Skip to content

Capabilities

Web search capability using Brave Search API (privacy-focused, free tier available).

How it works: - Research Agent has access to web search tools - Search results include titles, snippets, and URLs - Results are cited for factuality

Web Browsing

Interactive web browsing using Playwright with an accessibility-tree-based page understanding.

Why accessibility trees? - More reliable than XPath/CSS selectors which break on site changes - Works on JavaScript-heavy sites via Playwright's accessibility snapshot API - Captures semantic structure (roles, names, values) for precise interaction - Click targets use role + name references rather than fragile selectors

Browser Agent capabilities: - Navigate to URLs and extract content - Click, type, scroll interactions - Form filling and multi-step workflows - Accessibility tree extraction (interactive, forms, or full modes) - Screenshot capture for vision fallback

Cron Job Scheduling

Recurring task scheduling with APScheduler running within the main process.

Job types: - Tool Jobs: Execute a specific tool with fixed parameters - Prompt Jobs: Send a prompt to the coordinator for complex reasoning

Features: - Cron expressions for flexible scheduling - Enable/disable jobs without deletion - Execution history tracking - Run immediately option

Artifact Store

Durable file storage for generated artifacts (HTML pages, PDFs, DOCX, images, Python outputs) with public/private visibility, signed temporary links, and an optional passphrase gate.

End-to-end flow:

User: "Create a sales report and save it so I can share it later."

1. Coordinator → create_html(...)           # generates /tmp/…/report.html
2. Coordinator → store_artifact(            # persists to data/artifacts/<uuid>/report.html
       source_path="/tmp/…/report.html",
       title="Sales Report",
       make_public=False                    # private by default
   )
3. Assistant responds with the artifact ID and a link to the Artifacts tab.

Later, in the Artifacts tab:
4. User clicks "Make public"               # permanent link becomes shareable
   — or —
   User clicks "Copy link" while private   # mints a 300s signed temp link
   — or —
   User clicks "Set passphrase"            # adds a gate; visitors see a prompt

How it works: - The coordinator agent calls store_artifact with a source_path (e.g. a file path returned by create_html or python_execute) to persist it into data/artifacts/<uuid>/<filename> - Artifacts survive the nightly temp-dir cleanup because they live under data/, not tmp/ - Each artifact can be individually deleted by the user from the Artifacts tab

Visibility and sharing: - Private (default): only accessible via a 300s signed temporary link (?token=<HS256 JWT>) - Public: gets a stable permanent link at /artifact/{id} — no token required

Access check on every view request:

GET /artifact/{id}[?token=...]
        │
        ▼
  Is artifact public?
  ├─ Yes ──────────────────────────────────────────┐
  └─ No → valid ?token= present?                   │
           ├─ No  → 401 Unauthorized               │
           └─ Yes ──────────────────────────────── ┤
                                                   ▼
                                     Is passphrase gate set?
                                     ├─ No  → serve file
                                     └─ Yes → valid unlock cookie?
                                              ├─ Yes → serve file
                                              └─ No  → show gate page
                                                        (visitor enters passphrase,
                                                         POST /unlock sets cookie,
                                                         page reloads → serve file)

Passphrase gate (optional per artifact): - Owner sets a passphrase via the Artifacts tab; it is hashed with PBKDF2-SHA256 (200k iterations) and never stored in plaintext - Unauthenticated visitors see a branded gate page prompting for the passphrase - On success an httpOnly SameSite=Lax cookie (oa_artifact_{id}, 1h TTL) is set; subsequent views skip the prompt - Owner can change or remove the passphrase at any time; the API only exposes has_secret: bool

Artifacts tab (/artifacts): - Table view: Name, Type, Size, Visibility badge, Created date - Per-row actions: View, Copy link (permanent or 5-min temp), Make public/private, Set/Change/Remove passphrase, Delete

Coordinator tools:

Tool Purpose
store_artifact Persist a file (source_path) into the artifact store; returns artifact ID and link
search_artifacts Find previously stored artifacts by filename/title using a case-insensitive regex

Both are system tools assigned to the coordinator by default (always available, no integration toggle needed). Visibility, passphrase, and deletion are managed by the user via the Artifacts tab.

Future Task Scheduling

One-time task scheduling for future execution.

Use cases: - Reminders ("Remind me tomorrow at 3pm") - Scheduled actions ("Send this email Monday morning") - Deferred tasks ("Check my calendar tomorrow at 8am")

Monitoring UI

Web-based monitoring dashboard with:

System tab: - Health checks (database, LLM API, disk space) - Performance metrics - System logs viewer - Service connection status

Jobs tab: - Cron job management with enable/disable - Future task management - Execution history with status and duration

Single search interface across all integrated data sources.

Concept: - Keyword search using each service's native API - Semantic search using local embeddings (cosine similarity threshold: 0.3) - Results merged with Reciprocal Rank Fusion (k=60) - Automatic index updates (opportunistic + on-demand rebuild)

Tools: - unified_search - Search across all sources with hybrid keyword + semantic matching - reindex_search - Rebuild the semantic index from connected sources

Parameters: - query: Search query string (required) - sources: Filter by source (optional) — notion, gmail, outlook_email, outlook_files, onenote, nextcloud, memory - search_type: hybrid (default), keyword (API-only), or semantic (embedding-only) - limit: Maximum results per source (default: 10, max: 50)

Sources: - Notion — Pages and databases via Notion API - Gmail — Emails via Google Gmail API - Outlook Email — Emails via Microsoft Graph API - Outlook Files — OneDrive files via Microsoft Graph API (text extraction for .txt, .md, .docx, .pdf, etc.) - OneNote — Pages via Microsoft Graph API - Nextcloud — Files via WebDAV (text extraction for common formats) - Memory — Assistant's long-term facts indexed via system_index_memory_facts

Index Behavior: - Hybrid/semantic searches trigger automatic background reindex if index is empty - Keyword search results are opportunistically embedded in the background after each search - Semantic search requires LLM provider with embedding support