Neural Inverse Govern / Models and prompts
Which models exist for whom, and which prompts they run.
A model catalog you control, deny-by-default assignment to teams and projects, versioned prompts served through the gateway, and adoption and cost reporting without prompt content.
- Catalog
- Entered by you, with your prices
- Assignment
- Deny by default, deny wins
- Delivery
- Ed25519-signed model manifest
- Prompts
- Versions and labels, rendered by the gateway
- Reporting
- Metadata only, pseudonymised by default
Overview
Models and prompts is where you decide what the AI Gateway may serve. Platform teams connect providers, enter models and their prices, and assign them to the organisation, teams and projects. Applications reference prompts from a versioned library instead of carrying them in code. Insights reports who uses which model, MCP server, skill and harness pack, and what it costs, from usage metadata only.
What it does
What models and prompts covers.
The catalog and who may use it, the prompts that run against it, and what that use amounts to.
Provider connectionsAnthropic, OpenAI, Google, Azure OpenAI, Amazon Bedrock, Vertex AI, and self-hosted OpenAI-compatible or Ollama servers. Keys are write-only, held in Govern's vault, and handled by a permission separate from the one that configures the connection.
Your own catalogEach model is entered with its provider model id, uses, region, residency, the strictest classification it may serve, zero-retention flag, prices and limits. Govern does not discover models from providers, so the catalog works air-gapped.
Deny-by-default assignmentA model is allowed or denied for the organisation, a team or a project. It is usable only when allowed at a scope that applies to the person and denied at none. Leaving out the project never loosens a project's rules.
Classification and regionsRules per data classification require any model, a zero-retention model or a self-hosted one, optionally in listed regions. Restricted work defaults to self-hosted or zero-retention models.
Fallback chainsUp to twenty fallbacks per model. Each fallback is decided again from scratch, so it never reaches a model the caller could not use directly.
Defaults per useA default model for chat, agent, completion and embedding work at the organisation, team or project. The most specific scope wins, and a default the person may not use is skipped.
Signed model manifestThe Neural Inverse apps receive an Ed25519-signed list of what the person may use. It stays valid offline for a period you set, 24 hours by default and at most seven days.
Personal provider keysWhen you allow it, people store their own Anthropic, OpenAI or Google key. It is used only through the gateway, only for models you allowed, and only up to a classification you set.
Provider healthProbe a connection on demand or on a schedule, and see the last probes, latency and circuit state of its models, without a provider's error bodies.
Prompt libraryVersioned prompt templates with variables and organisation-wide values. Test, pin, fork and deprecate templates, and collect run statistics and feedback.
Prompt labelsApplications call a template by label, such as production, and the gateway renders it in OpenAI or Anthropic format. Moving a label promotes or rolls back a prompt without a client change.
Prompt runsEvery request made from a template is recorded with its version, label, the model that answered, measured tokens and cost, duration and outcome. Templates are also available to agents as an MCP tool.
Compare modelsA playground sends one request as the signed-in person to up to four models or routes side by side. Evaluation datasets sampled from sealed request logs score models and routes with exact, contains, regex, JSON schema and model-graded evaluators.
Model PassportA short-lived, Ed25519-signed attestation that binds an agent, and optionally its session and action, to the provider, model, version and hash it used, with a human-in-the-loop flag. Anyone with its id can verify it.
Adoption and cost insightsSummaries, time series and breakdowns by model, team, harness pack, MCP server and skill, with budgets and outcome signals such as changes kept, edited or discarded.
Metadata onlyUsage events carry identifiers, counts and outcomes, never prompts or code. Per-person views are pseudonymised by default, and exports of raw events are audited.
Bedrock models other than Anthropic models, and Gemini's native API, are not supported; Google and Vertex Gemini models are reached through their OpenAI-compatible endpoints. Spend on a personal key is not counted against spend budgets; its tokens are.
How it works
From a provider key to a model on someone's screen.
- ConnectAdd a provider connection and set its key, which nobody can read back.
- CatalogEnter each model with its prices, limits and the strictest classification it may serve.
- AssignAllow models for the organisation, teams or projects. Nothing is usable until you do.
- DeliverThe apps fetch the signed manifest of what each person may use.
- ServeThe AI Gateway decides every request against the same rules, rendering prompts by label.
- MeasureUsage, cost and outcomes arrive in Insights, without prompt content.
Who uses it
Built for every role that answers for AI.
Platform and AI enablement teamsConnect providers, curate the catalog and its prices, set defaults and fallbacks, and assign models.
Key custodiansSet and rotate provider keys under their own permission, without being able to change routing policy.
Application developersPublish prompt templates, move labels to promote or roll back, and compare models on their own traffic.
Engineering and finance leadsRead adoption and cost by team, model and tool, and export metadata for chargeback.
Enforced and recorded
Decided on the server, recorded with the change.
Assignments, classification rules and regions are enforced by the AI Gateway on every request. The manifest only tells the apps what the gateway will allow.
Enforced
Deny winsA deny at any applicable scope removes the model, for the requested model, every fallback and every route target.
ClassificationA request takes the strictest classification of its project, team, department and organisation. A client may raise it, never lower it.
Keys stay insideProvider keys are never returned. Changing a connection's address, region or settings drops its stored key, so nobody can redirect a key to a host they control.
Outbound guardCalls to providers refuse private, loopback, link-local and cloud metadata addresses unless your operator opens networks for self-hosted model servers.
Recorded
ConfigurationEvery change to connections, catalog, assignments, defaults and policy, in its own transaction. Key changes record the new version and the last four characters, never the key.
Prompt runsEach gateway request made from a template, with its version, label, model, tokens and cost.
PassportsEvery issued passport is stored and listed with a verification verdict.
Self-hosted and in your cloudProvider keys stay in the vault of your installation. Your operator decides which private networks count as self-hosted.
Air-gappedWith local models only: register on-premises inference servers and approve local models per data class. The catalog and prices are yours.
MixedRestricted work on self-hosted models, everything else on hosted providers, decided by classification rules.
Hosted by usWe run it for you, with the same controls and audit record.
Get started
Set it up with the documentation.
What's next
Being built now.
These are in development and not part of the product today.
Model registryRegister your own models and their versions, promote them through a lifecycle and record deployments.
Model safety evaluationsEvaluation suites and test cases scored per model, with comparisons, a leaderboard and stored red-team scenarios.
Golden sets and driftGolden question sets per project, run through a model against a baseline, to detect and attribute drift.
Passport revocationRevoke an issued passport, and keep a list of models approved for passport issuance.
Talk to our team.
Neural Inverse is available by request. We work with your engineering, security and procurement teams from the first conversation.