Every AI tool evaluated, what it actually does, and the moment it earns its place in a build. Filter by verdict or by the job you're trying to do — then grab the handoff prompt at the bottom to use any of it in another session.
Verdict
Job
Adopt — use it
Test — pilot cheaply first
On-call — niche
Watch — not usable yet
Skip — bookmark only
Signal — not a tool
LongHorizon-Harness
DreamX Team · open source
Adopt
A working shape for long, multi-step jobs. Instead of one swelling session, it splits work three ways: a Manager holding task state, an Executor doing one subtask in a fresh context then throwing its reasoning away, and an Auditor that independently verifies each result before it counts as fact.
Gains
Pass rate 51.8% → 80.7% on one benchmark; 3× on another
Adopt
The loop shape — not their Python harness
Pairs with
The cost pattern below
Reach for it when
Any long build where mistakes compound quietly — multi-page sites, the video and ad pipelines, big report generations. The key move: task state lives outside the working context, and nothing enters it unverified.
We already half-do this — the self-QA pass on the Tru-Vu pages was an audit step. The upgrade is making verification independent, not the same agent checking its own work.
Talking-head video from an audio track and a single picture. One-step generation fast enough to stream live — up to 200 frames per second — with the mouth and expression following the voice.
In → Out
Audio + one image → talking-head video
License
Check the repo — page is CC-BY-SA
Runs on
H200-class GPU → hosted only for us
Status
Code released, 2026 paper
Reach for it when
A mascot needs to talk — Rowdy fronting a GVEC promo, Sparky reading tonight's lineup. It needs only audio, which we already produce, so it's far less work per clip than motion transfer.
Head and face only — no full-body performance. Academic releases often carry non-commercial terms, so confirm the license before any client work.
Character motion transfer. Give it a reference image plus a driving video of a performance, and it renders your character doing that performance — face, expression, body, scene interaction. Also swaps characters into existing footage.
In → Out
Image + driving video → animated video
License
Apache-2.0 — commercially safe
Runs on
ComfyUI ~24GB GPU, or hosted per-gen
Status
v1 ships now; v2 weights pending
Reach for it when
A mascot needs full-body performance, or a character's identity has to hold across multiple ad shots where Seedance tends to drift.
Needs a driving video — real production overhead per clip. For a talking mascot, LeapTalk is the lighter path. Pilot hosted before any GPU purchase.
Singing voice synthesis from a written score. Hand it lyrics plus a melody — pitches, note lengths, tempo — and it sings them back expressively, with no phoneme timing work required.
In → Out
Lyrics + MIDI melody → sung audio
License
Unstated — verify before client use
Runs on
Open weights, Hugging Face
Reach for it when
A build explicitly wants an original sung jingle — a GVEC or Yard earworm for a reel — and a melody exists or can be written.
Sings only, no narration. Someone still has to compose the melody, so it's never the fast option.
Unified 3D system — generates 3D from text, answers questions about 3D scenes, edits models from plain-language instructions, and extracts individual parts on request. Outputs GLB.
In → Out
Text / 3D + instruction → GLB mesh
License
Unstated
Status
Code “Stay Tuned” — not released
Reach for it when
Any web-3D asset — product viewer, 3D mascot, hero object. GLB is exactly the right format for our model-viewer workflow.
Nothing downloadable yet. If a 3D asset is needed today, use the shipping Hunyuan3D 2.x line instead.
A very large frontier model — 2.4 trillion parameters, sparse mixture-of-experts, roughly a million tokens of context. A foundation model, not a tool you bolt on.
Access
API now; open weights promised, undated
License
None published yet
Context
~1M tokens
Reach for it when
Only really interesting as a cheap engine for low-stakes stages of our own pipelines — mechanical steps that shouldn't burn a premium model.
The open-weights release has no date, no license, and no confirmation of which checkpoint ships. Don't plan around it; re-check when weights actually land.
Meta's terminal coding agent with persistent background agents — a direct competitor to the setup we already work in. Closed weights, priced per token through Meta's developer API.
Access
Hosted API, pay-as-you-go
Weights
Closed — no self-hosting or fine-tuning
Reach for it when
Awareness only. Worth knowing the landscape; nothing here fills a gap we currently have.
Its discounted tier trades roughly 90% off for data sharing. Never route client code through that — the savings aren't worth the confidentiality exposure.
Text-to-CAD. Describe a mechanical part in plain English and it produces real dimensioned geometry — STEP and STL files ready for a CNC machine or 3D printer — plus the source that built it.
In → Out
Plain English → STEP / STL + Python
License
MIT — commercially safe
Runs on
Standalone Python app on port 8000
Reach for it when
Only for a manufacturing, machine-shop, or product-design client where a printable part file is the actual deliverable.
Makes mechanical parts, not web-display 3D — wrong output for our sites. And it's a whole app to host, not a skill to drop in.
A medical imaging model — reads 2D and 3D scans, answers clinical questions, and drafts radiology-style reports. Technically solid and openly licensed.
Domain
Clinical decision support
License
Apache-2.0
Reach for it when
Never, as things stand. This is a hard pass rather than a bookmark.
A web studio has no business in a diagnostic loop. Anything touching patient care carries regulatory and liability weight that belongs with counsel and clinicians — even if a medical client asks.
A robot control foundation model — watches through a camera, takes a spoken instruction, and outputs the physical movements to carry it out. Trained on 100,000+ hours of real manipulation.
Domain
Robot manipulation
License
Apache-2.0
Reach for it when
No realistic path to a Vivere build. Impressive work in a field we don't operate in.
Requires physical robot hardware to mean anything at all.
Not a tool — an announcement. An internal model produced ten new results on open problems that had stood for a decade or more, across sphere packing, group theory, cryptography and more, with proofs formalized in Lean so anyone can machine-check them.
Cost
~$2,000 in tokens, per OpenAI
Verified
Lean 4 certificates published
Why it's here
Nothing to adopt — but a real marker of where frontier reasoning is heading, and the machine-checkable proof angle is the interesting part: verification, not just generation.
Formal peer review still pending at time of writing.
Several of these were passes. Their architecture wasn't. Two ideas transfer straight into how our own long builds should run:
Cost — from Multi-Agent-CAD
Give each stage only what it needs: split a job into steps and pass a small structured payload between them rather than dragging the whole history through every call. And try plain code before reaching for a model — code → small model → big model, in that order. That restructuring alone cut their token use roughly 116×.
Reliability — from LongHorizon-Harness
Keep task state outside the working context, and let nothing into it unverified. Execute each subtask fresh, discard the reasoning, and have something other than the executor confirm the result before it counts as done.
Together they're the same instinct from two directions: small contexts, cheap executors, verified handoffs.
Use this anywhere
Handoff prompts for other sessions
Paste one of these at the start of any other Claude Code session. The agent will read this catalog live and apply only what genuinely fits — no need to remember what's in the list.
Universal — any session
Read the Vivere Capability Catalog at https://vivere-capability-catalog.pages.dev before we start.
It lists every AI tool/technique we've evaluated, with a verdict and a "reach for it when" trigger for each.
Rules for using it:
- Only suggest items marked ADOPT, TEST, or ON-CALL. Ignore WATCH, SKIP, and SIGNAL items unless I ask.
- Apply the two harvested architecture patterns (cost + reliability) to any long multi-step work in this session by default.
- If something in the catalog fits what I'm doing, say so proactively and tell me the caveat. Don't silently adopt anything, spend credits, or stand up infrastructure without asking me first.
- If nothing fits, say nothing about it and just do the work.
Campaign / content-render session (GVEC, The Yard)
Read the Vivere Capability Catalog at https://vivere-capability-catalog.pages.dev before we start.
I'm working on campaigns and content renders in this session. Use the catalog to flag optimizations as we go, with these rules:
- Only suggest ADOPT / TEST / ON-CALL items. Ignore WATCH, SKIP and SIGNAL.
- Mascot work: if a mascot needs to TALK, flag LeapTalk (audio + one image, no driving video). If it needs FULL-BODY performance or identity held across multiple shots, flag Wan-Animate. Both are TEST-status — hosted-endpoint pilot only, never a GPU purchase, and confirm licensing before anything client-facing ships.
- Original sung jingle: flag VocalRender (ON-CALL, niche — needs a composed melody, sings only, license unverified).
- Long render/build runs: apply the reliability pattern by default — keep task state outside the working context, execute subtasks in fresh contexts, and verify results independently rather than self-checking.
- Keep using our existing proven pipeline (HyperFrames, Higgsfield/Seedance, Blotato) as the default. These are additive for specific gaps, not replacements.
- Flag suggestions, then wait for my go-ahead. Never spend credits or publish without asking.
Web build / site session
Read the Vivere Capability Catalog at https://vivere-capability-catalog.pages.dev before we start.
I'm doing site/web build work. Use it as follows:
- Only suggest ADOPT / TEST / ON-CALL items.
- Apply both harvested patterns to any multi-page or multi-step build: narrow context per stage with small structured handoffs, plain code before a model call, task state held outside execution, and an independent verification pass before anything counts as done.
- 3D assets: Hunyuan3D-Buffalo is WATCH (code unreleased) — if I need 3D today, say so and use the shipping Hunyuan3D 2.x line instead, then run output through our normal GLB optimization budget.
- Flag suggestions and wait for my go-ahead before adopting anything new.