Planverity speaks the OpenAI chat-completions shape, so the tools you already run connect without a rewrite — n8n, Make, Cursor, LangChain, the OpenAI SDKs. Claude Code connects too, through the Anthropic Messages shape. Streaming included.
The fastest way to put a governed gateway behind work that already runs. An automation platform is where AI spend tends to escape first, because a workflow calls a model on a schedule and nobody is watching the invoice.
| Tool | How it connects | Status |
|---|---|---|
| n8n Run end to end against a real n8n MCP Server Trigger: discovery, import, classification, policy filter, ranking, then a real execution. On a 60-tool catalogue, 798 forbidden-tool exposures became 0 and schema tokens dropped 80% before any ranking was applied. | Point an OpenAI node at the gateway's base URL, or connect an MCP Server Trigger and let Planverity discover its tools. | Run end to end |
| Make | Use the OpenAI module with a custom base URL. | Works by protocol |
| Zapier | Any action that accepts an OpenAI-compatible endpoint and key. | Works by protocol |
| Retool | Configure an OpenAI resource with the gateway base URL. | Works by protocol |
| Flowise and Langflow | Set the OpenAI base URL on the model node. | Works by protocol |
There is no Planverity SDK to adopt, deliberately. A client library is one more thing to version, and the value here is that you do not change your code — you change one URL.
| Tool | How it connects | Status |
|---|---|---|
| OpenAI SDKs (Python, Node, and the rest) | Set base_url. Streaming and tool calls included. | Run end to end |
| LangChain and LlamaIndex | Pass the gateway as the OpenAI-compatible endpoint on the model. | Works by protocol |
| Vercel AI SDK | Configure the OpenAI provider with a custom baseURL. | Works by protocol |
| Anything that accepts a base URL | The gateway speaks the OpenAI chat-completions shape, so compatibility is a property of the protocol rather than a list we maintain. | Works by protocol |
Developer tooling is usually the least governed AI in a company: every engineer holds a key, and nothing records what was sent. Routing assistants through the gateway puts the same policy and budget around them as production traffic.
| Tool | How it connects | Status |
|---|---|---|
| Cursor | Set a custom OpenAI base URL and key in settings. | Works by protocol |
| Continue, Cline, and OpenAI-compatible assistants | Configure the provider as OpenAI with the gateway's base URL. | Works by protocol |
| Claude Code The gateway accepts the Anthropic Messages shape at /v1/messages, including streaming. ANTHROPIC_MODEL must name a pool (auto, or pool://your-pool) rather than a model: a bare model name is refused rather than reinterpreted, because silently treating one as a pool would defeat the primitive that decides which models are permitted at all. | Set ANTHROPIC_BASE_URL to the gateway and ANTHROPIC_MODEL to a pool alias such as auto. | Run end to end |
Planverity speaks MCP Streamable HTTP and can discover a server's catalogue directly — then governs which of those tools a model is ever offered.
| Tool | How it connects | Status |
|---|---|---|
| Any MCP server over Streamable HTTP Discovery imports capability, never authority. Tools arrive unclassified — a class no policy grants — so a freshly synced catalogue is exposed to nothing until an operator classifies it. A changed schema resets that classification, because the contract changed and the earlier judgement may no longer hold. | Register the server, sync its catalogue, classify the tools, publish a tool pool. | Run end to end |
A workflow that calls a model on a schedule is where AI spend escapes first, because nobody is watching it. Behind the gateway it inherits the same approved models, spend caps and refusal records as production traffic.
Connect an MCP server and its tools arrive unclassified — a class no policy grants. Nothing is exposed until someone classifies it, so a destructive tool cannot reach a model merely because a server advertised it.
Every request leaves a manifest: the model chosen, the policy and pool in force, what was refused, and what it cost. That holds whether the caller was your application, an automation or a coding assistant.
One base-URL change. 5,000 free executions, no card required — and if something you use is missing from this list, it most likely already works.
Get an API key