OpenFang

Hands

7 Built-in Packages + Custom Hands

Overview
What It Is
Core Concepts
Architecture
Security (16 Layers)
Hands
Ecosystem
Getting Started
Benchmarks
Whitepaper
Sources

Hands: Autonomous Capability Packages (Agents that work for you)

Hands are OpenFang's most "productized" innovation: packaging a class of repeatable, autonomous, and observable work into a deployable unit.

Hands vs Traditional Agents

DimensionTraditional AgentHand
TriggerYou send a messagecron/schedule, events, manual activation
LifecycleRequest-ResponseLong-running, with checkpoints, supports pause/resume
ObservabilityReading chat logsDashboard metrics + status API
ReproducibilityHighly dependent on prompt phrasingManifest + versioned playbook

Three-Layer Structure: Why It "Feels More Like a Job Role"

From docs/hands:

  1. HAND.toml: The job description (tools, settings, schedule, dashboard metrics)
  2. System Prompt (playbook): The training manual (multi-stage processes, decision trees, error recovery, quality gates)
  3. SKILL.md: Domain knowledge base (reference knowledge, rather than instructions)

Additionally, all three layers of content are compiled into the binary via include_str!() during the build process, making their delivery and versioning more like system components.

7 Built-in Hands (Official/README)

HandCategoryWhat It Does (Summary)
ClipContentCuts long videos into short clips: selects segments, adds subtitles, thumbnails, optional dubbing, distribution
LeadDataDaily lead generation: discovery-enrichment-scoring-deduplication-export, gradually learns your ICP
CollectorIntelligenceOSINT-style target monitoring: change detection, sentiment/trends, knowledge graphs, critical alerts
PredictorForecastingSuperforecasting: evidence chains + probabilities + Brier score calibration + contrarian pattern generation
ResearcherResearchMulti-source cross-validation + CRAAP evaluation + APA citations, multi-language reports
TwitterDistributionX management: multi-type content rotation, scheduling, interaction, approval queues, performance tracking
BrowserAutomationPlaywright-based web flow automation (including forced "purchase approval gates")

Hand Lifecycle Commands (docs/hands)

bash
openfang hand activate researcheropenfang hand status researcheropenfang hand pause researcheropenfang hand resume researcheropenfang hand deactivate researcher

How to Customize a Hand (Summary of docs/hands)

You need a directory containing:

  • HAND.toml
  • system-prompt.md
  • SKILL.md (optional but highly recommended)

The key is writing the playbook as an executable workflow: divided by Phases, with explicit conditional branches, error recovery, and quality thresholds.