Every model call,
controlled and on the record
From org structure and roles to API keys, audit logs and data redaction, TokenHub builds enterprise-grade governance into every layer of the gateway. It runs on your own servers: teams get scoped credentials, while provider keys, logs and sensitive data never leave your company.
- Multi-level RBAC
- Single sign-on
- End-to-end audit trail
- Self-hosted · air-gap ready
-
AuthenticationValid API key · source IP on the allowlistPassed
-
Permission checkDeveloper · project “Knowledge Q&A” · may call qwen-plusPassed
-
Quotas & rate limits62% of monthly quota used · 41 / 300 RPM nowPassed
-
Content safety1 phone number found, redacted before forwardingRedacted
-
Audit logCaller, model, tokens and cost recordedLogged
Four levels that mirror how your company works
Access and usage roll up through organization → team → project → member. Admins at the top set the security baseline; everyone below manages their own scope without losing control or getting stuck in approvals.
- Organization
- One identity source and sign-in policy, global security rules, and centrally managed provider keys
- Team
- Maps to a department or business unit, with its own budget and quotas; team admins manage their own members
- Project
- Issues API keys and sets the model allowlist, attached knowledge bases and rate limits
- Member
- Holds a separate role in each project; personal usage and every action are fully traceable
-
R&D
-
Code Assistant
-
WWangAdmin CChenDeveloper LLinDeveloper
-
-
Knowledge Q&A
-
-
Customer Support
-
Support Bot
-
Built-in roles out of the box, custom roles when you need them
Grant roles at the organization, team or project level. Permissions inherit and narrow automatically as you go down.
Super Admin
The org owner. Manages identity sources, global security policy, provider keys and overall budgets.
Org level · 1–2 peopleAdmin
Manages members, models, knowledge bases, budgets and API keys for the teams or projects they own.
Team / project levelDeveloper
Calls models, uses knowledge bases and manages their own API keys inside authorized projects.
Project levelViewer
Sees usage reports and settings but can't change anything. A good fit for finance, auditors and business owners.
Any levelNeed more than the built-ins? Mix and match permissions
For example, let “Data Labelers” upload and search knowledge base documents only, or let Finance view and export usage and cost reports only. Permissions go down to individual actions and can be scoped to where they apply.
Permission matrix (built-in roles)
| Permission | Super Admin | Admin | Developer | Viewer |
|---|---|---|---|---|
| Models | ||||
| Add and configure model channels (provider keys / local endpoints) | ||||
| Configure routing, failover and the model allowlist | ||||
| Call models / use the Playground | ||||
| Knowledge bases | ||||
| Create / delete knowledge bases | ||||
| Upload and manage documents | ||||
| Search and Q&A | ||||
| Users & permissions | ||||
| Invite / remove members | ||||
| Assign roles, create custom roles | ||||
| Issue and revoke API keys | ||||
| Configure SSO and identity sources | ||||
| Usage & cost | ||||
| View usage and cost reports | ||||
| Set budgets and quotas | ||||
| Export usage and cost reports | ||||
| Audit | ||||
| View audit logs | ||||
| Export audit logs | ||||
| Set log retention and external forwarding | ||||
- Allowed
- Partial: limited to their own teams / projects, or their own data
- Not allowed
Use the accounts you already have. Onboarding and offboarding just happen.
Connect your identity provider and employees sign in to TokenHub with their work accounts. Org structure and membership sync on a schedule, so when someone leaves, their access and any API keys they own are shut off too.
- Works with standard protocols and popular workplace platforms
- Maps departments or groups to teams and default roles automatically
- Require SSO and turn off password sign-in
- Run multiple identity sources side by side for parent companies and subsidiaries
Standard protocols
Workplace platforms
Every key has clear boundaries
Teams only ever hold keys issued by TokenHub. Your raw provider keys stay encrypted inside the gateway you deploy and are never handed out. Each key is bound to a project, a set of models and an access policy.
Per-project keys
Issue separate keys per project and environment (prod / test), with usage attributed automatically.
Expiration
Set an expiry date and get reminded before it hits. No more “master keys” that live forever.
IP allowlists
Only accept calls from specific egress IPs or ranges, so a leaked key is useless outside your network.
Rate limits & quotas
Cap RPM, TPM and monthly quota so a runaway script can't blow through your budget. For per-employee keys, see Employee Quotas.
One-click revoke
Spot something odd? Revoke the key and it stops working within seconds. The action lands in the audit log.
-
support-bot-prod
th-prod-••••-7f3aExpires 2027-03-31Active -
code-assistant-team
th-dev-••••-c21eExpires in 90 daysActive -
knowledge-qa-test
th-test-••••-09b4Expires in 7 daysExpiring soon -
contractor-temp
th-tmp-••••-5d10Revoked Oct 8Revoked
Every call leaves a trail
Who called which model, when, and what it cost. Model calls and admin actions are all logged, searchable and exportable.
| Time | Member | Project | Event | Model | Tokens | Provider cost | Result |
|---|---|---|---|---|---|---|---|
| 10-14 10:42:18 | Chen | Code Assistant | Model call | deepseek-chat | 2,184 | $0.0008 | Success |
| 10-14 10:41:55 | Support bot | Support Bot | Model call | qwen-plus | 1,026 | $0.0003 | Success |
| 10-14 10:39:07 | Lin | Knowledge Q&A | KB search + call | local/qwen2.5-14b | 3,410 | $0.0000 | Success |
| 10-14 10:36:44 | Zhao | Knowledge Q&A | Model call | qwen-plus | — | — | Blocked |
| 10-14 10:30:12 | Wang | Code Assistant | Revoked API key th-tmp-••••-5d10 | — | — | — | Admin action |
| 10-14 10:12:03 | System admin | — | Set Lin's role to Developer | — | — | — | Admin action |
* UI and data are samples, shown to illustrate how the feature works
Sensitive data stays in. Harmful content stays out.
Requests are scanned and redacted before they reach a model, responses are reviewed before they come back, and everything is encrypted in transit and at rest.
Sensitive terms & PII redaction
Please review the loan application for customer John Smith (phone (555) 013-8000, ID number 110101-1990-0101-XXXX) and send the result to john.smith@example.com.
Please review the loan application for customer [NAME_1] (phone [PHONE_1], ID number [ID_1]) and send the result to [EMAIL_1].
When the model responds, placeholders can be swapped back to the original values before the answer reaches an authorized employee. Redaction rules can be set per project and per model: for example, always redact for cloud models and relax the rules for local models on your internal network. All personal details in this example are fictitious.
Sensitive terms & PII redaction
Built-in detectors for phone numbers, ID numbers, bank cards, emails and more, plus regex and custom word lists. On a match, redact, block or just alert.
Content moderation
Checks both inputs and outputs for policy violations, confidential material and out-of-scope requests. Matching requests are blocked and logged, and third-party moderation services can plug in.
Encryption in transit & at rest
HTTPS / TLS end to end. Provider keys, API keys and sensitive log fields are encrypted at rest, and you decide whether raw requests and responses are stored at all.
Runs on your servers, not ours
TokenHub is software you download and deploy yourself. There's no hosted service. Your data, keys and logs stay on your own servers. Pick the setup that fits your scale and network.
Single server
One machine and a few commands. Great for individuals, small teams and proofs of concept.
- Up and running with Docker / Docker Compose
- Config and data volumes kept separate for easy upgrades and backups
- Connect cloud and local models side by side
Cluster / high availability
Scale out across multiple instances in your own data center or private cloud to handle production traffic.
- Multiple nodes behind a load balancer, no single point of failure
- Runs on Kubernetes and other container platforms
- Plugs into your identity, monitoring and logging stack
Air-gapped deployment
Local models, local knowledge bases and a local vector store. Data never leaves your internal network.
- Fully functional with no internet access
- Offline installer and offline upgrades
- Built for isolated government and enterprise networks
Compliance & certifications Coming soon
We're still compiling our compliance and certification details and will publish them here once confirmed. If your procurement or security review needs a completed questionnaire or a write-up of how the software handles data, get in touch.
Contact us →Access & security questions
Will our provider API keys be exposed to teams or employees?
No. Raw provider keys are configured in the gateway only by a Super Admin or an authorized Admin, and they're stored encrypted. Teams and employees only hold keys issued by TokenHub, which you can revoke or rotate at any time without touching anything on the provider side.
Can it connect to our existing AD / LDAP and org structure?
Yes. TokenHub supports LDAP / AD, OIDC and SAML 2.0, as well as WeCom, DingTalk and Feishu (Lark). Your org structure syncs on a schedule, and departments or groups map automatically to TokenHub teams and default roles.
How granular can permissions get?
Down to “can this member, in this project, call this model or access this knowledge base.” Built-in roles cover the common cases, and custom roles let you combine individual permissions and scope them to an organization, team or project.
How long are audit logs kept? Can we export them or send them to our log platform?
You set the retention period. Logs can be filtered by member, project, model, time and more, exported to CSV, or forwarded via webhooks or log shipping to the logging and security analytics tools you already use.
Does redaction hurt answer quality?
Redaction swaps in meaningful placeholders (like “[NAME_1]”) and keeps the sentence structure intact, so most Q&A and summarization tasks are barely affected. Original values can be restored in the response based on your policy. If a workflow truly needs the raw text, enable redaction for cloud models only and relax it for local models on your internal network.
What do we need to deploy TokenHub?
The gateway runs on an ordinary Linux server. A single server only needs Docker, and cluster and air-gapped setups are supported too. You only need GPUs if you plan to run local LLMs. Exact sizing depends on your concurrency and model choice. See the download page for installation steps, and reach out through our commercial licensing inquiry to plan a commercial deployment.
Does TokenHub ever see our business data?
No. TokenHub is software you download and run yourself. There's no hosted service and no account to sign up for. Requests, knowledge bases, logs and keys all live on your own servers. When you call a cloud model, the request goes straight from your server to the provider you configured, and you pay that provider directly.
Let the gateway handle governance. You focus on the business.
Free for personal and non-commercial use. For commercial deployments, contact us about a commercial license and we'll help you plan around your identity setup, compliance needs and infrastructure.