Share the behavior, not just the prompt.
Joseki WrapperHub is a registry for LLM behavior packages — prompts, fine-tunes, adapters, evals, and safety policies in one reusable format. Trust through community.
Read the concept →What it looks like
One package. One verified source of truth.
Joseki:WrapperHub becomes the place where I can find a Prompt for legal analysis and the Data Pack of 2026 tax code changes, and I can see that 500 other lawyers verified it worked this morning.
- 01
Prompt
A reviewed PromptSpec for the task — inputs, outputs, examples, compatible models.
- 02
Data Pack
A scoped dataset bundled alongside the prompt. Versioned, hashed, and licensed like code.
- 03
Liveness
Recency-weighted Works/Broken reports from real users — per model, updated in real time.
The problem
AI teams have outgrown one-off prompts.
Teams need reusable, testable, portable behavior across models, apps, and organizations. But today, prompts live in docs. Fine-tune configs sit in notebooks. Eval suites are buried in repos. Safety rules exist only in vendor dashboards.
Every model switch means rebuilding from scratch. Every handoff means copying prompts into Slack. Every audit means digging through git history. There is no standard way to package, version, or share the full behavior of an LLM deployment.
The solution
One package. Everything your model needs.
Joseki WrapperHub lets you create Behavior Packages — versioned bundles that capture everything about how an LLM should behave.
- PromptSpecSystem prompts, templates, and examples
- FineTuneSpecTraining recipes and configurations
- EvalSpecEvaluation suites and benchmarks
- SafetySpecSafety policies and guardrails
- LicenseSpecUsage terms and attribution
- InstallSpecRuntime setup and dependencies
- EvidencePackProvenance and audit artifacts
What you can publish
Four building blocks for LLM behavior
| Block | What it carries |
|---|---|
| Prompt packages | Reusable prompts with examples, metadata, and version history. Install them instead of copying between projects. |
| Fine-tune recipes | Portable training recipes for closed and open model providers. One recipe, any backend. |
| Adapter registry | LoRA adapters with compatible model manifests, artifact hashes, and runtime instructions. |
| Security scanning | Prompt-injection scans, license checks, and provenance tracking on every published version. |
Live signal
“Works on GPT-4o. Broken on Claude 3.5.”
Forget stale “Last Updated” timestamps. Every wrapper carries a Liveness Score — a recency-weighted aggregate of Works and Broken reports from real users, broken down per model.
| Model | Score | Works | Broken |
|---|---|---|---|
| gpt-4o | 97 / 100 | 342 | 3 |
| claude-3.5-sonnet | 48 / 100 | 51 | 62 |
| llama-3.3-70b | 12 / 100 | 4 | 28 |
Creators who maintain their wrappers earn reputation. Reputation compounds into a Go-style kyu / dan rank, the same way Stack Overflow rewards consistent contributors.
For developers
A CLI that speaks your language
Install, inspect, and evaluate behavior packages before using them in production.
$joseki install customer-support-classifier$joseki inspect customer-support-classifier$joseki eval customer-support-classifier
Built for trust
Governance is not optional
Joseki is built for teams that care about security, repeatability, and trust. Every package carries its own evidence trail.
- Package signing with verifiable authorship
- Artifact hashes for tamper-proof integrity
- License metadata that travels with every version
- Automated safety checks on every version
- Evidence packs for audit and provenance
Pricing
Free for the open registry. Paid for the private vault.
Joseki is freemium. Public Behavior Packages are free forever. Premium unlocks private vaults for teams, larger data packs, and priority eval runs.
Premium
$24 / month
Private vaults, 5 GB data packs, priority eval runs, verified badge. Two seats included.
| Plan | Free | Sponsored |
|---|---|---|
| Price | $0 | By request |
| Includes | Public wrappers, votes, Liveness reports, kyu/dan rank. | Featured registry slot once we cross 5k monthly reviewers. |
Stay in the loop
Subscribe for updates
We'll let you know when new features, wrappers, and eval tools ship — no spam.