Including the ones with uncomfortable answers. An FAQ that only handles softballs tells you the rest of the site is marketing.
An AI gateway with a policy engine in front of it. Your application sends one OpenAI-compatible request; Planverity decides which models are permitted for that request, picks the cheapest one that clears your quality floor, falls over to another if it fails, enforces your spend caps, and records what ran and why. It is bought most often by people running AI for several clients or teams at once, because each gets its own policy, budget and approved models.
Calling a provider directly gives you one model, one key, one bill and no record of why anything happened. Planverity adds the parts you would otherwise build: routing across providers, automatic fallback, spend caps that hold under concurrency, tool permissions, and an execution manifest per request. If you run one model successfully and have no governance or multi-client requirement, the honest answer is that you may not need this yet.
On our own benchmark it has the best cost-per-success of any strategy we tested, but it does not dominate simply using one good mid-tier model. Against always-Sonnet it offers a dial: 95.1% success at half the cost, or equal success at about 5% more. It clearly beats always-premium and a hand-rolled task-tier heuristic. We will not quote you a savings percentage before measuring your workload.
One base-URL change. Planverity speaks the OpenAI chat-completions shape, so the OpenAI SDKs, LangChain, LlamaIndex, the Vercel AI SDK, n8n, Make and Cursor work unmodified, streaming included. There is deliberately no Planverity SDK to adopt.
Yes. Set ANTHROPIC_BASE_URL to the gateway and ANTHROPIC_MODEL to a pool alias such as auto. The gateway accepts the Anthropic Messages shape at /v1/messages, including streaming. ANTHROPIC_MODEL must name a pool rather than a model, because a bare model name is refused rather than silently reinterpreted as one.
Anthropic, OpenAI, Google Gemini, DeepSeek, Z.ai and Ollama — 19 models that can actually serve a request today. That is a smaller catalogue than gateways advertising hundreds, and it is the number of endpoints we have priced, benchmarked and can route to, rather than the number we could theoretically reach.
Yes, and that is the default. Under bring-your-own-key the provider bills you directly with no markup, and your credentials are resolved at the execution boundary rather than held in the request path. You can also add prepaid Managed credit, priced as the provider's cost plus a stated access fee, never hidden inside a marked-up token rate.
Policy filters the tool list before anything ranks it, so relevance can only ever reorder an already-permitted set. Tools discovered from an MCP server arrive unclassified — a class no policy grants — so a freshly synced catalogue is exposed to nothing until an operator classifies it. Measured on a 60-tool catalogue across 38 tasks: 798 forbidden-tool exposures with no gateway, zero with one.
Yes, and the mechanism matters. Summing spend then comparing loses under concurrency: two requests both read 95 of a 100 cap, both estimate 3, both pass. The check and the reservation happen in one transaction behind a lock on the budget scope, priced at your max_tokens rather than an estimate, with the headroom returned once the real cost is known. Verified live: six concurrent reservations against remaining budget, two granted, four refused, cap not breached.
No, and this is the invariant the whole system is built on. Policy decides what is permitted; routing, ranking and any semantic feature optimise only inside that. Region and provider rules filter the candidate set before any scoring, so a cheaper model in the wrong jurisdiction is not a trade-off to weigh — it is not a candidate.
The control plane runs in Frankfurt and its database in Ireland — both in the EU. Provider calls go wherever your policy permits. Planverity the company is established in the United Kingdom, which is a different fact from where the software runs, and if your requirement is an EU-established processor rather than EU-hosted processing, that distinction is the one that matters.
No. An execution record holds metadata, token counts, cost, the policy and pool in force, and what was refused — never the text of the request or the reply.
No. We hold no certification and make no certification claim. Row-level security is enabled with deny-all defaults on every table, API keys are stored only as hashes, tenancy is enforced at the query rather than the handler, and the security page states what is implemented and what is not.
Free is $0 for one organisation and 10,000 executions a month. Pro is $99 a month for 3 organisations and 250,000 executions. Agency is $299 a month for 25 organisations and 1,000,000 executions. Enterprise is contract-priced with unlimited organisations. Overage is $0.50 per 10,000 executions on Pro and Agency. Pricing is on the control plane, never a markup on your inference.
One request you send to the gateway, whatever it takes to satisfy it. A fallback to a second model after the first fails is still one execution — charging per attempt would price you for our reliability engineering rather than your usage. A request refused by policy or a spend cap never reaches a provider and is not counted, and dry-run planning is free.
Streamed tool calls are not reassembled. SSO and SCIM are not built. Semantic Execution Memory, PII filtering of request bodies, and the human-approval execution chain are designed but not shipped. There is no evaluator running on live traffic, so verified quality is not computable — the console states plan-compliance instead and labels it as such. Each platform page says where its own fit is partial.
Ask directly, or bring a real workload to a briefing and we will measure it.