Build agent workflows that route, validate, review, and remember every run.
AgentOS is a TypeScript orchestration layer for composing specialist agents across Gemini, Kimi, DeepSeek, and offline adapters. It gives developers a local-first runtime, a dashboard, static docs, and a Pages-ready public console.
Router-first
Tasks are routed by risk, latency, context, multimodal needs, and confidence.
Schema-native
Every agent output is validated with Zod and retried when structure breaks.
Audit-ready
Runs are stored with prompts, routing traces, model choices, confidence, and review status.
Quickstart
Run the whole stack locally without provider keys. Offline adapters keep reference runs deterministic and free.
git clone https://github.com/fintie/agentos.git
cd agentos
npm install
cp .env.example .env
npm test
npm run dashboardhttp://localhost:4317 to use the AgentOS desktop, run individual agents, inspect workflows, and review evaluation records.What You Can Build
Aged care workflows
Parse shift notes, draft care reports, review compliance, and generate follow-ups.
STEM tutoring
Give feedback, check correctness, suggest next practice, and produce parent summaries.
Voice operations
Summarize transcripts, plan follow-ups, detect escalations, and prepare CRM notes.
Developer agents
Turn requirements into architecture plans, review code risk, and produce implementation notes.
Architecture
Agent definition
Metadata, prompt builder, task type, risk level, and Zod schema live in the agent registry.
Routing decision
The router picks a model family and records candidates, estimated cost, rule id, and escalation state.
Structured execution
The orchestrator calls the selected adapter and validates the response before returning parsed output.
Evaluation record
Each run lands in memory or Prisma-backed storage for review, comparison, and audit.
Agents
Agents are reusable units that describe their job, prompt, expected output, and default routing context.
| Agent | Purpose | Typical model |
|---|---|---|
CareNoteAgent | Generate care summaries from structured shift data. | Kimi or offline |
ComplianceReviewAgent | Review safety, compliance, and escalation requirements. | DeepSeek or offline |
DeveloperAgent | Create implementation plans and architecture guidance. | Kimi or offline |
JudgeAgent | Score and validate outputs as an LLM judge. | DeepSeek or offline |
Workflows
Workflows compose agents into vertical flows. The dashboard exposes runnable workflow pages for Care, STEM, Voice, Developer, Review, and Incident-style tasks.
raw input
-> specialist parser
-> domain agent
-> review agent
-> report or action plan
-> evaluation logModel Routing
Routing rules are explicit and inspectable. AgentOS can start cheap and fast, then escalate when confidence drops, risk rises, schemas fail, or a workflow requires stronger validation.
Gemini 3 Flash
Fast, low-latency, multimodal and interactive work.
Kimi K2.6
Long-context planning, coding workflows, and multi-agent reasoning.
DeepSeek V4 Pro
Deep reasoning, compliance review, final validation, and judging.
Evaluation Log
The evaluation layer records model name, agent name, prompt version, input hash, parsed output, confidence, review model, human-review status, and routing trace.
npm run seed
npm run dashboard
# open Evaluation log in the consoleExtend AgentOS
Add new capabilities by creating agents, workflows, adapters, or evaluation backends. Keep changes small and schema-backed so other developers can safely compose them.
Add an agent
Create metadata, prompt builder, schema, and example input in the registry/catalog.
Add a workflow
Compose existing agents in src/workflows and expose it through the catalog.
Add a model
Implement the shared adapter interface and update routing rules.
Deploy
The public Pages site is static. Build it with deterministic reference data, then push the generated root files.
npm run build:pages
git add dashboard docs.html index.html scripts/build-static.ts
git commit -m "Add AgentOS docs page"
git push origin mainContribute
AgentOS is easiest to improve through focused pull requests: one agent, one workflow, one adapter, or one docs improvement at a time.