An agent runtime that runs on your hardware.
Most AI automation tools are cloud services that rent you a workflow engine. AGNT is a local-first runtime you install — agents with persistent memory, a visual workflow builder, autonomous goals, MCP tool access, and a verifiable receipt for every action taken.
Everything AGNT does, and where to read about it
AGNT is composed of six primitives that stack. Agents reason and remember. Workflows make execution deterministic. Goals let work run unattended. Skills, plugins, and widgets extend all three.
Agents
Persistent AI workers with their own memory, assigned tools, credit limits, and model routing. They remember prior runs and improve against them instead of starting cold every session.
Explore agents → Execution layerWorkflows
A visual node graph with 119+ built-in tool nodes — triggers, branches, loops, HTTP, filesystem, database, code execution. Deterministic where you need certainty, agentic where you need judgment.
Explore workflows → Autonomy layerGoals
Describe an outcome and let AGNT plan, execute, evaluate, replan, and repeat until success criteria are met — or until it reports honestly that they were not.
Explore goals → Tool layerMCP & Tools
A full Model Context Protocol client, so any MCP server becomes callable by any agent — on top of AGNT's own native tool library and your custom tools.
Explore MCP → ArchitectureLocal-First
Your files, credentials, execution history, and receipts stay on your machine. Community Core ships with zero telemetry and can run fully air-gapped against a local model.
Why local-first → DeploymentSelf-Hosted
Run AGNT on your own server or in Docker. No per-task metering, no execution caps, no vendor holding your automation hostage behind a billing tier.
Self-hosting guide →Rented workflow engine vs. owned runtime
This is the architectural fork that determines everything else — what you can automate, what it costs at scale, and whether your credentials ever leave your machine.
You rent execution
Your automation lives on someone else's infrastructure, priced per task or per operation.
- Metered per task — costs scale with success
- Credentials stored on the vendor's servers
- Cannot reach your local filesystem
- Vendor outage stops your operations
- Execution history is the vendor's to keep
You own the runtime
The engine is installed software. Your hardware, your rules, your data.
- Unlimited local executions, no metering
- Credentials in your OS keychain
- Full local filesystem and shell access
- Runs offline against Ollama or LM Studio
- Every run leaves a local SQLite receipt
What ships in the box
Pick your model provider
AGNT routes to OpenAI, Anthropic, Gemini, Grok, Groq, DeepSeek, Mistral, Qwen, or an entirely local model via Ollama and LM Studio. Swap providers without rewriting a single workflow.
Where to go next
Compare AGNT
Honest, architecture-level comparisons against Zapier, n8n, Make, and ChatGPT AgentKit — including where each one is the better choice.
See comparisons → ConnectIntegrations
GitHub, Slack, Notion, Gmail, Postgres, Stripe, Google Workspace and more — connected through an encrypted local credential vault.
Browse integrations → TrustSecurity architecture
Zero telemetry, OS-keychain credential storage, human approval gates on high-consequence actions, and receipt-grade audit trails.
Read the architecture →