GUI Tour

The built-in web GUI gives you a full control panel for every part of Soulacy — agents, chat, schedules, memory, knowledge, and operations — without touching a single YAML file.

Open it at:

http://localhost:18789

If pages show 🔒 Authentication required, click the 🔑 button in the sidebar footer and paste the server.api_key from ~/.soulacy/config.yaml.

The sidebar is split into three groups: main (day-to-day work), ops (integrations), and system (observability).

Dashboard

Your at-a-glance health view: gateway status and version, agent counts, and a Live Event Log streaming every runtime event over WebSocket, with filter presets (All / Errors / Tools / LLM / Messages).

The Launch Readiness panel scores every load-bearing capability (providers, secrets, channels, Studio contracts, schedules, deployment, ops alerts, SLOs, release, docs) and shows blocker / warning counts with click-through to the offending page. The schedules_ready first-class check reports total / enabled / delivering / overdue with actionable next-steps — so an agent that saved cleanly but never fires (because e.g. no default outbound bot is configured) shows up here instead of in silence.

Try first: send a chat message in another tab and watch the events appear live.

First Run

The guided setup path for a fresh workspace. It helps you choose a first focus, configure the essentials, and points you toward a starter template or Studio flow that matches what you are trying to build.

Try first: pick the focus closest to your use case, then follow the suggested next action.

Studio

The one-stop agent development surface. Describe an agent in plain language, let Studio draft the workflow, inspect the generated plan, edit nodes and connections on the canvas, run dry/live tests, validate integrity, and use self-heal when a real run fails. Studio can author Workflow, ReAct, and Plan-Execute agents, then save them as normal SOUL.yaml agents.

• Generate has a Streamed / Wizard split-button. Streamed (default) runs all five pipeline phases (clarify_intent → choose_strategy → build_graph → validate → repair) in one shot and streams a live transcript below the canvas. Wizard opens the refinement modal with a Wizard-steps breadcrumb so you can inspect and edit intermediate output. Persist the default under the Studio model modal → Generate presentation.
• Runtime intent presets in the model modal: Fast local (small local model, tight timeouts, high loop cap), Reliable local (patient timeouts, the default), Cloud quality (long total budget, applied even for cloud providers).
• Contract panel for reasoning agents runs blocker/warning checks on the system prompt, tool allowlist, peer graph, prompt hygiene, step budget, channel delivery, LLM fit, capability scope, persona consistency, and builtin scope — issues surface before Save.
• Debug in Studio on a failed run in Activity pre-fills the failing input into the test bench, loads the structured run trace, and routes any heal result through an Apply this fix / Cancel diff panel; the canvas only changes on confirm.
• Build until it works now surfaces two sections above the attempt log: Needs your input (external blockers Studio can't fix — bad credential, bot not invited, rate limit, invalid destination) and What Studio changed (plain-language rollup of each attempt's edits).

Try first: type "an agent that summarizes RSS articles every morning", generate a draft, run Dry run, then click any node to inspect its inputs and outputs.

Agents

The full agent editor: identity, trigger, system prompt, LLM provider/model (with live model lists), memory scopes, reasoning strategy, brain memory layers, channels, tools, skills, knowledge bases, and peer agents.

• Tool Python file fields have a 📂 Browse picker that lists every script in the tool catalog — pick instead of typing paths.
• Chip-picker fields (channels, skills, knowledge bases, peer agents, built-ins) have a ▾ browse dropdown populated from what actually exists, so typos are impossible.
• Validate runs the same checks as sy agent validate and shows findings inline. Invalid schedule.cron strings are refused at Save with the parser's own explanation (5-field, 6-field-with-seconds, and @descriptor shapes are accepted).
• 💬 Test opens an inline playground; </> API shows copy-paste cURL/Python/JS snippets for the agent.
• Capability audit modal — a save that would escalate an agent's tier (ReadOnly → Active → Privileged) while it has interactive channel bindings pops a blocking modal listing the tier diff, warnings, and affected bindings. Confirming re-submits with an acknowledgement header; cancelling leaves the previous shape in place.

Try first: select an agent, click Validate, then 💬 Test and send it a message.

Templates

A catalog of ready-made agents: four agentic workflows (Meeting Minutes & Action Items, Smart Inbox Triage, Competitor & Market Monitor, Document Compliance Auditor) plus starters. Install opens a guided checklist with readiness checks, a mock test, and optional scheduled-output routing. See Templates.

Try first: install Meeting Minutes & Action Items and paste a transcript into Chat.

Chat

The Chat Tester: pick an agent, send messages, and watch the Thinking panel narrate every LLM call, tool call, and reasoning step live. Each reply carries a per-turn token/cost delta. Hover any bubble to reveal ⑂ Fork and branch the conversation from that point; the 🎤 button starts a realtime voice session. See Chat and Voice.

Try first: ask anything, then expand Thinking under the reply.

Brain Mem

The Brain Memory explorer for long-term agent memory: an Episodic timeline (search, write, clear), a Procedural rulebook editor with markdown preview — including version History, line-level Diff vs current, one-click Roll back, and a 🔒 Lock toggle that freezes the rules — and a Context Preview tab showing exactly what gets injected into the system prompt. See Memory.

Try first: open the Procedural tab and click ⧗ History.

Knowledge

Create and manage RAG knowledge bases: upload .md/.txt/.pdf/.docx files or paste text, inspect chunk counts, and run a Test search against any KB before wiring it to an agent. See Knowledge bases.

Try first: create a KB, drop in one PDF, and test-search a phrase from it.

Workboard

A kanban board (Todo → Running → Needs Review → Done → Failed) where each task can be assigned to an agent and run with ▶. Tasks carry owner, priority, tags, and due dates; the editor shows run history, downloadable artifacts, and comments/review notes. See Workboard.

Try first: create a task, assign an agent, click ▶ Run.

Channels (ops)

Status and configuration for every channel adapter — Telegram, Discord, Slack, WhatsApp, WhatsApp Web (with QR pairing), email (SMTP), Microsoft Teams, Google Chat, generic webhooks, and the always-on HTTP channel. Edit credentials, map bots to agents (multi-bot supported), and enable/disable adapters; a restart banner appears when changes need a gateway restart.

• Guided setup cards for Telegram / Slack / Discord / WhatsApp / WhatsApp Web / email / Teams / Google Chat / HTTP — step-by-step with per-field hints keyed off the actual adapter config keys, and a Test delivery button at the bottom.
• Diagnose on any mapping returns a plain-language category and fix: bad_token, bot_not_invited, missing_scope, invalid_destination, forbidden, rate_limited. SMTP-specific classifications (authentication failed, STARTTLS required, relay denied / SPF, recipient rejected, quota, message rejected, TLS handshake) so raw 535 5.7.8 / 550 5.1.1 codes don't bubble up as "unknown".
• Test on the card runs a real send when configured, or a dry-precondition check otherwise; the result renders on the card with the same friendly category/reason/fix.

Try first: open a channel's Edit dialog to see which agent each bot routes to.

Automations (ops)

All cron agents in one table: next run, last run, missed-run policy, live "Running…" status, and countdowns. ▶ Run triggers immediately, 📋 History opens a slide-out of past manual, channel, and scheduled runs with outputs and delivery status, 👁 Watch jumps to the live action log. See Schedules.

• ⟳ auto-replayed chip appears on the Agent cell whenever the gateway missed a scheduled fire and replayed it at startup (within the missed_startup_window); hover for the missed/replayed timestamps and how much late the run was.
• Save-time cron validation — the Edit modal refuses invalid expressions with the parser's own message ("expected exactly 5 fields, found 3") instead of silently succeeding and leaving the "Next run" column blank.

Try first: click 📋 History on any cron agent.

Skills (ops)

Browse installed Agent Skills (SKILL.md instruction packs), read their full instructions and resources, and install new ones — ⚡ From AgenticSkills installs directly from an agenticskills.io URL, and ➕ Skill sources lets you paste any URL (a skills.sh-style directory, a Soulacy package registry, or a Git host), have it reviewed and identified, then add it as a registry source.

Try first: click ➕ Skill sources, paste https://www.skills.sh/, and hit 🔍 Review.

MCP (ops)

Manage MCP servers: + New Server for manual stdio/HTTP specs (command, args, env, headers) with a connection test, or the Glama provisioner to import a server spec from a Glama URL.

Try first: expand an existing server row to see its tools and status.

Plugins (ops)

Install plugins from a git URL, a sha256-checksummed archive, or a local directory. Every install is staged first — you review the requested permissions, credentials, and security findings before approving.

The Agents → Import modal (for agent packages) surfaces the new v2 schema: namespaced ids (owner/name), calendar versioning (YYYY.MM.DD), and an install-time requirements gate that refuses import when a required provider / channel / secret / MCP server is missing, with a red Missing requirements banner and an I understand — import anyway checkbox for local experiments. Legacy v1 packages import with a deprecation warning until the 2027-06-01 cutoff. See Packaging.

Try first: paste a plugin source and click Clone & review (nothing activates until you approve).

Providers (ops)

LLM provider credentials and defaults: add keys for OpenAI, Anthropic, Google, Groq, Mistral, OpenRouter, Together, DeepSeek, Grok, or point at local Ollama; pick default models from live model lists; run connection tests; tune provider-specific options (Ollama context window, Anthropic extended thinking, and more).

Test connection on any card now surfaces a structured Provider Doctor diagnosis when the call fails: friendly reason (bold), concrete fix (secondary line), and a Show raw error toggle for the underlying provider response. Categories cover the shapes the drivers actually return: bad_key, region_blocked, rate_limited, overloaded, model_not_found, context_too_large, quota_exceeded, provider_down, bad_endpoint, network, local_unreachable. See Troubleshooting → LLM providers for the full category → fix table.

Try first: click a provider's test button to verify it answers.

Activity (system)

The per-agent action log from the SQLite action store: every run, LLM call, tool call/result, reasoning step, reply, and error with timestamps and summaries. Filter by type, enable ▶ Watch (2s) for live polling, and see per-run token/cost totals. See Dashboard & Activity.

The Running now strip at the top polls /activity/running every 3s and renders one card per in-flight session — agent name, Xm Ys in flight, silent Ys, last: <event_type>. A session that hasn't emitted anything for more than 5 minutes turns red and shows a per-last-event-type reason + fix ("waiting on the LLM provider for 6m 12s — check Providers for rate-limit/overload"). Click a card to deep-link into that agent/session's event stream.

Try first: select your busiest agent and click ▶ Watch (2s).

Config (system)

Edit the live runtime configuration: log level/format, Python interpreter, tool timeout, max turns, max concurrent sessions, default LLM provider, agent/skill directories, and per-plugin settings. After saving, a banner offers a one-click Restart Gateway.

Try first: check that python_bin points at the interpreter where your tool dependencies are installed.

Logs (system)

The gateway log tail: last 100–5000 lines, free-text filter, severity filter (error/warn/info/debug) with per-level counts, line wrap toggle, and 3-second auto-refresh.

Try first: filter on error after anything misbehaves — it is usually the fastest diagnosis.

Tip

Everything in the GUI is also scriptable: the CLI (sy) and REST API (/api/v1/...) drive the same endpoints the pages use. Run sy doctor if any page shows unexpected errors.