AGNT Agent OS
A local-first runtime for durable AI agents, repeatable workflows, and long-running goals.
- Source repo
- agnt-gg/agnt
- Stars
- ★ 565
- Last updated
- 1d ago
- License
- NOASSERTION
- Primary language
- JavaScript
- FA score
- 51/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An automation builder who wants timer, webhook, Slack, or Discord events to start traceable AI workflows.
- Not a fit if
- Teams needing public multi-tenant hosting or zero-trust tenant isolation
- Large enterprises with 50+ concurrent users on one shared instance
- Users wanting a zero-config cloud chatbot with no local setup
- Website
- agnt.gg ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-09latest release v0.6.8 (2026-11-07)
- Source review
- 51/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
AGNT is a local-first agent operating system composed of an Electron desktop shell, a Vue 3 frontend, and an Express backend on port 3333. It organizes work around agents, visual DAG workflows, and long-running goals, with SQLite and the filesystem providing local persistence. The runtime includes workflow nodes, triggers, plugins, MCP, skills, evaluations, execution records, and real-time streaming interfaces. It can run from source, as a headless self-hosted service, or in Docker; the Full image includes browser automation while Lite omits it. It fits individuals, small trusted teams, and developers who want local control while combining judgment-driven chat with repeatable automation, rather than public multi-tenant SaaS deployments.
Users create and run agents, workflows, and goals through the desktop interface or local API. The agent runtime handles persistent chat, memory, tool calls, skills, streaming output, traces, and token and cost accounting; the workflow engine executes visual DAGs with branches, checkpoints, nested workflows, and triggers such as timers and webhooks; and the goal engine plans, decomposes tasks, executes, evaluates, replans, pauses, resumes, and reverts work. AGNT can call its listed model providers, MCP tools, plugin tools, custom HTTP APIs, and integrations including Slack, Discord, Gmail, and Telegram. Messages, tool calls, errors, outputs, and evaluations can become insights, memory, and Skills. Its backend exposes route families including /api/agents, /api/workflows, /api/goals, /api/plugins, /api/mcp, and /api/skills, with live updates over SSE and Socket.IO.
- An automation builder who wants timer, webhook, Slack, or Discord events to start traceable AI workflows.
- A small product team that needs research, copy, engineering, and evaluation work delegated in parallel while retaining execution records and human approval steps.
- A self-hosting user who wants agents, workflows, credentials, and execution data to remain on a personal machine, homelab server, VPS, or Raspberry Pi.
- A developer combining visual process automation with OpenAI, Anthropic, Gemini, or a custom OpenAI-compatible endpoint.
- An integration developer who needs to extend a local workspace with MCP or packaged
.agntplugins that add tools, triggers, widgets, or workflow nodes.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Ports
3333- Environment variables
JWT_SECRETSESSION_SECRETENCRYPTION_KEY- Config files
.env.env.example- Runtime
- Node.js 18+
- Install notes cover
- Docker
How do you install or deploy this agent?
Source installation requires Node.js 18+, npm 9+, and Git:
git clone https://github.com/agnt-gg/agnt.git
cd agnt
npm installcd frontend && npm install && cd ..
npm startThe backend runs at http://localhost:3333. For frontend hot reload, run cd frontend && npm run dev in one terminal and npm start in another. For a production web build, run cd frontend && npm run build && cd .., then npm start. Remote models and external services require the relevant provider API key, CLI authorization, or integration credentials in AGNT. For Docker or production setup, create .env from .env.example and set JWT_SECRET, SESSION_SECRET, and ENCRYPTION_KEY.
How do you use this agent?
After startup, open http://localhost:3333 and choose among the Agents, Workflows, and Goals execution modes. Use the workflow canvas to combine nodes, branches, checkpoints, and triggers for repeatable work; create a goal for work that needs planning, evaluation, and resumption; and run an agent where judgment and tool use are needed. Other local programs can access resources through the /api surface, including /api/agents, /api/workflows, /api/executions, and /api/goals; no copyable API request example is provided. To create a plugin, use node build-plugin.js my-awesome-plugin from backend/plugins, then install the resulting .agnt package through Marketplace or place it in <data-dir>/plugins/installed/.
What are this agent's strengths and limitations?
- Combines persistent agents, visual workflows, and long-running goals in one runtime instead of delivering only one execution model.
- Its local Express API, SQLite and filesystem storage, SSE, and Socket.IO provide a scriptable and observable automation boundary.
- Ships with 60+ workflow nodes and tool actions, then extends through
.agntplugins, MCP, plugin tools, and custom HTTP APIs. - Supports desktop, Docker, headless servers, VPS deployments, homelabs, and Raspberry Pi, with Full and Lite images that make the browser-automation tradeoff explicit.
- The documented target is a trusted local workspace, not public multi-tenant SaaS, zero-trust tenant isolation, or large enterprises with 50+ concurrent users.
- Source deployment requires Node.js, npm, and Git; production Docker deployments also require the operator to manage persistent volumes, an access layer, and generated secrets.
- Remote models, messaging, and SaaS integrations depend on user-supplied API keys, OAuth or CLI authorization, and external services, whose network use and costs vary by provider.
- Lite removes Chromium, Puppeteer, and Playwright browser automation, so web scraping or browser automation requires the larger Full variant.
- Repository metadata states NOASSERTION while the README says Custom License; adopters should inspect LICENSE.md before use.
How does this agent compare with similar options?
Against LangChain, AutoGen, or CrewAI, AGNT presents itself as more than a framework by adding a desktop UI, local backend, workflow canvas, goals, storage, and runtime APIs. Against n8n or Zapier, it emphasizes persistent agents, memory, SkillForge, MCP, provider authentication, and local traces; against OpenWebUI-style chat apps, chat is only one execution mode. The README also compares it with Hermes Agent, OpenClaw, and Claude Code, positioning AGNT around visual workflows, long-running goals, plugins, evaluations, dashboards, and durable local state.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AGNT Agent OS This agent | 51 · Major gaps | Desktop appFree + model costs | ★ 565 | 1d ago | JavaScript | Codex · Claude Code · OpenAI API · Claude API |
| Auto Company | 59 · Major gaps | CLIFree + model costs | ★ 3.2k | 21d ago | Python | Codex · Claude Code |
| LobeHub — Chief Agent Operator for Multi-Agent Work | 42 · Major gaps | Self-hosted serviceFree + model costs | ★ 85k | today | TypeScript | ChatGPT · OpenAI API · Claude API |
| EpicStaff Agent Workflow Platform | 78 · Good | Self-hosted service | ★ 360 | today | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes local storage, traces, human approval, and rollback; plugin builds require capability and domain declarations, and the Electron source defines a permission allowlist. These provide visible boundaries and review cues, so data-flow transparency and rollback score 2. Deductions: the product spans filesystem access, networking, external APIs, browser automation, and plugins, creating a broad permission surface; approval is described as a feature, but the sample does not show that every consequential action requires it. The API orchestrator skill reads in-process credentials and imports AuthManager directly, while the Figma skill writes a token to local MCP configuration, so sensitive-data handling and least privilege score 1. The sample provides no dependency-audit evidence, so dependency security scores 0. The license names a licensor and CLAUDE.md names an author, but publisher identity is unverified; source attribution scores 1.
The source and documentation describe diagnostic logs, health checks, error handling, and connection-failure guidance, so failure messages score 2. Deductions: the README says 0.6.6 while CLAUDE.md and several install instructions still say 0.5.0, and multi-user support descriptions differ. Installation also depends on npm, Docker images, and external AI services, with no availability guarantees or dependency pinning evidence in this sample; self-consistency and dependency availability score 1.
The README and quickstarts distinguish personal desktop, multi-device, self-hosted, and small-group scenarios, and cover Windows, macOS, Linux, Docker, VPS, and Raspberry Pi, supporting scores of 3 for audience and scenarios and 2 for environment fit. The plugin skill has specific trigger language and requires permission declarations, so trigger precision scores 2. Deduction: the breadth of capabilities leaves some boundaries underspecified, and self-hosting documents differ on multi-user support and authentication, so capability boundaries score 2.
The sample includes a README, installation guides, an API documentation entry point, plugin skills, and multiple quickstarts; navigation, installation, examples, and limitation notes are reasonably developed, scoring 2. The license explicitly identifies the software as source-available and describes commercial, resale, and hosting limits, scoring 2. Deductions: version labels conflict across the README, CLAUDE.md, and quickstarts; the sample does not show a clear changelog or maintenance owner/update commitment. CLAUDE.md includes development guidance, but its boundary between product instructions and contributor guidance is not fully clear, and version differences weaken naming stability.
The product combines agents, workflows, goals, plugins, APIs, and traces in one runtime, and the README gives concrete usage modes, supporting a marginal-value score of 2. Deductions: output quality and cost-benefit are presented mainly as feature and performance claims, without measured benefit, resource comparison, or independent verification in the sample. Deployment and maintenance require local dependencies, model services, or Docker, so cost-benefit scores 1 and output usability scores 2.
The README, source, and guides support some claims about architecture, installation, routes, and permissions. Deductions: many feature-count, provider-support, and self-improvement claims lack itemized implementation evidence or corroboration in the sample; conflicting document versions weaken cross-source checks; marketing claims are not consistently separated from observable facts. All three verifiability criteria score 1.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Version labels conflict in the sample (README says 0.6.6 while other files still say 0.5.0); verify the target revision and matching instructions before deployment.
- Plugins, scripts, and external services can involve network, filesystem, and credential access; review actual permissions, token storage, and external data flows before deploying or extending the product.
- Quickstarts and the newer self-hosting guidance do not fully agree on multi-user authentication; verify authentication and access controls for the target revision before exposing a deployment publicly.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Full Changelog: https://github.com/agnt-gg/agnt/compare/v0.6.7...v0.6.8View on GitHub ↗ -
Release notes excerpt
What's Changed * Make a run visible to every client, and survive a restart by @rimusz in https://github.com/agnt-gg/agnt/pull/62 * fix(skills): a by-id route must check who owns the row by @rimusz in https://github.com/agnt-gg/agnt/pull/68 * fix(models): a save must not erase the columns it does not name by @rimusz in https://github.com/agnt-gg/agnt/pull/67 * test(plugins): the update-status test raced the millisecond clock by @rimusz in https://github.com/agnt-gg/agnt/pull/74 …View on GitHub ↗ -
v0.6.6 0.6.6
FAQ
Does AGNT upload my workspace data?
Can it run without the desktop application?
What does it cost to run?
How should I choose between Lite and Full?
Which port does AGNT Agent OS use?
Is AGNT Agent OS open source?
Is AGNT Agent OS safe to use?
Is AGNT Agent OS still maintained?
Who makes AGNT Agent OS?
What is AGNT Agent OS's official website?
Thanks — we use this to improve the page.