Holaryn Agent User Guide

Holaryn Agent is a proprietary, local-first AI agent from Synergentic, Inc.: its own agent loop,
tools, memory, scheduler, and web dashboard, all running on your machine with no account required.
This guide covers everything from installation to multi-machine peering.

You can read this guide three ways: right here in the repository, inside the app (press F1 or pick Help → User Guide, and click the small ? icons in Settings), or from the command line with /user-guide (add search <query> to search it).

Getting started

Introduction — what Holaryn Agent is and how it thinks about autonomy, events, and privacy.
Installation — the desktop app, authorized private-source installs, extras, and updates.
Self-host with Docker Compose — hardened single-host deployment,
automatic TLS, profiles, diagnostics, updates, encrypted backup, restore, and rollback.
Quick start — from zero to your first completed task.
First-success onboarding — resumable desktop and CLI setup, verification, approval demo, and repair diagnostics.
Living capability matrix — maturity, platform and surface support, prerequisites, limits, privacy, and verification evidence.
Capability Center — inspect effective session authority, denial reasons,
provenance, and safe source actions during the Capability Fabric private preview.
Running the agent — attended sessions, the persistent host, OS services, and the state directory.
CLI reference — every holaryn subcommand and flag.
Repository instructions — target-scoped AGENTS/CLAUDE guidance, precedence, safety, and explainability.
Repository intelligence — revision-scoped definitions, references,
diagnostics, context packs, language-service health, cache controls, and safe fallback behavior.
Automatic code validation — discover tiered project checks, review
repository commands, run bounded validation, and inspect revision-bound evidence.
Isolated coding worktrees — workspace modes, disk use, exact validation
and integration, conflict handling, crash recovery, archive, and safe cleanup.
Visual Studio Code and ACP — create and reattach editor sessions,
bounded context, stale-edit protection, approvals, and troubleshooting.
Context references — attach typed sources with freshness, provenance, trust, and validated citations.
File attachments — ingest, search, select, cite, recover, and delete structured evidence from files and folders.
Evidence-gated self-improvement — inspect proposals, gates, independent reviews, bounded rollouts, rollback, and privacy controls.
Batch and dataset execution — import versioned tables, preflight budgets, run isolated row tasks, retry failures, and export linked results.

The interface

The web interface — the sixteen tabs, the composer, @-mentions, and accessibility.
Browser agent — built-in semantic browser automation, profiles, approvals,
live supervision, trace export, and repair.
Native computer use — accessibility-first Windows application control,
selected-window privacy, emergency takeover, traces, support limits, and troubleshooting.
Language and region — interface, conversation, and voice language choices, regional formats, reviewed safety translations, and pseudo-locales.
Annotations — anchor, discuss, assign, remap, resolve, and act on exact review targets.
Canvas — live visual surfaces the agent renders beside the chat.
Professional artifacts — create, validate, preview, revise, compare, and download versioned DOCX, XLSX, PPTX, and PDF files.
Settings — the complete reference for every configuration option.
Providers & models — providers, models, nicknames, and switching.
Connected Apps — scoped mail, calendar, and file accounts, exact action previews, health, activity, and complete revoke.
Adaptive model routing — opt-in task-aware selection, hard constraints,
dry-run diagnostics, telemetry, and rollback.
Provider recovery — opt-in retry, eligible-model failover, durable
circuits, unsafe replay stops, diagnostics, and rollback.
Prompt caching — deterministic prefix reuse, retention controls, isolation,
and diagnostics.
Local models — hardware fit, verified files, Ollama/llama.cpp lifecycle, roles, benchmarks, and safe removal.
Keyboard shortcuts — defaults, the Ctrl+. reference, and rebinding.
Sounds — per-event audio cues and custom clips.
The desktop app — window, tray, native menus, and auto-update.

Working with the agent

Autonomy & approvals — postures, the approval gate, policies, and dry-run plans.
Organization governance — enterprise identity, tenant scope, RBAC, organization policy, retention, and audit.
Slash commands — built-ins, your own commands, sequences, and command packs.
Memory — what the agent remembers and how recall works.
Privacy-reviewed training data — select successful runs, inspect
redactions, and explicitly create local fine-tuning JSONL with provenance.
Skills — teachable procedures the agent can learn and reuse.
Trusted marketplace — inspect and manage signed skills, connectors,
plugins, assets, and isolated executable extensions across public, private, and offline
registries.
Executable extensions — inspect, configure, and explicitly enable isolated
out-of-process SDK extensions.
Messaging channels — bind external chat identities, inspect effective
policy and delivery health, and use stop/recovery controls.
Tools & MCP — native tools, MCP servers, images, and execution backends.
MCP Apps — isolated interactive tool UIs, approval boundaries, inspection, and recovery.
Scheduler & automation — unattended jobs, triggers, and how they stay safe.
Workflows & reusable runbooks — typed multi-step automation, recipes, approvals,
triggers, durable recovery, and visual supervision.
Event sources — durable webhook, folder, timer, and connector signals with health and replay.
Lifecycle hooks — deterministic policy and automation boundaries,
trusted project configuration, handler isolation, dry-runs, and redacted traces.

Collaboration

Subagents & coding jobs — delegated runs and conducting coding CLIs.
Agent teams — linked agents that collaborate on one goal.
Task packages — reviewed, signed, portable task handoffs with immutable
provenance, clean-profile dry runs, local remaps, and reproducible evidence.
Mobile supervision — pair an installable pocket console to monitor,
approve, steer, stop, receive alerts, and recover a lost device.
Device capability nodes — securely use a paired phone for an
approved camera capture, location snapshot, notification, or selected file.
Peers & networking — your agents on multiple machines, discovery, and pairing.

Reference

Voice — speech input and output.
Security & privacy — the trust model, secrets, and network exposure.
Encrypted local state — opt-in envelope encryption, migration,
key rotation, backup/restore, recovery responsibilities, and troubleshooting.
Scoped secret broker — opaque references, exact JIT leases,
trusted injection, rotation, revocation, redaction, and break-glass.
Troubleshooting — common problems and their fixes.
Run usage diagnostics — correlated timing, token, retry, and optional cost records without conversation content.
HolarynBench — deterministic and opt-in live capability evaluation, baselines, reports, and honest score interpretation.
Dynamic tool discovery — bounded capability search, exact-schema loading, scoped policy, and optional restricted multi-call orchestration.

← Previous Use Holaryn in Visual Studio Code Next → Installation