No control over access
People sign up for AI tools on their own, share provider keys between teams, and build assistants on personal accounts with company documents. Nobody can say who has access to what, or take it back when someone leaves.
Tenmore Gate · AI governance for the enterprise
Say yes to AI. Say no to shadow AI.
Tenmore Gate is the single, governed entry point to every AI model your company uses. IT decides who uses which models, sensitive data stays inside, changes are tested before they reach everyone, and every dollar is accounted for. Agents built from your teams’ documents are managed the same way. Run it in our cloud or inside your own network.
- Works with every major model provider and your own models
- Sign-in and user provisioning with Microsoft Entra ID
- Exports to Splunk, Sentinel and Datadog
- Self-hosted or cloud
Works with
Amazon Bedrock
Azure OpenAI
Google Gemini
Anthropic
OpenAI
Mistral
Cohere
your own servers
and more
The problem
AI is already in your company. Most of it is outside IT's view.
Teams adopt AI faster than policy can follow. The result is the same three risks in almost every organization.
Sensitive data leaves
Customer records, contracts and credentials get pasted into prompts and sent to outside providers, with no record that it happened.
Costs nobody owns
AI bills arrive as a few large numbers on several invoices. Finance can't tell which team, project or person they belong to.
Agents
An agent for every team, managed like the rest of your AI.
Give a team an agent that knows its handbook, price list or runbook. Tenmore Gate builds it from the team’s documents and tests it against them before anyone uses it. From then on it’s managed like any gateway: admins decide who may use it, every change is versioned, and it runs on approved models with the same filters, redaction, cost tracking and audit log.
- Built from a team’s own documents, with no AI expertise needed
- Tested against those documents before it goes live, with the score to show for it
- Rebuilt when the documents change, with the new test score next to the last one before you update it
- Created, changed and shared only by admins, with every change in the audit log
- Experts can write an agent’s skills themselves, and download any agent’s files
HR policy assistant · built from 6 documentsready
Test questions answered correctly18 of 20
What it knows
- Parental leave
- Vacation and holidays
- Expense claims
- Remote work
- Benefits enrolment
- Couldn’t read page 14 of Benefits.pdf
- Removed 2 personal phone numbers
app.tenmore.com/w/acme/chat/hr-policyTry it
Example: an agent built from an HR handbook.
On your website, too
Make an agent public and it answers visitors on your website, from public information only. It has its own spending cap and usage limits, can’t reach company tools, tells every visitor it’s an AI, and switches off at once. The assistant on this page runs on Tenmore Gate.
What Tenmore Gate does
One control point for all of your company's AI.
Govern access
- Approve which models each team can use
- Company sign-in and automatic provisioning
- Revoke access in one place
- Every admin change in the audit log
Protect data
- Policies written in plain English
- Stop or redact sensitive data
- Checks answers before people read them
- A record of every policy decision
Build safe agents
- Agents for any team, from its documents
- Agents use company tools through Tenmore Gate
- Each person's own permissions apply
- Risky actions stopped before they run
Change safely
- Every change versioned
- Test new models and policies first
- Roll out to a pilot group
- Roll back in one step
Control cost
- Spend per team and per person
- Budgets with alerts or hard limits
- One view across every provider
- Mix local and frontier models
Policies
Write the policy the way you'd say it.
Describe what shouldn't leave the company. Tenmore Gate turns your description into a set of narrow yes/no questions and asks a fast decision model about every request. Narrow questions are far more accurate than one broad instruction, and you never have to write them.
- Set the threshold that matches your risk, then block, log, or both
- Keyword lists for the things that are simply off-limits
- Reads documents and text files attached to the prompt
- The same filters on answers, before anyone reads them
Redaction
Sensitive values come out. The work goes on.
When blocking is too blunt, Tenmore Gate replaces the value and lets the request through. It works on prompts before the model sees them, on answers before your people read them, and on tool results before they reach the model.
- Works on streamed answers too
- Add your own patterns, such as project code names, and test them before saving
- Logs count each kind of value found. The values themselves are never kept
Output filter · redactstreamed
From the modelSend the invoice to dana@northwind.io, charge card 4539 1488 0343 6467 and use key sk-ant-api03-Qx7… for the retry.
To your userSend the invoice to [email], charge card [card] and use key [secret] for the retry.
secret API keys, tokens
private-key
card Luhn
iban mod-97
national-id
email
phone
ip
custom your patterns
Tools and agents
Let agents use tools without handing over the keys.
Add MCP servers to a gateway and Tenmore Gate runs their tools itself. The model provider never sees the server's address or credentials, and your servers don't need to be on the internet. Every call is checked before it runs, and every result before the model reads it.
- Each person signs in to tools with their own account, or the team shares one
- Pick exactly which tools a gateway offers
- Write tool rules in prose, like "never delete anything in GitHub"
- A blocked call never runs; the model is told why and can explain
- Also checks the tool calls your own app makes
Tool activity · one request3 servers
- ✓ linear__search_issues(query: "checkout timeout")12 results · run as maya@acme.com
- ✓ github__get_pull_request(repo: "web", number: 482)result redacted: 1 × [secret]
- ✕ github__merge_pull_request(repo: "web", number: 482)Blocked by policy: "No merges to main from agents"
- ✓ slack__post_message(channel: "#release")allowed · logged
Example: one agent request.
Change management
Manage AI the way you manage software: tested, staged and reversible.
A new model, a stricter policy or a new tool can change what hundreds of people get back. In Tenmore Gate, every change to a gateway is a new version. Test it, release it to a pilot group, compare, then promote it to everyone, with each step in the audit log.
- Drafts, staging for admins, gradual rollout to chosen groups, then production
- Evaluations: your own test prompts, run against any stage
- Compare the stages side by side in Monitoring: errors, latency, cost
- Go back to the previous version in one step
Cost and monitoring
Know what AI costs, down to the person.
Tenmore Gate records tokens and cost on every call, with monthly totals per person and per gateway. Prices for common models are filled in for you.
- A call log with status, latency, tokens, cost and the reason for each failure
- Every filter and redaction event, in one place
- Budgets per workspace, gateway or person, with alerts or a hard limit
- Everyone can see their own usage
Models
Any model. The same rules.
Tenmore Gate isn't tied to any provider. It connects directly to the major clouds and model makers, to any service with an OpenAI- or Anthropic-compatible API, and to models on your own servers. Every one of them gets the same access rules, policies, redaction and cost tracking.
Direct connections
- Amazon BedrockClaude, Llama, Mistral, Nova and more
- Azure OpenAIOpenAI models in your Azure tenant
- OpenAIGPT and reasoning models
- AnthropicClaude models
- CohereCommand models
- Jev or equivalentdecision models for policy checks
Compatible providers
Any service with an OpenAI- or Anthropic-compatible API, for example:
Google GeminiMistralxAIDeepSeekGroqTogether AIFireworksOpenRouter
Your own servers
Open-weight models on hardware you control, inside your network:
vLLMOllamallama.cppLM StudioSGLangMLX
- Apps call models in whichever format they already speak; Tenmore Gate translates
- Switch the model behind a gateway without anyone changing code
- Pair models from different providers as primary and fallback
Deployment
Run it behind your firewall, on your own models.
Tenmore Gate ships as a container. Run it inside your network with your own database, and point it at models you host yourself. Prompts, answers, tool calls and policy checks can all stay on hardware you control.
- Local models on vLLM, Ollama or MLX, on internal addresses
- Policy checks with a local decision model, so even the filters stay in-house
- Internal MCP servers, with no public address needed
- Mix local and cloud models: keep sensitive work in-house, use a frontier model where you allow it
- The same filters, redaction and cost tracking whichever model answers
Gateway "internal" · inside acme-corp networkexample
- ● "local" → GLM-4.7-Flash on mlx_lm.serverhttp://gpu-01.acme.internal:8200 · tool calls supported
- ● decision model → Jev or an equivalent, self-hostedhttp://gpu-01.acme.internal:8100 · input and output filters
- ● MCP → jira.acme.internal, wiki.acme.internalno public address · each person signs in
- ● "frontier" → Claude via Amazon Bedrockallowed for the research team only · redaction on
Example setup.
Integrations
Fits the tools your security and operations teams already use.
Tenmore Gate sends what it sees to your SIEM, your observability platform and your alerting tools, as it happens. Prompts and answers are never included, only who, what, when and the outcome.
- Splunk, Microsoft Sentinel, and the OCSF format many other SIEMs read
- OpenTelemetry, for Datadog, Grafana, New Relic and others
- Alerts to Slack, Microsoft Teams and PagerDuty; webhooks for anything else
- Archives in S3, Google Cloud Storage or Azure Blob
- Users added and removed automatically from Microsoft Entra ID (SCIM)
- An audit log of every admin change
Integrations · acmeexample
- ● Splunk · OCSFpolicy and admin events · delivered 2 minutes ago
- ● Datadog · OpenTelemetryevery model call as a trace · delivered just now
- ● Microsoft Teams · #ai-governanceblocked requests and budget alerts
- ● PagerDutymodel failures on production gateways
- ● S3 · s3://acme-audit/gate/all events, hourly archive
Compatibility
No rewrites for your developers. No lock-in for you.
Existing apps and tools keep working: developers change one URL and use a personal key. Tenmore Gate translates between API formats, so an app written for one provider can use a model from another, and callers can't reach a model you didn't approve.
- Works with the OpenAI and Anthropic SDKs, and the many tools built on them
- A fallback model answers when the first is down, overloaded or refuses
- Automatic retries
from openai import OpenAI client = OpenAI( base_url="https://api.tenmore.com/acme/gateway/assistants", api_key="tmgw_…", # your personal gateway key ) reply = client.responses.create( model="smart", # a name your admin chose input="Summarize the attached contract.", )
from anthropic import Anthropic client = Anthropic( base_url="https://api.tenmore.com/acme/gateway/claude", api_key="tmgw_…", ) message = client.messages.create( model="smart", max_tokens=1024, messages=[{"role": "user", "content": "Draft the release notes."}], )
curl https://api.tenmore.com/acme/gateway/assistants/chat/completions \ -H "Authorization: Bearer tmgw_…" \ -H "Content-Type: application/json" \ -d '{"model": "fast", "stream": true, "messages": [{"role": "user", "content": "Hello"}]}'
Performance
Tenmore gateway overhead ~10 ms
In a typical 2-second call, about 10 ms is Tenmore Gate; the rest is the model.
~10 msTenmore Gate's own time per request (median; under 15 ms at the 95th percentile)
< 5 msper keyword, pattern, redaction or tool-rule filter
~150 msper decision-model policy check; several policies share one check
Livestreams pass through as they arrive, unless an output filter must be able to block the answer
Medians measured 2026-10-01 on AWS, at low load; cold starts excluded. Measure your own gateways with the Filter latency button on any gateway's page.
Under the hood · for your architects
What happens to one request.
Every call passes the same checks, in the same order, whichever model answers. This is one example request, step by step.
POST /acme/gateway/assistants/responses
- ✓accessmaya@acme.com may use assistants
- ✓stagerollout, revision 14 · Maya is in the pilot group
- ✓model"smart" → Claude Sonnet, Anthropic
- ✓input filterCustomer financial data · p(yes) 0.04, blocks at 0.60
- ◆redaction1 × [email] 1 × [secret] before the model sees them
- ⇄translationOpenAI Responses → Anthropic Messages
- ↻fallback529 overloaded → Gemini, Google, before a byte reached Maya
- ✓tool filtergithub__merge_pull_request blocked by "No merges to main"
- $usage1,842 in · 411 out · $0.0097 to Maya, gateway "assistants"
Streamed answer, as Maya sees itI've drafted the reply to [email] and left the key [secret] out of the excerpt. I can't merge the pull request: that's blocked by your team's policy.
Example request. Names, prices and scores are illustrative.
- Access. Tenmore Gate checks the person may use this gateway, picks the version they should get (production, or a rollout they're part of), and maps the model name the admin chose to an approved provider model.
- Input policy. Filters check the prompt, including attached documents and tool results. Sensitive values are replaced before the model sees them.
- Routing. Tenmore Gate translates between APIs when needed, retries, and switches to the fallback model if the first one fails.
- Tools. Agent tool calls are checked against your rules before they run.
- Output policy. The answer is checked and redacted as it streams back.
- Record. Tokens and cost count against the person, the gateway and any budgets, and the event goes to your integrations. The user’s prompt itself is not stored.
Security and privacy
Built by people who'd rather not hold your data.
Tenmore Gate sees every request, so it's designed to remember as little as possible.
User prompts are never stored
Tenmore Gate records which policy acted, for whom and when, with a score or count. Never the text of a user’s prompt or its answer.
A complete audit trail
Every policy decision and every admin change, with before and after values, kept as long as you choose and exported to your SIEM.
Your identity provider
Single sign-on and automatic provisioning with Microsoft Entra ID, or passwords with two-step verification. Sign-ins are logged.
Provider keys locked away
Encrypted at rest and never shown again, not even to admins. People use personal gateway keys you can revoke at any time.
Approved models only
People and apps can only reach the models IT approved for them.
Your region or your network
Use our cloud, or self-host Tenmore Gate as a container next to internal models and tools. Nothing has to touch the public internet.
Compared
More than a proxy.
Most LLM proxies route traffic and count tokens. Tenmore Gate governs it.
| Capability | Typical LLM proxy | Tenmore Gate |
|---|---|---|
| One endpoint for several providers | ✓ | ✓ |
| Cost tracking | Per key, user or team | Per person and gateway |
| Access by person, with allow and deny lists | Some | ✓ |
| Policies written in plain English | — | ✓ |
| Agents built from your documents, under the same rules | — | ✓ |
| Public agents for your website, with spending caps and limits per visitor | — | ✓ |
| Decision-model filters with tunable thresholds | Rare | ✓ |
| Redaction on streamed answers and tool results | Rare | ✓ |
| Runs MCP tools itself, with per-person sign-in | Some | ✓ |
| Tool rules that block a call before it runs | Some | ✓ |
| Translation between provider API formats | Common | ✓ |
| Self-hosted behind your firewall, with local models and local policy checks | Some | ✓ |
| Staged rollouts and evaluations for model and policy changes | Rare | ✓ |
| Audit log of admin changes, exported to Splunk or Sentinel | Some | ✓ |
| Budgets with alerts and hard limits | Common | ✓ |
| User prompts never stored | Varies | ✓ |
Use cases
How teams use Tenmore Gate.
Five situations most organizations meet as AI spreads, from the person responsible to the result.
Model rolloutsRoll out a new model without surprisesTest on your own work, release to a pilot group, then everyone.
Internal agentsAn agent for every teamBuilt from each team’s documents, shared with the right people.
Public agentsAnswer your website’s visitorsDay and night, within a budget and limits you set.
Coding agentsCoding agents, within limitsTool rules, personal keys and a cost per developer.
Sensitive workSensitive work on your own hardwareContracts stay on the company’s own model.
Get started
Put your company's AI behind one gate.
Try Tenmore Gate with your own models and teams, ask us anything about it, or write to us and a human will reply.