Research
Decision briefs, competitor monitoring, source-backed analysis.
Your AI. A workspace for the whole job.
Give AI the job.
Not just the question.
AGNT is a free, local-first agent operating system. Code, research, create, and automate in one workspace—bring your supported AI subscriptions, API keys, or local models, and turn useful work into tools and workflows you can use again.
Free local app · Windows, macOS, and Linux · Model access and usage depend on your provider.
Builders from teams and universities including
Organization names reflect individual affiliation. No partnership or endorsement is implied.
This is AGNT
More than an answer. A place to do the work.
Research, code, agents, and repeatable workflows. Explore the actual AGNT workspace below.
A conversation that produces something.
See a rich response inside AGNT, with saved outputs, workflows, and integrations close at hand.
A workspace for the agents you build.
Organize your agents, inspect their configuration, and talk to them without leaving the workspace.
Give repeatable work a repeatable structure.
Connect triggers, tools, and code in a visual workflow. Inspect the configuration alongside the graph.
One canvas. More than one kind of work.
Explore AGNT’s visual workspace with workflow nodes, writing, charts, code, images, and audio side by side.
Actual AGNT product screenshots · Click a view to explore or enlarge.
The AI you already pay for
ChatGPT
Claude
Google AI
Grok
Cursor
Kimi Codeand moreOne workspace. Your choice.
ChatGPTClaudeGoogle AIGrokCursorKimi Codeand moreSupported connections. Provider plan limits and terms apply.
Start with a job.
Leave with something useful.
Run a single operation. Keep it only if the result earns another run.
Operations
KPI reports, reconciliation, meeting follow-up, queue triage.
Customer work
Lead qualification, outreach, support research, proposals.
Engineering
Issue investigation, code changes, checks, and release work.
GmailGoogle DriveGoogle SheetsGitHubSlackDiscordNotionCustom APIsOllamaOpenAI-compatible models
Keep your AI.
Expand what it can do.
You do not have to abandon the AI you already like. Bring supported model access and choose your execution options. Keep the work around them together in AGNT.
01 / SUBSCRIPTIONS
Use what you already pay for.
Connect supported AI subscriptions alongside your API keys. Put your existing access to work instead of starting with another token bill.
ChatGPTClaudeGoogle AIGrokCursor
Kimi Codeand more02 / AGENT ENGINES
Your agents. Working together.
Keep the execution engines you trust. AGNT can delegate to Grok CLI, Codex, Claude Code, Hermes, and OpenCode through CLI or SDK integrations.
Explore the execution options03 / MODELS
The right model for the job.
Frontier models, fast specialists, or a model on your own machine. Choose from 20+ providers and local runtimes in the same workspace.
OpenRouter
Ollama
LM StudioSupported plans, authentication methods, and integration features vary. Provider limits and terms still apply. CLI and SDK delegation are not identical to native AGNT execution.
Where AGNT fits
Stop rebuilding your process around your AI.
Use deterministic automation when every step is known. Use AGNT when the job needs context, judgment, tools, approvals, and a durable record.
Not another chat tab
AGNT carries work through tools instead of stopping at advice.
It does the work →Not a black box
Sources, actions, approvals, cost, and failures remain visible.
Every step stays visible →Not cloud-only
Community Core runs locally. Cloud removes operational overhead.
Local first, cloud optional →Outcome first.
Automation second.
You do not need to design an agent before you know whether the job is useful.
01Name the finish line
Describe what should exist when the work is done.
02Approve access
AGNT requests only the tools and authority the operation needs.
03Inspect the receipt
Review the output, evidence, cost, and any unresolved failure.
Meet Annie · Your local AI assistant
Say hello to Annie.
The agent that builds your agents / workflows / tools / skills / then puts your AI to work and hands you the receipt.
🧠Persistent memory
Pick up where you left off.
Preferences, corrections, facts, and workflow learnings persist across every conversation. Annie gets better the more you work with her.
⚡100+ tools
Tools that act on your work.
Web search, scraping, file operations, code execution, image and video generation, email, spreadsheets, Git, shell — plus every plugin tool in the registry.
🎯30+ skills
Expert-level on demand.
Skills activate contextually: frontend design, research synthesis, plugin building, codebase analysis, media pipelines, and more.
🔄Async execution
Keep work moving in the background.
Queue tasks to run in parallel, on intervals, or after a delay. Annie monitors, polls, and iterates while you do other things.
📂Full filesystem
Your machine. Your files.
Read, write, edit, and organize files anywhere on your system. Create projects, generate code, produce reports — all local, all yours.
🌐API orchestration
Your accounts. Her hands.
Annie uses your stored OAuth tokens to call GitHub, Gmail, Slack, Notion, Stripe, and any connected provider directly.
Give AI a job.
Get the proof.
Start with the result you need. AGNT plans the work, uses approved tools, and stops when the job is complete or blocked.
Request
“Build this week’s operating report and flag anything that needs a decision.”
agnt · run #48210:00
- planScope approved — 6 steps, 3 tools, 1 gate
- postgresPull weekly ops metrics
- stripeRevenue + failed charges
- sheetsCross-check against budget
- gateAd spend 34% over budget — routed to you
- docsReport drafted → Google Docs
Receipt sealed — every action, source & cost recorded
A blocked run stays blocked. AGNT does not label partial work complete.
Builders don’t
hold back.
Unedited words from the AGNT community — developers, operators, and first-timers who ran a real job and saw the receipt. Spelling preserved exactly as written.
“So far you’ve accomplished what AutoGPT did not for me: have actual workflows running, not on my own computer, in a non-containerized environment, that would otherwise require me to invest more time and effort setting up.”
AE
Automation engineer
Verified testimonial
“AGNT is the best orchestration harness I’ve come across. I recommend all of you try this out! This interactive demo is so impressive — just off scraping our website, Annie built this whole interactive page. Real-world use cases included.”
CB
Community builder
AGNT Discord · Jun 25, 2026
“After using a lot of different systems I can tell you the way AGNT works when people see it run is an eye opener. The key is to see it run as someone building.”
EO
Early operator
AGNT Discord · Jun 23, 2026
“On a high level the workflow builder seems great, and the UI is super nice. Really looking to accelerate my workflows and run more parallel agents.”
WB
Workflow builder
AGNT Discord · Jun 29, 2026
“I have a skill that I run after each workflow that patches the workflow skill — but this is automatic and the insights are really good! Cool.”
PU
Power user
AGNT Discord · Jun 25, 2026
“I’m a normal user, no IT background — my job is nurse, haha. But I’m learning and want to go deeper into AI. AGNT works well for me.”
NL
Nurse & AI learner
AGNT Discord · Jun 8, 2026
Authentic community feedback, preserved as written. Handles shown where a public testimonial was given; other members attributed by role and date to protect privacy. Not paid endorsements.
Make your AI budget go further
More work.
Less AI spend.
Cache-aware execution and supported subscriptions can change the economics of agent work. Here is one recorded workload, with the numbers behind it.
98.75%
below the uncached API equivalent
Caching + subscription economics.
Not a caching-only claim or a guaranteed bill reduction.
Uncached API equivalent$8,012.00
Cache-adjusted API equivalent$1,121.07
86.01% lower estimated API cost from recorded cache reuse.
Allocated subscription fee$100.00
One full monthly fee allocated to this recorded usage.
792.71M recorded processed tokens / Sep 6, 2026 snapshot
Cache-aware by design.
AGNT keeps reusable instructions, tool definitions, and conversation prefixes stable to help supported providers reuse cached input.
See the economics.
Inspect usage and estimated cache savings. Keep metered API costs separate from subscription fees and API-equivalent comparisons.
Choose how you run.
Use supported subscriptions, metered APIs, or local models. External agent engines retain their own caching and billing behavior.
Read the comparison methodology
This comparison uses 792,710,774 recorded processed tokens in AGNT’s Codex/Astra-labelled usage through September 6, 2026: 790,588,439 input tokens, including 765,659,136 cached input tokens, and 2,122,335 output tokens. Processed tokens include context reused across requests; they are not unique words.
Assumed API rates: $10 per million uncached input tokens, $1 per million cached input tokens, and $50 per million output tokens. These are calculation assumptions, not a newly verified public rate card. The cached rate reconciles to the stored estimate.
The comparison allocates one full $100 monthly subscription fee to the recorded usage. It is not a matched billing-cycle invoice, a controlled benchmark, or a guarantee of future savings. Recorded usage has completeness and model-attribution limitations. Caching accounts for 86.01% of the uncached API-equivalent reduction; subscription pricing accounts for the additional difference. Native provider caching is also available outside AGNT. Results depend on the model, plan, workload, and cache reuse.
Caching + subscription economics. See the breakdown
Start free.
Go Pro and your agent gets seven services.
Community Core runs free on your machine. Pro hosts your agent and includes everything it works with, each with a monthly allowance. One price, no extra service fees.
- The coreHosted instanceYour agent on a dedicated machine, at yourname.agnt.gg.
- 01ModelsA 1M-context model with tools, vision and streaming.
- 02SearchTitles, links, snippets and sources for every query.
- 03ScrapeAny page or file as markdown: PDF, Word, Excel, slides.
- 04SandboxAn isolated VM per task. Run, export, destroy.
- 05MailAn inbox your agent owns. Send, receive, reply.
- 06WebhooksA public URL. Stored, signed, replayable.
Community Core
$0 forever
Do useful work on your own machine.
- RunsOn your machine · unlimited local runs
- ModelsYour own keys or local
- SearchNot included
- ScrapeYour own browser
- SandboxNot included
- MailNot included
- WebhooksNot included
Agents, workflows, tools and memory.
Download freeAGNT Pro
$29/mo
Your agent, hosted. Everything included.
- Hosted instanceSleeps when idle · up to 8 active hrs/day
- Models5M credits
- Search150 searches
- Scrape2,250 pages
- Sandbox100 min
- Mail1 inbox · 1,000 units
- Webhooks10 endpoints · 1,000 units
- Mobile1 phone · 500 texts
Bought separately: $62.50/mo
Pro + Always-On
$49/mo
Pro, awake around the clock.
- Hosted instanceNever sleeps
- Models15M credits
- Search500 searches
- Scrape7,000 pages
- Sandbox300 min
- Mail5 inboxes · 5,000 units
- Webhooks25 endpoints · 5,000 units
- Mobile2 phones · 2,000 texts
Bought separately: $149.50/mo
AGNT Team
$99/mo
Always-On for three people, one shared instance.
- Hosted instanceNever sleeps · shared vault · audit receipts
- Models25M credits
- Search1,400 searches
- Scrape18,000 pages
- Sandbox600 min
- Mail15 inboxes · 20,000 units
- Webhooks100 endpoints · 20,000 units
- Mobile5 phones · 5,000 texts
Bought separately: $345/mo
Enterprise and Managed Operations
Talk to AGNTSSO and access controls, private or hybrid deployment, security review and negotiated support. Includes the Team allowances.
Monthly allowances. You pay the plan price and nothing else; nothing runs past an allowance unless you authorize it. Cancel anytime and keep Core. Every allowance, compared
Deploy anywhere
Runs where you run.
The same runtime, receipts, and vault everywhere — native desktop apps, a single Docker command, your own servers, or AGNT-operated cloud.
Desktop apps
Native builds with local model support and the OS-keychain vault.
Docker
Headless on any box — NAS, VPS, homelab — in one command.
$ docker run -d -p 3333:3333 ghcr.io/agnt-gg/agnt
Self-hosted
Full stack on your servers. Source available — audit it and extend it.
AGNT Cloud
Zero-ops hosting with Auth Manager keeping every token fresh.
Docker / Headless CLI:
docker run -d -p 3333:3333 ghcr.io/agnt-gg/agnt:latestWhat do you want
to accomplish first?
Download AGNT, connect supported model access, and start with one useful task. These prompts need no additional service accounts.
Research and reports
Build a source-linked brief comparing three products I name. Separate observed facts from recommendations. Cite primary sources and flag anything you cannot verify. Ask me which products first.
Code and engineering
Review a project I choose, read-only. Identify one concrete issue, explain the evidence, and propose the smallest fix and verification plan. Do not edit files. Ask me for the project path first.
Repeatable work
Using only sample data you generate, demonstrate a repeatable report calculation. Run it twice, check that the results match, and show the inputs, method, and output. Do not connect external accounts.
Review the output before acting on it. Model and tool availability depend on your setup.
Every major model.
Your choice, every job.
AGNT connects to 20+ AI providers — frontier cloud models and fully local runtimes side by side. Bring supported subscriptions or API keys, choose a model per task, and keep your options open. Provider plan limits and terms apply.
OpenAIAnthropicGoogle GeminixAI Grok
DeepSeek
Meta Llama
Mistral AI
QwenMoonshot Kimi
Z.AI GLM
MiniMaxOllamaLM Studio
Hugging Face
Cohere
Perplexity
Groq
Cerebras
Together AIOpenRouter
Replicate
NVIDIA NIM
Azure OpenAI
Amazon BedrockModel names and logos are the trademarks of their respective owners and indicate compatibility, not endorsement.
50+ integrations,
built in on day one.
119 native tool nodes span the whole stack — from GitHub and Slack to Stripe and Postgres — plus a Zapier bridge to 6,000+ apps and a client for any MCP server. No glue code, no brittle webhook plumbing.
50+ native integrations119 tool nodes6,000+ apps via Zapier∞ via MCP
GitHub
Docker
Vercel
Netlify
Sentry
Linear
Jira
PostgreSQL
Redis
Gmail
Google Sheets
Google Drive
Google Slides
Google Calendar
Microsoft 365
YouTube
Slack
Discord
Telegram
X (Twitter)
Twilio
Notion
Obsidian
Calendly
Dropbox
Asana
Trello
Salesforce
HubSpot
Stripe
Firecrawl
Airtable
ElevenLabs
UnsplashZapier
QuickBooksLogos are the trademarks of their respective owners. Additional providers reachable through the Zapier bridge and MCP.
Do not start from a blank chat.
Browse ready-to-run agents, tools, workflows, skills, plugins, and widgets. Every marketplace page gives you the idea, the working template, and a direct path to run it in AGNT.
AgentsAutonomous workers with tools and memory.Install ToolsSingle actions any agent can call.Install WorkflowsMulti-step pipelines with approvals.Install SkillsExpert playbooks, loaded on demand.Install PluginsNode packs that extend the runtime.Install WidgetsLive dashboard cards on your data.Install
01 · DISCOVER
Start with the job
Find proven systems for reporting, research, support, engineering, finance, and more.
02 · DOWNLOAD
Get the whole asset
Install an inspectable package with its configuration, sample input, permissions, and version history.
03 · ADAPT
Make every part yours
Change the model, tools, routing, approvals, output, and schedule after the first run.
The platform
Every execution primitive.
One accountable runtime.
Agents for judgment, goals for long-running outcomes, workflows for determinism — all sharing the same tools, vault, approval gates, and receipts.
Agents — a workforce, not a chatbot
Persistent workers with scoped tools, memory, and credit budgets. Each one owns a job and reports back with evidence.
Research Agent
Support Triage
Finance Ops
Long-running goals
Define the outcome and its success criteria. AGNT iterates — execute, evaluate, replan — for hours or days until the goal passes.
iter 6evaluated — 71% of criteria met · replanned
iter 7executing — step 3 of 5 · cross-checking sources
execute→evaluate→replan↺
Workflows for the deterministic half
Timers, webhooks, and app events trigger node pipelines with retries and approval gates — when every step is known in advance.
webhook→enrich→approve→post→receipt
paused · waiting on approval from @dana · 2m
Controls your security team will sign off on
The same runtime scales from one laptop to a team without changing how work is verified.
- Approval gates on sensitive actions
- A receipt for every run — actions, cost, failures
- Credentials in the OS keychain, injected at run time
- Self-host, desktop, or Docker — your infrastructure
Auth Manager · Cloud
Connect once.
Stop babysitting OAuth.
Paid Cloud plans include AGNT Auth Manager for supported providers. It monitors connection health and refreshes eligible tokens before routine expiration.
Googleoauth2 · a91f…4e07 · scopes: mail, calendar, drive✓ refreshed 2m ago
Slackoauth2 · 77c2…b1aa · scopes: chat:write, channels:read✓ refreshed 1h ago
Notionoauth2 · 30de…9c51 · provider requires re-loginreauth neededAlert sent
AGNT Black Label
The black-glove service.We run the operation. You read the receipts.
For companies that want a defined production operation owned end to end — designed, hardened, and answered for by the team that builds AGNT itself.
We own the outcome, not the tooling
A defined operation with agreed success criteria and reviews — you hold us accountable for results, not software licenses.
Accountability, contractedProduction-grade without the hires
Skip recruiting an automation team. The people who build AGNT design, harden, and operate the system for you.
Zero added headcountEvery run stays verifiable
The same receipts, approval gates, and audit trail — your team keeps full visibility into every action we take.
Full audit trailExplore Managed ServicesAnnual engagements start at $100K
Research on how agents work.
We publish leading work on agent architectures, evaluation, verification, and the practical engineering of accountable AI operations. Written by the people building AGNT.
Pillar guide · Start here
AI harness engineering: how to build AI agents your business can trust
The permissions, memory, evaluations, cost limits and recovery that make agents safe on business data, with a tested starter and twenty companion guides.
Essay
Everything as Code: the quiet reformation of creative work
Guide
The 100 best AI agent skills in 2026
Whitepaper
AGNT Memory: self-improving agent infrastructure
Field note
The NN-shaped hammer: when to learn and when to write the code
Guide
The complete guide to AI agent architectures in 2026
Research
AI agent evaluation and the gold standard system
AGNT at a glance
The facts, in one place.
| What it is | A local-first agent operating system: agents with their own memory, tools and budgets, visual workflows, autonomous goals, skills and MCP tools, with a receipt for every run. |
|---|---|
| Made by | AGNT Labs, agnt.gg |
| Runs on | Windows, macOS and Linux desktop apps and Docker. Reachable from the web and the AGNT Chat mobile app, paired by QR code, locally or through a remote server. |
| Price | Community Core: free for personal use and small businesses, unlimited local runs. AGNT Pro $29/mo, Pro + Always-On $49/mo and AGNT Team $99/mo with 3 seats, each including a hosted instance plus Models, Search, Scrape, Sandbox, Mail and Webhooks. Enterprise: custom. See pricing |
| License | Source available under the AGNT Community Core License: free for personal use, education, nonprofits and businesses under $1M revenue and 10 people. Not OSI open source. |
| Models | 20+ providers, supported AI subscriptions, and local models through Ollama and LM Studio, chosen per agent. |
| Agent engines | Delegates work to Codex, Claude Code, Hermes, Grok CLI and OpenCode. |
| Connects to | 50+ built-in integrations, any MCP server, any REST API, and 6,000+ apps through Zapier. |
| Your data | Runs on your machine. External models receive only what an approved step sends; keys stay in your operating system keychain. |
| Compared with | Claude Code · OpenAI Codex · OpenClaw · Hermes Agent · n8n · all agentic systems |
FAQ
Common questions.
Clear boundaries before you commit — what is free, what stays local, and what happens when things fail.
Read the docs →Is AGNT really free?
Community Core is free for personal use, education, nonprofits and businesses under $1M revenue and 10 people. Larger businesses need a paid plan, and paid plans also cover AGNT-operated cloud services, managed connections, teams, and support.
Can I use the AI subscriptions I already have?
Yes. AGNT supports subscription-backed connections alongside API keys and local models. Availability depends on the provider, plan, authentication method, and integration. Your provider’s usage limits and terms still apply; connecting a subscription does not make it unlimited.
Is AGNT a harness, or does it work with other harnesses?
Both. AGNT runs its own model-and-tool loop and can delegate work to external engines such as Grok CLI, Codex, Claude Code, Hermes, and OpenCode. It can use installed CLIs and SDK integrations, so you can keep your coding agents and coordinate broader work around them. Features, session controls, and billing behavior vary by integration.
How much can AGNT reduce AI costs?
It depends on your workload and how you connect. In the recorded comparison above, caching reduced the uncached API equivalent by 86.01%. Allocating a full $100 subscription fee to that usage put it 98.75% below the uncached API equivalent. That combines cache reuse and subscription economics — not a universal discount or a caching-only saving. Read the methodology.
Can I use Ollama or other local models?
Yes — Ollama and LM Studio run through a compatible local endpoint, fully offline. You can also bring your own commercial provider key and mix local and cloud models per task.
What happens when a run fails?
The receipt records the failed step, error, retries, partial output, and whether it is safe to resume. Nothing is silently swallowed.
Does local-first mean no data leaves my machine?
Only if you choose local models and local tools. External models and APIs receive the data required for the approved operation — and the receipt shows exactly what was sent where.
Where are my API keys and tokens stored?
Fully local by default — in your operating system's encrypted keychain, on your machine, injected at execution time and never written to logs or prompts. Optional Cloud Sync keeps your OAuth tokens fresh across devices, encrypted end to end.
Can an agent act without my approval?
Only within the authority you granted. Sensitive actions — sending email, changing records, spending money — can be gated behind explicit approval steps, and every action lands in the receipt.
What if the tool I need is not built in?
Bridge to 6,000+ apps through Zapier, connect any MCP server, call any REST API directly, or build a custom tool in Tool Forge — no plugin release cycle required.
Can I self-host everything?
Yes. Run the desktop app or the Docker image on your own hardware. The source is public under the AGNT Community Core License: read it, audit it, and modify it for your own use. It is source available rather than OSI open source: paid setup, services and resale for others need a Certified Partner license, and white-label, OEM or hosting for others needs an Enterprise agreement.
What do you want to stop
doing the hard way?
Start with a task. Keep what works. Build on it next time.