T-Llama combines a local Onramp, private gateway, Entra-backed control plane, budget enforcement, guardrails, and analytics so company AI traffic is usable, governed, and attributable.
Every layer of the mesh, from the endpoint to evidence.
T-Llama Onramp
The Onramp is the local process that tools talk to. It exposes compatible local AI APIs, authenticates local callers, applies endpoint guardrails, and forwards governed traffic to the private gateway.
Local OpenAI-compatible /v1 API
Anthropic-compatible local API for Claude Code
OpenAI Responses local API for Codex
Local Session Credential authentication
Metadata-only local usage events
Local Session Credentials
Local Session Credentials work like API keys for local tools, but they are not provider keys and they do not authorize direct gateway access.
Endpoint-only by design
Tool-bound attribution by default
Immutable scope — mint a new credential to change project/tool/budget
Token values shown once; stored metadata and hashes only
OS-protected storage on macOS
Native launch integrations
Launch supported coding tools through T-Llama without hand-copying base URLs or keys.
tllama launch claude
tllama launch codex
Reuse-or-mint local credential flow
Project attribution with --project
Optional model pinning, centrally validated
Restore flow for T-Llama-owned Codex profile
Private gateway
The gateway is the central enforcement plane for identity, model policy, budgets, provider routing, and key custody.
tllama-gate fronting Bifrost
Entra JWT validation
Identity-to-virtual-key mapping
Provider keys stay gateway-side
Inference-only route surface
Gateway reachability diagnostic
Fail-closed when the private route is unreachable
Diagnostic echo and contract evidence
T-Llama includes deterministic contract paths so private-gateway behavior can be tested without live provider spend.
Gateway-side tllama-echo diagnostic model
Phase 1A contract report in CI
Phase 2A launch conformance
Phase 2C contract report for management, enforcement, and analytics
Metadata-only artifacts
Guardrails
The endpoint blocks high-confidence secrets before requests leave the user's machine.
Pre-egress deterministic scan
Secret fingerprints instead of raw values
Guardrail failure class distinct from auth and gateway errors
Applies even to diagnostic requests
Management API
Every admin capability is an API first. The portal and future integrations use the same governed endpoints.
/control/v1 Management API
Entra role-gated admin writes
Human delegated tokens and application-token path
Projects, members, allow-lists, budgets, personal allowance
Me-scoped budget reads for user budget displays
Admin Portal
The Admin Portal gives administrators a visual control plane without creating a portal-only backend.
Entra-protected sign-in
Project creation and archive
Member management by UPN
Per-project model allow-lists
Project budget update
Tenant default personal allowance and user overrides
Analytics page
Budget enforcement and snapshots
T-Llama separates authoritative enforcement from user-friendly display. The gateway evaluates live budget state on every request; clients receive display-only snapshots and piggybacked state.
Gateway budgets for projects and personal usage
Local budgets can be stricter, not looser
Smallest applicable budget wins
Budget state header for active traffic
CLI budget display with spent and remaining amounts
Usage analytics
Metadata-first analytics show who used which project, tool, and model, how many requests ran, what they cost, and where denials happened.
Request accounting rows written by the gateway
Analytics API endpoints
Portal headline tiles
Spend over time
Group by project, user, model, or tool
Cost Confidence badges
Budget denial panel
Onramp Dashboard and tray
The endpoint UI track makes T-Llama visible outside the terminal: sign-in, budget visibility, credential minting, usage views, and always-on Onramp readiness.
Tokened loopback dashboard shell shipped
Browser Entra sign-in flow shipped
Budget snapshot foundation shipped
Credential mint/revoke/usage pages in flight
Tray app with login startup, restart, and status planned
How traffic flows.
Tools talk to the local Onramp. The Onramp authenticates local callers, applies pre-egress checks, and forwards over the private path. The gateway validates identity, rechecks central policy, enforces budgets, routes to providers, and writes metadata-first evidence.