AgentOS Workbench

Visual debugging and orchestration dashboard for AgentOS agents.

Website · Docs · GitHub · Discord · npm

---

AgentOS Workbench is a React + Vite dashboard for inspecting, debugging, and orchestrating AgentOS agent sessions. It provides a zero-config cockpit for streaming AgentOS chunks, session timelines, multi-agent coordination, evaluation benchmarks, and plan lifecycle management.

Built on @framers/agentos, the open-source TypeScript runtime for building production AI agents with cognitive memory, HEXACO personality, multi-agent orchestration, and runtime tool forging.

Features

Feature	Description
Session Inspector	Color-coded timeline rendering of streaming AgentOS chunks (text deltas, tool calls, tool results, workflow updates, agency updates, errors)
Compose	Request composer for prototyping agent turns, replaying transcripts, and testing multi-turn conversations
Multi-Agent Dashboard	Visualization of 6 coordination strategies (sequential, parallel, debate, review loop, hierarchical, graph DAG)
Adaptive Execution	Task-outcome KPI tracking, fail-open overrides, tool-exposure recovery state from the adaptive execution runtime
Evaluation	Benchmark runner for testing agent quality, response accuracy, and guardrail effectiveness
Planning	Plan lifecycle management with checkpoint history, fork/restore, and runtime-backed graph-run inspection
RAG Workspace	Live retrieval + demo-backed document-library fallbacks across 7 vector backends
Runtime Inspector	Inspect exports from generateText, generateImage, AgentGraph, workflow(), and mission()

Quick Start

# 1. Clone and install
git clone https://github.com/framersai/agentos-workbench.git
cd agentos-workbench
pnpm install

# 2. Configure environment
cp .env.example .env.local
# Edit .env.local with your backend URL and API keys

# 3. Start the backend
pnpm --filter backend dev

# 4. Start the workbench
pnpm dev
# Opens at http://localhost:5175

Configuration

Frontend Environment

# Option A: Explicit API base URL
VITE_API_URL=http://localhost:3001

# Option B: Same-origin /api/* with dev proxy
VITE_BACKEND_PORT=3001
VITE_BACKEND_HOST=localhost
VITE_BACKEND_PROTOCOL=http

Backend Environment

AGENTOS_WORKBENCH_BACKEND_PORT=3001
AGENTOS_WORKBENCH_BACKEND_HOST=0.0.0.0
AGENTOS_WORKBENCH_EVALUATION_STORE_PATH=../.data/evaluation-store.json
AGENTOS_WORKBENCH_PLANNING_STORE_PATH=../.data/planning-store.json

Provider API keys (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) should be set in the backend environment. AgentOS supports 21 LLM providers with automatic fallback chains.

GMIs, Agents, and Agencies

AgentOS uses a three-layer cognitive architecture:

Layer	Purpose	Configuration
GMI (Generalized Mind Instance)	Persona prompts, memory policies, tool permissions, language preferences, guardrail hooks	Reusable cognitive cores versioned and exported across apps
Agent	Product surface (labels, icons, availability) wrapping a GMI	Preserves GMI cognition and policy
Agency	Coordinates multiple GMIs via 6 workflow strategies	Visualized via WORKFLOW_UPDATE and AGENCY_UPDATE events

API Endpoints

Method	Endpoint	Description
POST	/api/agentos/chat	Send a turn (messages, mode, optional workflow)
GET	/api/agentos/stream	SSE stream for incremental updates
GET	/api/agentos/personas	List personas (filters: capability, tier, search)
GET	/api/agentos/workflows/definitions	List workflow definitions
POST	/api/agentos/agency/workflow/start	Start agency workflow
GET	/api/agentos/graph-runs	List persisted runtime graph-run records
GET	/api/agentos/graph-runs/:runId	Inspect a single graph-run record
GET	/api/evaluation/runs	List evaluation runs
POST	/api/evaluation/run	Start a new evaluation run
GET	/api/planning/plans	List persisted plans
POST	/api/planning/plans	Create a new plan

See backend/docs/index.html for the generated backend route documentation.

Storage and Data

• Local-first: all data stored in your browser via IndexedDB (no server writes)
• Stored entities: personas (remote + local), agencies, sessions (timeline events)
• Export: per-session from the timeline header, or all data from Settings > Data > "Export all"
• Import: Settings > Data > "Import..." (schema: agentos-workbench-export-v1)
• Clear: Settings > Data > "Clear storage"

See CLIENT_STORAGE_AND_EXPORTS.md for details.

Scripts

pnpm dev              # Vite dev server at http://localhost:5175
pnpm build            # Production build (emits dist/)
pnpm preview          # Preview production build
pnpm lint             # ESLint
pnpm typecheck        # TypeScript type checking
pnpm e2e              # All Playwright test suites
pnpm e2e:chromium     # Chromium only
pnpm e2e:firefox      # Firefox only
pnpm e2e:webkit       # WebKit (serialized for stability)
pnpm bundle:report    # Bundle size analysis
pnpm bundle:check     # Enforce bundle size budgets
pnpm build:check      # Build + bundle report + budget enforcement

AgentOS Ecosystem

Package	Description	Links
@framers/agentos	Core TypeScript AI agent runtime	GitHub · Docs
@framers/sql-storage-adapter	SQL persistence for agent memory and sessions	npm
AgentOS Workbench	Visual debugging dashboard (this repo)	GitHub
AgentOS Docs	Guides, tutorials, and TypeDoc API reference	docs.agentos.sh
Wilds.ai	AI game worlds powered by AgentOS	wilds.ai

License

• AgentOS core (@framers/agentos) — Apache 2.0
• Workbench — MIT

---

  ·  

Built by Manic Agency LLC / Frame.dev