bf4d-1ecc2e9f2dfd	1	2	Visual-tradition test
Bones and All	@antidote / artist model	88a9631b-f355-4631-a115-c892b5e31b82	b1f5b030-5335-482c-b069-6e3cb29d52ed	1	2	Visual-tradition test

The final tradition was Cut From Elsewhere, trained on Sierra Juujuumama's original physical collages. Its tactile, layered, sacred-object language supported the project's worn municipal relics without requiring the artist model to reproduce one literal subject category.

Each final portrait was selected from two generated candidates. Selection remained an editorial decision rather than an automatic first-output acceptance.

Recorded Saint Draft style-test executions:

• Cut From Elsewhere — 77761392-1d54-491a-a6a4-3a6e4885c76e
• Bones and All — b1789836-6fae-456d-858e-991ded0ab8af
• Collage Around Me — 2d3f84ac-a158-4912-92a0-a974dc53dd3c

Image-to-video generation

Model	Model ID	Adapter ID	Nodes	Outputs	Role
Kling Video v2.6 Pro	249177de-df51-498e-bf9a-abab2ee6a6cf	1a102e60-0263-4b0d-88e0-0aaef4789d6d	62	60	One canonical five-second manifestation for each god

The prompt grammar asked for one object-specific event rather than default camera movement. Examples included mildew behaving like weather, a drawer revealing an obstruction, mugs exchanging handles, muted words becoming beads, and a receipt becoming a ferry.

The final coverage invariant was:

60 gods → 60 canonical portraits → 60 canonical manifestations → 0 decode failures

Speech generation

The complete session includes auditions, early liturgy evidence, and the final feature-film cast.

Model	Model ID	Adapter ID	Session executions
Gemini 3.1 Flash TTS	9abdcf50-5e37-4f5c-bc7f-d9d54babcdd6	0058c2d2-274b-4ed9-a20e-7f290e7b476c	44
ElevenLabs Turbo v2.5	546bf826-ed1e-4901-b1a0-ae7403098401	dde7df58-050e-4e65-b664-4d26db4c6fbc	39
MiniMax Speech 2.8 Turbo	6b25d7b9-ca34-4caa-8107-a457716863aa	70c04ce5-771f-4b4b-bd34-5e377d9f4026	19
MiniMax Speech 2.8 HD	afa71476-5d33-424d-893e-432aec6c8f8c	5f2eac0c-686b-4809-96dc-b1d41372c9e4	11
Chatterbox HD	759b9328-7658-459f-a22d-15a932e0b217	68ec61e3-9d30-4a1b-93a3-59092d9fa871	6
Total			119

Final recurring cast:

Evidence channel	Model	Preset voice(s)	Final allocation
Private answering machine	ElevenLabs Turbo v2.5	River	28 units
Municipal intercom	MiniMax Speech 2.8 Turbo	Patient_Man	19 units
Conflicting witnesses	Gemini 3.1 Flash TTS	Charon + Aoede	13 units / 26 performances

The final evidence film therefore used 60 voice units and 73 raw performances.

Spend

Media category	Actual TITLES spend
Images and artist-model tests	$3.187894
Sixty successful manifestations	$27.300000
Pre-evidence prayers, production speech, and auditions	$0.830098
Final evidence-film speech	$0.729782
Complete project	$32.047774

Additional cost decisions:

• Two moderated Lost Glove video attempts were refunded and excluded.
• Refunded execution IDs: e2806e2e-dfaf-428e-8431-cab21a6e5e72 and f5137c21-ea27-4c4d-8f7a-bae553fb6ece.
• Successful restrained Lost Glove output: edc37289-1057-4eb6-b974-ce2265c37629.
• A 60-second TITLES music generation was quoted at $1.04 and intentionally not run.
• Ten chapter scores and sixty audio relics were instead created procedurally and locally for $0.00 generation cost.
• The final sixty-god evidence-film speech pass cost $0.729782.
• The final speech pass was projected at $0.562230; the actual cast/performance mix cost $0.729782.
• Nothing was published publicly by the agent.

MCP operations used

The workflow used TITLES MCP capabilities for:

1. Signed-in account and output inspection
2. Public feed and artist-practice reconnaissance
3. Session creation and persistent session retrieval
4. Model search and model-detail inspection
5. Operator-schema and input-constraint discovery
6. Text-to-image generation with explicit model/adapter selection
7. Image-to-video animation from selected portrait outputs
8. Text-to-speech generation across several providers
9. Generic execution runs where exact provider/model/voice pinning was required
10. Execution waiting and terminal-state checks
11. Output retrieval and local asset download
12. Session-wide provenance inspection
13. Moderation/refund recovery

No TITLES publish operation was called.

0→1 production pattern

1. Account and practice reconnaissance

The agent inspected the artist's existing TITLES work before proposing a new project. The resulting practice description emphasized tactile synthetic world-building, deadpan humor, handmade materiality, and a balance between whimsy and the uncanny.

2. World-system design

A sixty-concept bible was written before large-scale generation. Every god received:

• Domain
• Offering
• Taboo
• Miracle
• Material presentation
• Potential motion event
• Role within one of ten liturgies

This made prompting a consequence of the world rather than a sequence of unrelated image requests.

3. Visual-model test

Several artist models were tested against the same representative premise. Cut From Elsewhere was selected and credited as the project's consistent visual tradition.

4. Serialized generation and selection

The collection was produced in six-god liturgies. Each phase included canonical portraits, selected manifestations, local scores, evidence speech where justified, and a verified chapter film. A midpoint review identified the risk of formula hardening and shifted the project toward more varied civic, digital, domestic, and office domains.

5. Complete motion coverage

After the initial liturgy films, a coverage audit showed that 32 gods still lacked individual canonical videos. TITLES generated the missing transformations in bounded batches, producing a numbered 01–60 motion library.

6. Individual sound coverage

Local Python generated one nonverbal five-second material relic per god. The website then synchronized each relic with its corresponding manifestation.

7. Voice-system correction

An early polished synthetic-narrator direction was rejected. A sixteen-option audition reel led to the governing principle:

Speech is evidence, not narration.

The final system used private messages, municipal notices, and conflicting witnesses rather than an omniscient voice.

8. Evidence film

Existing assets were re-authored into a second film using:

moving portrait → canonical incident → moving final-frame residue → next moving slot

The feature became a municipal record rather than a compilation reel.

9. Website translation

The original Cabinet exposed each god as a hover/focus manifestation with a synchronized relic. The Evidence Cabinet turned the film into a temporal index: the playhead travels through the grid, each god becomes active at its edit time, and tiles seek directly into the master.

10. Deployment handoff

The two websites and all 181 runtime media files—60 stills, 60 manifestations, 60 relics, and one complete evidence master—were copied into a sanitized, checksummed, static-host package with no API keys or TITLES runtime dependency.

What this use case demonstrates

Artist models can become a world-level visual tradition

The artist model did not merely decorate one prompt. It held a sixty-work cosmology together while each object, domain, and miracle varied.

MCP is most valuable when it retains exact selectors and lineage

For this project, model IDs, adapter IDs, voice names, execution IDs, output IDs, and actual costs were production data—not implementation trivia.

Generation quality and production completeness are different metrics

A successful output did not guarantee:

• Complete 60/60 coverage
• Correct aspect ratio
• Semantic voice timing
• Continuous motion
• Consistent typography
• Browser-ready interaction
• Deployability

The project required repeated audits and conform passes after generation had technically succeeded.

Human review changed the work's form

The most consequential revisions came from direct review:

• Reject polished narration
• Use speech as evidence
• Cancel unnecessary first/last-frame extensions
• Restore native 4:5 presentation
• Replace floating lower thirds with website-style slots
• Eliminate the return to the opening still
• Move between gods while both remain in motion
• Turn the grid into the film's timeline

MCP enabled rapid realization; artistic judgment determined what deserved to remain.

Product-roadmap takeaway

The project suggests a natural TITLES adoption ladder:

Model discovery
→ canonical output generation
→ batch/coverage management
→ cast and provenance management
→ editorial relationship authoring
→ collection publishing
→ deployment export

TITLES already handled the first two layers powerfully. The strongest roadmap opportunity is to make the later layers first-class without erasing the ability to hand off to local specialist tools.