Admin Console Surface

This page is the technical inventory for the Admin Console surface. Start with Use the Admin Console for the operator workflow; use this page when you need screen coverage and endpoint mapping.

The Admin Console is the browser control plane for a running awaken-server: configure providers and models, edit prompts and tool descriptions, assign MCP tools, tune reminders and deferred-tool policy, preview a draft, then publish the next registry snapshot. Starting the server + console is covered in Use the Admin Console.

Screenshots

The screenshots show representative console states. A running console reads values from your backend APIs; if a subsystem is not wired, the corresponding surface shows a disabled or unavailable notice.

Admin dashboard showing live workload, agent activity, recent audit events, provider and MCP health, and current scope metadata. — Dashboard: live workload, health, audit activity, and read-only scope.

Agent editor with model selection, system prompt, tools, plugins, delegates, history, save controls, and preview chat. — Agent editor: prompts, tools, plugins, delegates, history, and draft preview.

Agents list with filters, model and plugin metadata, and runtime inference statistics. — Agents list: filters, model/plugin metadata, and runtime stats.

Infrastructure and resources

Providers screen listing Anthropic, Vertex, and local Ollama providers with adapter, base URL, API key status, and test/edit/delete actions. — Providers: upstream adapter, credential status, and connection test entry point.

Models screen listing stable model ids, provider ids, upstream model names, modalities, context windows, and actions. — Models: stable runtime ids mapped to provider-backed upstream models.

MCP Servers screen listing filesystem and Linear servers with transport, live status, restart policy, tool count, and actions. — MCP Servers: transport config, live status, and restart actions.

MCP server detail screen showing filesystem live status, restart button, command, exposed tools, prompts, and resources. — MCP detail: handshake, restart, and exposed inventory.

Tools screen listing built-in and MCP tools with source badges, descriptions, and edit actions. — Tools: discovered catalog and model-facing descriptions.

Skills screen listing reusable skill instructions with invocation mode, context mode, allowed tools, and source paths. — Skills: reusable instructions and allowed tool context.

Observe and evaluate

Audit Log screen showing recent update, publish, restart, and create events with actor, resource, timestamp, and change summary. — Audit Log: config history, restore context, and operator attribution.

Datasets screen under Observe with dataset list, fixture counts, and create/delete actions. — Datasets: captured trace fixtures grouped for replay.

Eval Runs screen listing completed eval jobs with dataset, mode, status, fixture count, and pass/fail summary. — Eval Runs: execution records for dataset replays.

Eval run detail screen with pass rate, failure count, and per-fixture report rows. — Eval run detail: per-fixture pass/fail output.

Eval Reports screen with upload cards for a new NDJSON report and an optional baseline report. — Eval Reports: offline NDJSON review and baseline comparison.

Assistant and shortcuts

Admin Assistant panel with capability description, suggested agent chips, and an input for describing an agent or asking about config. — Admin Assistant: guided agent creation and config help.

Create A2A server form with server id, base URL, timeout, optional target, options JSON, and bearer token controls. — A2A server setup: remote agent-card discovery configuration.

Command palette overlay with search input and quick links to agents, providers, models, tools, and observe screens. — Command palette: keyboard navigation across console surfaces.

Screens and endpoints

Each screen is a thin client over admin REST routes (all behind the admin bearer token). For request/response shapes see the HTTP API.

Screen	Reads / writes
Dashboard	`GET /v1/capabilities`, `/v1/system/info`, `/v1/audit-log`, `/v1/runs/summary`, runtime stats
Agents (list + editor)	`GET/POST/PUT /v1/config/agents`, validate `POST /v1/config/agents/validate`, draft preview `POST /v1/ai-sdk/agent-previews/runs`, restore `POST /v1/config/agents/:id/restore`, stats `GET /v1/agents/:id/runtime-stats`
Providers	`GET/POST/PUT/DELETE /v1/config/providers`, test `POST /v1/providers/:id/test`
Models	`GET/POST/PUT/DELETE /v1/config/models`
MCP Servers	`…/config/mcp-servers`, restart `POST /v1/mcp-servers/:id/restart`, inventory `GET /v1/mcp-servers/:id/inventory`
A2A Servers	`…/config/a2a-servers`, status `GET /v1/a2a-servers/:id/status`
Skills / Tools	`GET /v1/config/skills` (read-only), tool catalog from `/v1/capabilities`
Admin Assistant	run `POST /v1/admin/assistant/runs`, policy `GET/PUT /v1/admin/assistant/config`
Audit Log	`GET /v1/audit-log`
Datasets / Eval Runs	`…/eval/datasets` (+ `/:id/items`), `…/eval/runs` (+ `/:id`)
Eval Reports	offline NDJSON upload (no backend call)

The provider→model→agent setup workflow, the editor tabs, and wiring a saved agent to a frontend are covered in the how-tos: Use the Admin Console, Configure Agent Behavior, and AI SDK frontend integration.

Provider credentials and MCP credentials are intentionally separate. Providers feed model execution; MCP server credentials belong to that MCP transport (env for stdio, URL/config for HTTP), and agent access is controlled through tool selection plus optional permission rules. The Admin Assistant unlocks only after the first provider-backed model is configured; its tools are server-locked and do not appear in the normal tool registry.

Operate, Trace, And Evaluate

Dashboard shows live workload, provider/MCP health, recent audit events, optional runtime stats, and read-only scope_id.
Recent runs on a saved agent opens persisted traces when trace routes are enabled.
Datasets capture trace fixtures for evaluation.
Eval Runs execute datasets against configured agents and models.
Eval Reports view NDJSON reports and baseline diffs in the browser.

Trace and eval payloads may contain prompts, tool arguments, and model replies. Protect the admin bearer token and route access accordingly.

Version History And Pinning

Every config save records metadata and, when audit logging is wired, appears in the Audit Log. Agent History lets you inspect diffs and restore a previous snapshot back into the editing store.

Restore is intentionally a review step: after restoring, save/publish the resource again when that restored payload should become active for new runs. When the server is wired with a versioned registry store, published runtime registry snapshots are immutable and durable runs carry a resolution_id so resume and replay can reselect the same graph.

REST-only Features (No Console UI Yet)

The console focuses on configuration. Some surfaces are intentionally REST-only today — drive them from curl or your own scripts with the same admin bearer token (see the HTTP API for request shapes):

Surface	What	Endpoints
Threads & runs	list / create / cancel / inspect messages	`/v1/threads`, `/v1/runs`
HITL decisions	resume / cancel a suspended tool call	`POST /v1/runs/:id/decision`
Mailbox	peek / push inter-agent dispatches	mailbox routes
Skill CRUD	the console lists skills but does not edit them	`/v1/config/skills`
Config diagnostics	registry-wide validation report (no screen renders it yet)	`GET /v1/config/diagnostics`

Scope

scope_id is shown as read-only system metadata. The browser does not choose scope directly; the server resolves scope from the trusted HttpScopeProvider for each request. Hosted products should switch tenant/workspace scope in their auth/provider layer and display the resolved value in the console.

Use the Admin Console - operator walkthrough
Configure Agent Behavior - full tuning surface
HTTP API - request and response reference