TITLES MCP vs. local production — capability map

Co-authored project document. Product observations describe the TITLES MCP/operator surface used during this production on 2026-07-31; they should not be read as claims about unreleased or untested TITLES capabilities.

Executive view

TITLES supplied the high-value generative and provenance layer: model discovery, artist-model selection, still generation, image-to-video manifestation, speech synthesis, execution tracking, outputs, and cost-bearing model access. Local tools supplied the production layer required to turn those generations into a coherent sixty-work system: editorial selection, deterministic sound, voice treatment, sequencing, semantic synchronization, graphic design, responsive interaction, verification, and deployment packaging.

The project succeeded because the two layers were complementary. TITLES created canonical media; local tooling converted canonical media into authored forms.

Capability split

Production need	TITLES MCP / TITLES models	Local tools
Model discovery	Search and inspect artist and generic models, operators, adapters, allowed voices and parameters	Compare outputs and record the chosen tradition
Persistent generative workspace	One TITLES session containing all image, video, and speech nodes	Local folder system and cross-media manifest
Canonical portraits	txt2ImgNode; two candidates per god	Human/agent selection, naming, download, sequential filenames
Canonical manifestations	img2VidNode; reference-led five-second clips	Coverage matrix, retries, local fallback study, standardized library
Voice performances	txt2SpeechNode; multiple providers and preset voices	Casting system, script/channel assignment, processing, assembly, timing
Audio relics	No paid generation used	Sixty deterministic five-second nonverbal relics synthesized in Python
Chapter scores	TITLES music generation was quoted but intentionally not run	Ten procedural scores synthesized locally at zero generation cost
Voice world-building	Raw TTS performances	Answering-machine, municipal-intercom, and shared damaged-witness DSP chains in FFmpeg
Film editorial	Source assets available	Still holds, detail crops, moving portraits, final-frame residue, split screens, slot-scroll transitions, chapter structure
Semantic synchronization	Raw media durations and scripts available	Per-god cue manifest; impossible phrase → manifestation and relic; unresolved phrase → residue
Typography	Not used for title generation	ImageMagick-rendered cards, website-slot framing, explicit line breaks and metadata hierarchy
Mixing/mastering	Raw audio output	Resampling, normalization, sidechain ducking, limiting, AAC mastering
Quality assurance	Execution completion/output state	ffprobe, complete FFmpeg decode, frame hashes, audio levels, path/count invariants
Interactive Cabinet	Not part of the generation workflow used	HTML/CSS/JS still-to-motion/audio grid, lazy playback, keyboard/touch behavior
Evidence timeline site	Not part of the generation workflow used	Exact film cue map, click-to-seek, active/entering/past states, shared preview decoder, responsive QA
Hosting handoff	Individual cloud outputs	Self-contained static package, checksums, credential scan, byte-range deployment guide

What TITLES MCP made possible

1. Artist-model discovery with creditable provenance

The project could inspect live artist models and select a visual tradition rather than prompting an anonymous general model. The final sixty portraits use Cut From Elsewhere by @juujuumama, and the model/adapter remains attached to the project record.

This mattered aesthetically and economically: the model was not treated as a style adjective but as an authored contribution with a creator, training practice, model page, model ID, and adapter ID.

2. One persistent session for a large serialized world

Session:

f2739b27-bd62-4a36-ad50-cccb6569c290
The Cabinet of Minor Gods — Launch Pantheon

The session currently records:

• 62 text-to-image nodes / 124 image outputs
• 62 image-to-video nodes / 60 successful video outputs
• 119 text-to-speech nodes / 119 speech outputs
• 243 total operator instances

The two video nodes without outputs were moderated Lost Glove attempts that were refunded; the successful third prompt used restrained ambient material behavior.

3. Cross-provider speech casting

The MCP/operator layer exposed provider-specific model selectors and allowed voices. That enabled one recurring town-like cast rather than sixty unrelated default voices:

• River — ElevenLabs Turbo v2.5 — private answering-machine caller
• Patient_Man — MiniMax Speech 2.8 Turbo — municipal clerk
• Charon and Aoede — Gemini 3.1 Flash TTS — conflicting witnesses

Other speech models were useful during the wider audition process, including MiniMax Speech 2.8 HD and Chatterbox HD.

4. Programmatic generation at series scale

MCP calls allowed the agent to:

• Run exact operator/model/adapter combinations
• Await executions
• Inspect outputs and execution states
• Retrieve output IDs and downloadable assets
• Track paid runs and refunded failures
• Continue generation in the same session

This made a 60-image / 60-video / 119-speech-node project tractable without manually recreating every node in the studio UI.

Work completed locally because it was not exposed by the TITLES workflow used

Deterministic sound design

Sixty five-second nonverbal audio relics and ten chapter scores were synthesized locally. Deterministic generation offered reproducibility, zero marginal generation cost, and an exact one-relic-per-god naming invariant.

Diegetic voice processing

The raw TTS voices became in-world evidence only after local DSP:

• Telephone-band filtering
• Bit crushing
• Echo and damaged-machine acoustics
• Municipal presence EQ
• Shared two-witness recording treatment
• Speed correction
• Loudness normalization and resampling

Non-linear editing and conform

The finished film required operations beyond source generation:

• Intact still holds and object-specific crops
• Moving still inspection
• Canonical manifestation played once
• Extraction and animation of final decoded frames
• Before/after and witness split screens
• Frame-quantized cue records
• Cross-dissolves and 0.6-second cabinet-scroll transitions
• Ten chapter assemblies and a clean master
• Multiple editorial conform passes after human review

Graphic system and typography

ImageMagick generated the title cards, liturgy cards, closing card, borders, numbering, channel labels, names, and reinforcing domains. This was necessary both because the installed FFmpeg text filters were unreliable and because the graphic system needed to match the website rather than a generic video-title template.

Interactive and distributable experiences

Two static web experiences were authored locally:

1. The original Cabinet: still → hover/focus manifestation + synchronized audio relic.
2. Evidence Cabinet: sticky/stacked film player synchronized to a sixty-slot timeline grid.

The second experience required exact transition-aware seek times, one shared muted preview decoder, active/entering/past states, URL hashes, keyboard controls, responsive containment, and HTTP byte-range testing.

Roadmap opportunities suggested by the project

A. Project-level export and handoff

A first-class export could provide:

• Selected canonical outputs
• Stable sequential naming
• Original prompts and model/adapter selectors
• Execution/output IDs and costs
• A sanitized manifest
• Bulk download without expiring signed URLs
• Optional static gallery or website bundle

B. Provenance and spend ledger

The project manually maintained cumulative spend across phases, successful runs, quotes not approved, and refunded failures. A project ledger could expose:

• Spend by operator/model/media type
• Phase and batch totals
• Refunded/failed executions
• Selected versus rejected outputs
• Creator payout/credit information
• Exportable CSV/JSON

C. Batch orchestration as a first-class primitive

Useful controls would include:

• Generate N items from a manifest
• Bounded concurrency
• Automatic sequential download
• Rate-limit backoff
• Per-item retries and fallback prompts
• Resume from incomplete batches
• Coverage invariants such as 60/60

D. Editorial timeline and local-NLE bridge

A TITLES timeline or export bridge could support:

• Still holds, crops, pans, and masks
• Freeze-frame extraction and residue animation
• Audio/video cue placement
• Chapter assembly
• Crossfades and moving slot transitions
• Lower thirds and metadata cards
• Export to an editable FFmpeg/OTIO/EDL representation

The important roadmap lesson is that artists need to author relationships between outputs, not only outputs themselves.

E. Speech as a cast and recording system

Potential product primitives:

• Named recurring cast stored at project level
• Provider/model/adapter/voice pinning
• Multi-performance units
• Shared acoustic environments
• Telephone, answering-machine, intercom, archival-radio, and damaged-recorder presets
• Word/phrase timestamps for semantic editing
• Per-line cost and provenance

F. Audio effects and deterministic procedural sound

The Cabinet needed sixty object-specific effects, not sixty songs. Useful additions could include:

• Short material-event synthesis
• Parameterized procedural generators
• Reproducible seeds
• Loudness targets
• Batch one-sound-per-output workflows
• Stem export

G. Automated media QA

A project verifier could report:

• Missing/duplicate outputs
• Decode failures
• Duration and aspect-ratio mismatches
• Sample rate/channel mismatches
• Loudness and true peak
• Stream start offsets
• Coverage across a numbered collection

H. Publishable collection and timeline surfaces

The generated-media relationship was ultimately expressed through custom HTML. TITLES could offer exportable components for:

• Numbered grids
• Still-to-motion hover/focus
• One-active-audio enforcement
• Film-synchronized timeline grids
• Direct item hashes
• Responsive portrait players
• HLS/static-host deployment

Adoption lesson

The most productive division of labor was:

TITLES for model discovery, generative execution, artist-model lineage, outputs, and cost-bearing inference; local tools for authored systems, deterministic transformation, editing, verification, and deployment.

A future TITLES roadmap does not need to replace every local tool. The highest-leverage opportunity is to make the handoff between generation and authorship explicit, structured, reversible, and exportable.