Part of the ContextMemory docs. Back to README.
The Admin console (ContextMemory.Admin.Web, default http://localhost:5200) is a Blazor Server app that talks to the ContextMemory API over HTTP. UI copy is English; field help sits under each control.
Screenshots (from a live Admin at :5200):
| Dashboard |  |
| LLM backends (per tenant) |  |
| Agentic (sandbox / MCP / HITL knobs) |  |
| Playground |  |
| Skills & guardrails |  |
| Who uses it | Operators configuring tenants (LLM, memory, agentic, MCP, keys) |
| Auth | Master Key (ContextMemory:MasterKey) for admin APIs — not the app cm_live_… key |
| Where settings live | Browser local storage (overrides host defaults from Admin__* / Compose env) |
| Image | ghcr.io/kortexio/contextmemory-admin |
| Menu | Route | Purpose |
|---|---|---|
| Applications | / |
List tenants, open details / config |
| New application | /apps/new |
Register a tenant and mint an API key |
| Chat Lab | /chat |
Interactive chat against an app (memory + agentic) |
| Skills | /skills |
Shared skills catalog (+ view guardrail packs) |
| Settings | /settings |
API base URL, Master Key, health |
Per-app pages (from Applications):
| Page | Route | Purpose |
|---|---|---|
| Details | /apps/{appId} |
Telemetry, wiki/web-search stats, view/rotate API key |
| Config | /apps/{appId}/config |
LLM, wiki, Global Wiki, web search, rate limits, Agentic Gateway |
- Open Settings.
- Set API base URL:
- Local
dotnet run:http://localhost:5100 - Docker Compose / GHCR Admin container:
http://api:8080(server-side calls stay on the Docker network; the browser still opens Admin onlocalhost:5200)
- Local
- Paste the Master Key (demo:
cm_master_dev_key_change_me). - Click Save, then Test connection.
- Confirm Health: Ollama, Persistence, Apps loaded, optional Database.
Save stores values in this browser. Reset to defaults clears browser overrides and restores host defaults (Admin__DefaultApiBaseUrl, Admin__DefaultMasterKey). A link to public Prometheus metrics (GET /metrics) appears when connected.
- Cards per app: request count, active users, last-turn wiki pages included/total, source badge (
seedvs registered). - Actions: Details, Config, Chat Lab.
- New application / Refresh. Seed app
demo-devappears when the API starts with default config.
New application creates an isolated tenant (own API key, session wiki, optional Global Wiki).
| Field | Notes |
|---|---|
| Application name | Operator label only (not an HTTP header) |
| Domain | Prefix for generated X-App-Id (e.g. helpdesk → helpdesk-…) |
| Default language | BCP-47 for prompts (en-US, pt-PT, …) |
| LLM backend / model / endpoint / API key | Same semantics as Config (below) |
| Persona | Optional base system prompt |
After create, copy the API key immediately — it is shown once on this screen. Then open Config, Details, or Chat Lab.
/apps/{appId} shows:
- Requests, errors, active users, average latency
- Prompt/completion tokens, wiki truncated total
- Last-turn wiki memory (chars, pages included/on disk, compaction ok/err)
- Web search totals when used
- API key — show/copy; Generate new API key rotates it
For seed apps (appsettings), rotation lasts until API restart unless you also update ContextMemory:Apps:{appId}:ApiKey.
Use the app key in Chat Lab and client apps — never the Master Key.
/apps/{appId}/config patches runtime config (PATCH /admin/apps/{id}/config). Changes apply to new requests after Save changes.
| Control | Meaning |
|---|---|
| Default language | Locale in prompts |
| LLM model | Model id on the backend |
| Backend | ollama (default /v1), vllm, lmstudio, openai, openai-compatible, custom, ollama-native |
| Endpoint / API key | Optional per-app overrides; empty = host defaults |
| Max history messages | Recent turns kept in the prompt (wiki is separate) |
| Streaming enabled | Allow streaming on legacy /api/chat |
| Enable model thinking | llmThinkEnabled → reasoning_effort / think (off by default) |
Generation defaults (llmOptions) |
Tenant defaults merged into chat/generate unless the client overrides |
| → temperature / top_p / top_k | Sampling |
| → num_ctx | Ollama context window. With backend ollama, gateway auto-uses ollama-native so num_ctx applies (/v1 ignores it). |
| → num_predict | Max tokens to generate (max_tokens on OpenAI /v1) |
| → repeat_penalty / seed / stop / tfs_z / mirostat | Advanced Ollama options |
| → keep_alive / format | Ollama top-level fields. format=json is ignored on agentic turns with tools. |
Persisted in app_profiles.ConfigJson as llmOptions. Request body wins over tenant defaults for any field set.
Prefer clients calling POST /v1/chat/completions.
Max wiki context chars, max context tokens (agent compaction), compaction threshold (bytes), compaction min pages.
- Enable Global Wiki → digests inject + tools
wiki_search/wiki_grepfor documents under/apps/{id}/wiki/… - Max Global Wiki tool chars — budget per wiki tool call (
0= service default)
Ingest/digests are API-side; the Admin UI toggles availability and budget. Retrieval is dynamic context discovery (digests + FTS/tools), not vector RAG — see architecture-and-features.md.
Enable, mode (heuristic / llm / always / off), provider (tavily / brave), max results, max ephemeral context chars, persist-to-wiki, telemetry logging. Provider API keys live on the API host (ContextMemory:WebSearch), not per app.
App RPM, per-user RPM, TPM, agentic request weight, agentic tokens per iteration.
| Area | What you configure |
|---|---|
| Enable agentic loop | Multi-step tools on the same chat request |
| Prompt profile | auto / ollama / openai / claude / qwen / composer |
| Harness mode | auto / weak / strong — weak inlines evidence + aggressive repair; strong prefers lazy skills |
| Loop guardrails | Validation mode, network egress, max iterations, loop timeout, min answer length, confirmation keywords, allowed hosts, expected regexes, require exit 0, human review on max iterations |
| Skills & guardrail packs | Per-app checkboxes from the shared catalog (see Skills page). Omit selection → catalog defaults. Skills may be always_on / requestable rules. |
| Execution tools | self-hosted-sandbox → sandbox endpoint (Compose: http://sandbox-runtime:8080) or aca-session → ACA pool URL; runtimes shell/python/node/(custom); allowEgress. Sandbox output is always archived as a session artifact. |
| MCP integrations | http or stdio; name, URL/command+args, auth mode, credential ref, OAuth fields, allow/deny tool lists, timeout, enabled, allowEgress; max MCP tools per turn; schemas are lazy (tool_describe) |
Example stdio MCP is shown in a collapsible on the Config page. After editing MCP servers, rebuild the tool catalog via API if needed: POST /apps/{appId}/mcp/catalog/rebuild (Master Key or app auth as configured).
Markdown fields: base persona, business rules, format rules, wiki schema.
Two levels:
| Level | Where | Behaviour |
|---|---|---|
| Platform | /skills |
Seeded defaults shared by all apps. IsDefaultEnabled = always active for every tenant. Full CRUD for skills and guardrails (system items: edit OK, delete blocked). |
| Per app | /apps/{id}/policies |
App-owned inventory only. Full CRUD + enable toggles. Additive to platform defaults (cannot disable platform packs from the app). |
| Action | Notes |
|---|---|
| New / Edit skill | Id (slug, create-only), name, category, description, markdown editor (Write/Preview), sort order, default-on (platform) or enabled (app) |
| System skills / guardrails | Prompt/config editable; cannot delete |
| Download / Import | .skill.json / .guardrail.json (optional replace) |
| Runtime | Union of platform default-on + app enabled packs |
Legacy agentic.policyPacks selection is ignored.
Interactive tester for a tenant. Uses the app API key + X-App-Id (Master Key only for “Load app” metadata via admin APIs).
Connection panel
| Field | Role |
|---|---|
| App ID / User ID / Session ID | Headers; empty session → API generates |
| App API key | Bearer for chat |
| Model | Body model (match app config) |
| Chat / Generate | Legacy POST /api/chat or /api/generate |
| Streaming | NDJSON token stream |
| Show raw JSON | Debug / Inspector panel |
| Save locally / Load app / New session | Persist lab settings, pull app summary, reset session id |
Ollama options — optional per-request system prompt, temperature, top_p, top_k, num_ctx, repeat_penalty, num_predict, keep_alive, format.
Main pane
- Conversation bubbles with agentic timeline (LLM request, tool start/end, Compacting, Subagent start/complete, validation, HITL)
- HITL banner with
confirm/[CONFIRM:id]and copy token - Stop / Regenerate / Clear UI
Side panels (Chat mode)
| Panel | Content |
|---|---|
| Todos | Items from todo_write tool progress |
| Artifacts | artifactIds seen in the turn (sandbox, wiki, subagent, history compaction) |
| Wiki / discovery | Compaction + wiki tool steps for quick grounding checks |
| Inspector | Raw request/response when “Show raw JSON” is on |
Typical smoke path: Settings connected → open Chat Lab → demo-dev + demo key → send a message → same Session ID on the next turn to verify memory / rolling summary.
| Secret | Used for | Where |
|---|---|---|
Master Key (cm_master_…) |
List apps, patch config, rotate keys, skills catalog | Settings |
App API key (cm_live_…) |
Chat, wiki ingest, app-scoped MCP ops | Details / Chat Lab / clients |
| LLM API key | Downstream OpenAI-compatible backends | App Config (optional) |
| MCP / OAuth secrets | Integration tools | Config credential ref / OAuth fields (prefer credential store over inline tokens) |
| Symptom | Fix |
|---|---|
| Not connected / cannot list apps | Settings → correct API URL + Master Key → Test connection |
| 401 on admin calls | Master Key ≠ ContextMemory:MasterKey |
| Chat Lab 401 | Using Master Key as app key — use cm_live_… + X-App-Id |
| Compose Admin cannot reach API | From Admin container use http://api:8080, not localhost:5100 |
| Sidebar dead after upgrade | Hard refresh (Blazor admin-shell.js, not legacy AdminLTE jQuery) |
| Seed key rotation “lost” on restart | Persist new key in appsettings / env for seed apps |