AI and engine

The AI & engine group holds the model, decision and code runtimes: which providers your agents run on, the decision layer, code intelligence and the on-device runtimes.

Label	Route	Nav needs	Backend rule
Providers	/providers	any member	POST /api/providers, DELETE /api/providers/{name}, the ElevenLabs, Sarvam and SDLC credential writes and OCR credential need PERM_MANAGE_INTEGRATIONS
Decision providers	/capabilities/decision	any member	credential, policy and playground writes need PERM_MANAGE_INTEGRATIONS
Code Intel	/code-intel	any member, when the build has a Code Intel address	/api/code-intel/** needs the tenant-admin or agent-builder role
Browser models	/edge	any member	the manifest is public; relay routes need PERM_MANAGE_CHANNELS
Local code	/code	any member (page is also public)	none; no API call is made

Providers (/providers)

Connect and activate the LLM providers agents run on. This is the first stop for a new workspace: an agent cannot be created without at least one active provider, and the Agents page links back here when there is none. The page is a section titled with the active and total provider counts, one card per provider (Anthropic, OpenAI, Google Gemini, Mistral, Groq, OpenRouter, Ollama and others), each with a key form, a start hint and credential help. The same screen also holds the keys for adjacent services: ElevenLabs (with a voice list), Sarvam, and the SDLC credential with the code-execution template and a "run code" sandbox test. Keys are encrypted at rest with APP_ENCRYPTION_KEY (getting started).

On a personal install the page also has Set up a local model (/providers/local-model), a guided setup for a free model that runs on this computer through Ollama: it lists the models that fit the machine's memory, downloads the one you pick and makes agents use it. Its Speed and behaviour card sets how long the model thinks before answering, how long it stays loaded and its context window. A personal install with no provider yet sees a first-run card for this on the page. On lite and full installs the models the platform provides are set up by the operator, so these entries are not shown. See local models.

A link on the page opens AI Model Control (/ai-models), where the workspace picks a preferred provider and model per AI-touching feature instead of the platform default; changing it needs PERM_MANAGE_TENANT_SETTINGS (GET/PUT/DELETE /api/tenant/ai-model-preferences/{featureKey}).

API: GET /api/providers, POST /api/providers, DELETE /api/providers/{name}, /api/elevenlabs/credential, /api/elevenlabs/voices, /api/sarvam/credential, /api/sdlc/credential, /api/sdlc/sandbox/run-code, PUT /api/tenant/settings.

Decision providers (/capabilities/decision)

A decision provider is not a chat provider: it never generates text. It answers typed questions (a yes or no probability, a choice over options, a score over levels) and powers the decision layer: cost routing on chat turns and Quick Task agent selection. The page lets you see whose key the active decision provider is running on for your traffic (yours, the platform's, or none), bring your own key for any provider the catalog lists (which then takes precedence for the workspace), and experiment against the resolved key in a playground. /capabilities has no page of its own and redirects here; the old /jev URL does too. API: GET /api/capabilities, /api/capabilities/{cap}/credential, /api/capabilities/{cap}/policy, /api/capabilities/decision/status, /api/capabilities/decision/playground.

Code Intel (/code-intel)

The console for the code-intelligence service (a sibling backend service, reached with the app's token at the configured NEXT_PUBLIC_CODE_INTEL_URL). Tabs: Repositories (add a repository by git URL or path), Knowledge, Search (results use the shared pager), Graph, Execute and Governance. Per-agent repository access is set on the agent (PATCH /api/agents/{id}/code-intel-repo-access).

Browser models (/edge)

Sidebar label: Browser models (the page is the VegaDuta Edge console). It holds: a device verdict and a mode master toggle, the on-device model marketplace (real downloads with visible progress), installed-model management and a privacy band. The page itself explains when this device cannot run models locally. The catalog comes from the public GET /api/edge/manifest (cached for five minutes); operators edit that catalog in Platform admin. The relay used by the Edge relay feature is documented in Edge relay. For free local models see local-models.md.

Local code (/code)

A browser-local code assistant: open a folder from this device, index it in the tab, ask questions and apply suggested edits, with nothing leaving the browser. The route makes no API call and carries no session, which is why it is also a public route. The sidebar label comes from the localCode message namespace, not nav. Needs a browser with the File System Access API; otherwise the page says so.