Every page of the local control-plane UI, captured live from the running app.
| Application | http://localhost:7777 (hash-routed SPA, single port) |
|---|---|
| Repository | C:\Users\archi\Desktop\Testing\Code\Local-Claw-Agent\agent-main |
| Data root | C:\Users\archi\.prometheus (state.db, MEMORY.md, SOUL.md, skills/) |
| Captured | 6 October 2026, headless Chrome 145, 1920 px viewport, one capture per route |
| Coverage | 29 screenshots — 15 nav pages plus 14 sub-tabs, across 5 groups |
http://localhost:7777/#overview
The app's home screen as first shipped: live counters, the retrieval-gate split, the clickable architecture diagram, and the chat dock (browser chrome included).
CLI --
http://localhost:7777/#overview
Dashboard: six live counters (spend, avg turn, turns, tool calls, facts, events), the 100%/0% retrieval-gate split bar, and the clickable Harness / Loop / Memory / LLM-Ops architecture diagram.
http://localhost:7777/#gateway
Cross-channel inbox — web, phone (Telegram) and voice all answered by the same brain; every conversation is listed with channel badge, timestamp and reply.
The agent turn itself: user message, the GATE · SKIP decision with its reason, the reply, and the per-turn latency and cost.
Structure around the loop — live DAG topologies for the triage and gather workflows, drawn from the engine's own describe() so the picture cannot drift from the code, plus a one-click run.
The 8 capabilities the agent can call, grouped into flagship tasks, web search, self-management and experimental, each with the exact description the model reads.
http://localhost:7777/#tools/results
What tool calls actually wrote: calendar events from create_event (also mirrored to calendar.ics) and drafted messages sitting in the outbox.
http://localhost:7777/#tools/mcp
Model Context Protocol status and a 30-second connect recipe — install the extra, write .mcp.json, restart — to borrow tools from any external server.
http://localhost:7777/#database
The raw persistence layer: the state.db path and size on disk, per-table row counts with what each table holds, and the 14 FTS5 tables behind keyword top-k retrieval.
http://localhost:7777/#database/facts
Raw SQLite tables with column names as sticky headers; this is the uncurated altitude that the Memory tab renders from.
http://localhost:7777/#database/chat_log
Every message tagged by session_id — the table consolidation reads from to distil memories.
http://localhost:7777/#database/query
A read-only SQL console over state.db with three runnable example queries; the file is opened read-only so nothing typed here can change your data.
http://localhost:7777/#observability
What each turn did, cost and remembered — TODAY / 7 DAYS / ALL window switcher, four metric cards, and an expandable per-turn trace list.
http://localhost:7777/#observability/tools
Tool-call distribution across the window, split by origin (agent, Waku Memory, local) with the error count.
http://localhost:7777/#observability/memory
Retrieval counts versus writes kept — how often memory was consulted and how many facts survived the write pass.
http://localhost:7777/#observability/spend
The spend ledger: total, charged versus estimated dollars, per-model and per-day token usage priced at list price, sourced from usage.jsonl.
Five checks grade every turn (spend claim, one report, grounded numbers, errors handled, under budget) and the release gate reports 2,091 deterministic tests in 156 files plus 3 AI-judge suites.
The three pillars with live counts — Semantic, Episodic, Procedural — the retrieval-gate bar, and the on-disk files: state.db, MEMORY.md, SOUL.md and skills/.
http://localhost:7777/#memory/semantic
Durable, distilled facts about you and your people, each one editable and deletable straight from the UI.
http://localhost:7777/#memory/episodic
One dated summary per consolidation run — the always-small trail of what a session was about.
http://localhost:7777/#memory/skills
Procedural memory: 7 SKILL.md files loaded only when a message matches, each editable inline and saved back to its file on disk.
http://localhost:7777/#memory/soul
The persona file — the durable rules about how the agent should behave for this user.
http://localhost:7777/#memory/consolidation
Every 6 exchanges a cheap model reads the unconsolidated chat_log and distils it into durable facts plus one episode; shows 2 messages queued against a 12-message trigger threshold.
http://localhost:7777/#compare/models
One message, every brain at once on the same harness with isolated homes — graded by a referee model, with per-model gate latency, cost and tools as receipts.
http://localhost:7777/#compare/memory
One brain, five stores told the same 5 facts then quizzed with 4 questions, showing right versus wrong answers per store — including a deliberately corrected fact.
http://localhost:7777/#judgment
The same 15 cases (NOUL, CHOICE, SCORE) sent to 7 judge models at once, so you can see which judge is right on your own domain instead of guessing.
Provider card grid for 12 providers with logo, enabled / unconfigured status, and a per-provider EDIT modal for keys.
http://localhost:7777/#connections
Channels (Telegram, Discord, WhatsApp), productivity (Google and Apple Calendar) and memory backends (Notion, Mem0, Zep, LangMem, Supabase), each with a Configure action.
http://localhost:7777/#settings
Runtime switches for how a turn runs: sub-agent delegation, and graph-workflow triage that rebuilds the agent in-process with no restart.
