Self-Hosted AI Agent Templates: 6 You Can Watch Run

This week we published a run of short videos on our YouTube channel. Six of them record a single Heym workflow from the first click to the final output, and each of those workflows is a free template you can import today. Watching them back to back, I noticed they had turned into a small catalog of self-hosted AI agent templates that actually run, so I am putting them in one place, with a seventh video as a bonus.

An AI agent template is a saved workflow with the agent, its tools and its output already wired, so you add credentials and run it. The six below cover a first agent, questions over company files, invoice extraction, human-approved support replies, browser automation that repairs itself, and an agent you can trace step by step.

This post is for developers and operators who want a working agent this week, not a framework to study. It is not a ranking of agent platforms, and it is not our gallery either: to browse everything we publish, open our AI agent templates. You get the six templates with their recorded runs, what to change in each, how to combine them, and a bonus on testing a template's prompt before you trust it.

The short version of what you will find here:

• Most AI agent templates are listings, not runs. The pages that rank for "ai agent templates" are galleries with hundreds or thousands of entries. None of the editorial results we read on 27 September shows a template running
• Six templates, six recorded runs. Each section below embeds a 90-second video of the template working, lists its nodes, and says what to change first
• All six are free, self-hosted AI agent templates. They are JSON files for Heym, which you can run on your own machine, so your documents and emails stay on infrastructure you control
• Bonus: test a template before you trust it. A seventh video shows Evals running one prompt against several models and scoring every answer
• The value grows when templates call each other. An agent can call other workflows as tools up to five levels deep, and an MCP client can call a workflow directly
• Nearly 1,300 people starred the repo. The last section is a thank you, and a first word about what we are building next: Heym Work

Table of Contents

• What Is an AI Agent Template?
• Why Most AI Agent Templates Never Run
• The Six Templates at a Glance
• 1. Build Your First AI Agent from One Sentence
• 2. Chat with Company Files and Get the Line It Used
• 3. Invoice Data Extraction from Any PDF Layout
• 4. A Support Reply Agent That Waits for Your Approval
• 5. Browser Automation That Heals Itself
• 6. Watch an Agent Workflow Run, Step by Step
• Bonus: Test a Template's Prompt with Evals
• Composition: AI Agent Workflow Templates as Building Blocks
• Self-Hosted AI Agent Templates: Free and Source-Available
• Nearly 1,300 Stars, and What Comes Next
• Frequently Asked Questions

What Is an AI Agent Template?

Definition: An AI agent template is a saved, working workflow built around an AI agent. The input, the agent with its instructions and tools, and the output are already connected, so you import it, attach your own credentials, and run it.

It sits between a prompt template, which is only text, and an agent framework, which is code you build on. The difference matters when you search. Type "ai agent template" into Google and the autocomplete suggestions mix three different things: prompt templates, UI kits for agent landing pages, and workflow templates. Only the last kind does work on its own.

A useful template answers four questions without you opening a single node. What goes in? Which model and tools does the agent use? What comes out? And what do I need to configure before the first run? In Heym, the answers live on the canvas itself, in sticky notes next to the nodes they describe.

Prebuilt AI agents from SaaS vendors are a related idea with one big difference. A prebuilt agent runs on the vendor's platform. A template is a file you own, so you can read every instruction the agent receives, change it, and run it wherever you like.

If you are still deciding what to automate at all, start with our lists of AI agent use cases and AI workflow automation examples. This post assumes you already know the job and want a working starting point.

Why Most AI Agent Templates Never Run

On 27 September I read the top results for "ai agent templates" in three search engines. The first page is mostly galleries: Microsoft's Copilot agent templates, Beam's "200+" templates and Jotform's "7,000+" examples, next to community galleries that list thousands more. Scale is the selling point.

The editorial results are thinner. The one article in the top five, from a software agency, describes three agent templates and a bonus in 1,603 words. None of the four comes with a file to download or a video of it running. A Reddit thread in the same results is titled "I tested 50+ AI agent templates for my startup. Here are the 7 that actually..." That title is the search intent in one line: people have found plenty of templates that do not work.

Trust is the other half. In the 2025 Stack Overflow Developer Survey, 87% of respondents said they are concerned about the accuracy of information from AI agents, and 81% have concerns about the security and privacy of data when using them. A template that hides what the agent does, or ships your files to someone else's cloud, fails both tests.

Teams that do ship agents name the same blocker. LangChain's State of Agent Engineering report, published in June 2026 from 1,340 responses, found 57.3% of respondents running agents in production, and quality was the top barrier for 32%. So I held the six templates below to one rule: each has a recorded run you can watch before you import it, and each shows you what the agent did.

Before you import any AI agent template, from us or anyone else, run these five checks:

1. Is there a recorded run? A video or a trace of the template working, not a screenshot of the canvas.
2. Can you read the agent's instructions? If the system prompt is hidden, you cannot predict what it will do with your data.
3. Are the credentials listed up front? A template that needs four API keys you do not have is a weekend project, not a template.
4. Does it say what happens when it cannot answer? The Drive template below says "I don't know" instead of guessing, and that is written into its instructions.
5. Can a person stop it before it acts? Anything that emails customers or writes to a ledger needs a review step.

A template you cannot watch run is a listing, not a template.

The Six Templates at a Glance

Template	The job	Key nodes	Credentials you add
Build Your First AI Agent	Turn a message into a reply, such as a messy note into bullets	Input, Agent, Output	One model
Chat with Drive Docs	Answer questions from your PDFs and quote the line it used	Agent, two Drive tools, a PDF skill	One model
Invoice Data Extraction from PDF	Turn any invoice layout into the same JSON fields	Agent, HTTP tool to LlamaParse	Model, LlamaParse key
HITL Support Reply Agent	Draft a customer reply and wait for approval	Agent with human review, Slack	Model, Slack
Browser automation with auto heal	Drive a real browser and repair broken selectors mid-run	Playwright with AI step, Agent	Model
Decision Model AI Text Humanizer	Rewrite text in rounds and keep the best-scoring version	Decision nodes, Loop, Agent	Model, Decision Model

Every video is exactly 90 seconds, the bonus included, so all seven take ten and a half minutes. Setup time depends mostly on whether you already have the credentials.

1. Build Your First AI Agent from One Sentence

The first template is the smallest AI agent that is still useful. In the video, I start from an empty canvas and type one sentence into the AI Assistant: "Summarize this note into bullets." The assistant lays out three steps, YourMessage, ChatAgent and AgentReply, and a pasted messy note comes back as clean bullets.

The Build Your First AI Agent template is that same workflow, saved. ChatAgent keeps a rolling conversation buffer across manual runs in the same session, so follow-up messages stay in context. The shipped model is GLM-4.7, and any OpenAI-compatible model works in its place.

What to change first is the system instruction. The template ships as a general assistant, and the video changes it to a summarizer. Change it to a release-note writer, a ticket triager or a translator, and you have a different agent with the same three nodes.

When it outgrows manual runs, replace YourMessage with a Telegram or Slack trigger. If you want the longer tour of what an agent node can do, our guide on how to build an AI agent covers tools, memory and multi-agent setups on the same canvas.

2. Chat with Company Files and Get the Line It Used

Every team answers the same questions every week. How many remote days do we get? What is the travel allowance? The answers sit in policy PDFs nobody wants to open. The second template lets anyone ask in a chat and get an answer they can check.

In Chat with Drive Docs, userQuery takes the question from the Chat Portal. docAgent lists your files with the listDriveFiles tool, fetches one with readDriveFile, and runs the pdf_base64_to_text skill, a small Python tool that extracts text from every page with pypdf. The answer quotes the sentence it used.

The part I like most is the second question in the video. Asked whether employees can bring a dog to the office, the agent searches all four policies, finds nothing, and says it does not know. The instruction tells it to answer only from the extracted text, and that one rule is what makes the answers checkable.

"Drive" here is Heym's own Drive tab, not Google Drive, so the files live on your Heym server. Two notes from the template: scanned PDFs without a text layer need an OCR step first, and the Drive list tool ships capped at one result, so raise the limit before you load a real folder. For the retrieval side of this pattern, see our write-up on agentic RAG.

3. Invoice Data Extraction from Any PDF Layout

Invoices arrive as PDFs and someone copies the numbers into a spreadsheet by hand. The third template does the copying. Drop in a PDF link and the same columns fill in, even when the next vendor uses a completely different layout.

In Invoice Data Extraction from PDF, InvoiceURL takes the link. InvoiceExtractor, an agent, calls llamaParseAPI, an HTTP node attached as a tool that sends the PDF to LlamaParse and gets clean markdown back. The agent returns JSON with vendor_name, invoice_number, invoice_date, due_date, subtotal, tax, total, currency and line_items, and InvoiceData passes it on.

Setup is two credentials: a LlamaParse key from LlamaIndex Cloud in the HTTP node, and a model on the agent. LlamaParse handles OCR for scanned invoices, so this template does not need the extra step the Drive template does.

To extend it, add fields such as PO numbers or tax IDs to the JSON schema in the agent's instruction, then point the output at a DataTable, QuickBooks or Xero. For volume, put a Loop in front and pass each URL through the same workflow. Our comparison of OCR and vision models for document processing explains when to use which.

4. A Support Reply Agent That Waits for Your Approval

An agent can draft every reply in a busy support inbox. The question worth asking is who decides what actually reaches the customer. In the fourth template, the answer is you.

In the HITL Support Reply Agent, CustomerRequest provides a sample escalation. DraftReplyAgent writes the reply and calls request_human_review, which pauses the run. NotifyReviewSlack posts the review link, a summary and the full draft to Slack. After you decide, ApprovedReply returns the final text and the decision.

The review page gives you three buttons. Accept sends the draft as written. Edit & Continue sends it with your changes. Refuse stops it, and the agent records that no reply was sent. Nothing reaches the customer in any path you did not approve.

The template uses a sample request so it runs on the first try. In the video, a real email starts the run, and you can do the same by replacing CustomerRequest with the IMAP trigger. Keep the agent's JSON output off, since human review works with text output. Our guide to human-in-the-loop AI agents covers where else to put a review gate.

5. Browser Automation That Heals Itself

Browser automation breaks the day a website ships a redesign. The fifth video shows the fix happening while the run is still going: an MCP client starts a Heym workflow, a Playwright node drives a real browser, and when a selector fails, auto heal finds a sturdier one and carries on.

Auto heal works on the Playwright node's AI step. When a saved step fails twice, Heym sends the page HTML, a screenshot and the failed step to your model, gets back a role or text based locator such as role=button[name='Export CSV'], and retries. The working step is saved with the node's other steps, so the next run needs no model call at all.

The locators it reaches for are the ones Playwright's own locator guide recommends. It says CSS and XPath selectors "can break when the DOM structure changes" and advises "prioritizing user-facing attributes and explicit contracts such as page.getByRole()". Auto heal applies that advice at run time instead of in a code review. It only replaces selector-based actions such as click, type, fill, hover and select, so a step that fails for another reason, like an expired login, still needs your attention.

One honest note: the workflow in this video is a demonstration built on a fictional supplier portal, so there is no one-click copy of it. The closest importable start is the Playwright Visual AI Monitor template, which screenshots a page on a schedule and alerts Slack when a vision model sees something wrong. Add an AI step with auto heal to it, and read our AI browser automation guide for cookie restore and the sandbox.

6. Watch an Agent Workflow Run, Step by Step

The last template is the one I would show a skeptic. A draft scores 0.96 on an AI-writing check. We hand it to a workflow, follow the run from the first node to the last, and it comes back at 0.27. The point of the video is less the score than the fact that you can see every step that produced it.

In the Decision Model AI Text Humanizer, scoreBefore scores the whole message with a decision model. A Loop runs up to five rounds. In each one, humanizeAgent rewrites the text five sentences at a time and checks every batch with checkAi, a Decision node attached as a tool. scoreRound scores the result, and only a lower score replaces the best version so far.

Decision models such as Jev and Laya return a computed probability instead of generated text, which is why 0.27 after a round means the same thing as 0.96 before it. Our post on System One models explains how they produce those numbers.

Everything else in the video works on any template: the Execution Log with timing and output per node, each tool call inside the agent, traces with tokens, cost and latency, automatic retries with Ask AI explaining an error, pinned data to skip a slow step, and run history you can bring back to the canvas. This is what I mean by a template you can trust.

See AI agent observability for how the traces work. Watching a run tells you what happened once. The bonus below shows how to check that it keeps happening.

Bonus: Test a Template's Prompt with Evals

A recorded run proves a template worked once, with one model and one input. Before a template goes near customers, I want to know it keeps working when the input changes or when I swap the model. That is the job of the seventh video from the same week.

Evals is its own tab in Heym. You create a suite for a workflow, such as a template you just imported, and edit its system prompt right in the suite. Optimize Prompt asks a model for a sharper version. Test cases pair an input with the answer you expect: write them by hand, or let Generate Test Data create them from the suite prompt.

Then pick several models and run them side by side, as many times per test case as you need. Scoring is Exact Match, Contains or LLM-as-Judge. The judge can be an OpenAI, OpenAI-compatible or Gemini model, or a decision model that rates each answer on a five-level rubric, from "Completely off" to "Fully matches", and scores from the probabilities it computed.

Results show pass or fail per case, the actual answer next to the expected one, and accuracy, latency and tokens per model. Every run stays in history, and Export downloads it as JSON. It is also the quickest way to decide whether a template's default model is worth changing: run the same suite on the model you have and the one you are considering, and let the numbers decide.

For the method behind it, see our guides to AI agent evaluation and LLM as a judge. The Evals docs list every setting.

Composition: AI Agent Workflow Templates as Building Blocks

Here is the idea that ties the six together. A template is most useful after you stop treating it as a finished product and start treating it as a component. Heym gives you four ways to connect workflows, and all of them are already in the product.

Entry

Chat, Portal or MCP client

A person or an outside agent starts the run

Orchestrator

Agent with sub-workflows

Decides which workflow to call through call_sub_workflow

Components

Imported templates

Invoice extraction, Drive Q&A, browser runs, each tested on its own

Gate

Human review

The orchestrator pauses before anything reaches a customer or ledger

Other ways to connect workflows

Tools

Canvas nodes as tools

HTTP, Drive or Decision nodes the agent calls when it needs them

Steps

Execute node

Calls a workflow as a fixed step, with or without waiting

Agents call workflows as tools. An agent node's Sub-Workflows setting gives it a call_sub_workflow tool for the workflows you select, and calls can nest up to five levels deep. An approval agent can call the invoice template, summarize what it found, and request human review before anything is posted to your accounting system.

Canvas nodes become tools. Any supported node connected to an agent's tools handle becomes something the agent can call, which is exactly how the invoice template's HTTP node and the humanizer's Decision node work. The Execute node covers the other case, when a workflow should always run as a fixed step.

Workflows become MCP tools. Turn on MCP for a workflow and Claude, Cursor or any MCP client can call it, which is how the browser automation video starts. Heym Chat composes in the same spirit: the first-agent video ends with Chat running the Brave Search API and Website Loader templates to research the web before it builds.

The practical rule I follow: keep each workflow small enough to test on its own, then compose. When something goes wrong, the trace tells you which component failed, and you fix one template instead of untangling one giant workflow. For more patterns, see agentic design patterns and our explainer on MCP.

Self-Hosted AI Agent Templates: Free and Source-Available

People searching for open source AI agent templates, or for AI agent templates on GitHub, usually want three things: no cost, code they can read, and the option to run it themselves. Here is where the six stand, plainly.

What Heym is: Heym is a source-available, self-hosted AI workflow automation platform with a visual canvas for multi-agent pipelines, RAG and MCP. Everything this post says about the six templates comes from the shipped templates and documentation. Heym Work, in the last section, is the one roadmap item, and it is named, not described.

The templates are free JSON files. The platform they run on, Heym, is source-available under the MIT license with the Commons Clause, which lets you use, read, modify and self-host it at no cost, but not resell it as a commercial product. The code lives on GitHub, where you can read exactly what an agent node sends to a model. Most AI agent templates on GitHub are code you build on. These are finished workflows you import.

One caveat if you need open source AI agent templates in the strict OSI sense: the Commons Clause means Heym's license is not OSI approved. And unlike prebuilt AI agents that live on a vendor's cloud, these templates are files you keep.

Self-hosting is what makes self-hosted AI agent templates worth the extra setup. The support emails, policy PDFs and invoices they process never leave your server, except for the model calls you choose to make. Point the credentials at a model you host yourself and even those stay inside. Our article on self-hosted AI agents covers how to make a local instance reachable from outside when you need webhooks.

Getting a local instance running takes a few minutes with Docker; the quick start has the commands. Then every template in this post is a drag and drop away.

Nearly 1,300 Stars, and What Comes Next

Something personal before the FAQ. When I checked this morning, the Heym repository had 1,289 stars, 96 forks and 24 contributors. It was created on 28 March 2026. In the seven days to 26 September, 62 people starred it, and every one of those notifications still makes our day.

The contributions mean even more. Since 28 August, three people outside our team got 11 pull requests merged. Two landed this week: one that resolves shared resources for users in multiple teams, and one that uses the highest share permission for data tables. Bugs we had not hit yet, found and fixed by people we have never met. Thank you.

We are a small part of a much bigger wave. GitHub's Octoverse 2025 counted monthly contributors to generative AI projects rising from 68,000 in January 2024 to 200,000 by August 2025. More people want to build agents in the open, and more of them want to run those agents themselves.

That belief is behind what we are building next. It is called Heym Work, and it is coming soon. I am not going to describe it before it ships, but it comes from the same place as everything in this post: in the agentic era, the agents that do your work should run on infrastructure you own. If you want to hear first, star the repository and keep an eye on the changelog.

Until then, pick one template from the table above, import it, and run it once today. Then browse the rest of our AI agent templates, or all 195 in the full template gallery, for the next one. If you build something worth sharing, submit it from the gallery so the next person starts from a run instead of a guess.

Frequently Asked Questions

What is an AI agent template?

An AI agent template is a saved, working workflow built around an AI agent: the trigger or input, the agent with its instructions and tools, and the output, already wired together. You import it, attach your own credentials, and run it. It differs from a framework, which gives you code to build an agent, and from a prompt template, which is only the text you send to a model.

Are these AI agent templates free?

Yes. Every template in the Heym gallery is free to download as JSON and import into your own Heym instance. You pay only for what the workflow calls: your model provider, and in the invoice template, a LlamaParse key. Heym itself is source-available under the MIT license with the Commons Clause, so you can self-host it at no cost for your own use.

Can I self-host AI agent templates?

Yes. A template is a JSON file, and it runs wherever Heym runs, including a laptop, a VPS or a server inside your company network. The documents, invoices and support emails the templates handle then stay on infrastructure you control, and only the model calls leave it, unless you point the credentials at a model you host yourself.

How do I import an AI agent template into Heym?

Open the template page on heym.run, download or copy the JSON, then drag the file onto the Heym canvas. The workflow appears with every node and connection in place. Open each agent node, choose a credential, and run it once with the sample input before you change anything.

Can one AI agent template call another?

Yes. An agent node can call other workflows as tools through its Sub-Workflows setting, up to five levels deep, and the Execute node calls a workflow as a regular step. Workflows can also be exposed to Claude, Cursor or any MCP client as tools, and Heym Chat can run your existing workflows while it works. That is how small templates become larger systems.

How do I test an AI agent template before I rely on it?

Run its system prompt through an eval suite. In Heym Evals you pair test inputs with the answers you expect, or let a model generate them, then run several models side by side and score with Exact Match, Contains or LLM-as-Judge. The judge can be an OpenAI, OpenAI-compatible or Gemini model, or a decision model that scores each answer on a five-level rubric.

Which AI agent templates actually work?

The ones you can watch run before you import them. Check for a recorded run or trace, readable agent instructions, a clear list of credentials, a defined answer for when the agent does not know, and a human review step before anything reaches customers. Each of the six Heym templates in this article has a 90-second video of it working.

Which models do the templates need?

Any OpenAI-compatible chat model works for the agent nodes, and you pick it on the credential. The humanizer template also needs a Decision Model credential, such as Jev, or Laya running on your own hardware. The browser automation AI step needs a vision-capable model if you want it to read screenshots.

Sources

• Stack Overflow, 2025 Developer Survey: AI, 2025. Concern about the accuracy of AI agent information (87%) and about data security and privacy when using agents (81%).
• LangChain, State of Agent Engineering, June 2026. 1,340 respondents, 57.3% with agents in production, quality as the top barrier for 32%.
• Playwright, Locators, read 27 September 2026. The recommendation to prefer user-facing locators over CSS and XPath.
• GitHub, Octoverse 2025, October 2025. Monthly contributors to generative AI projects.
• Heym, GitHub repository, read 27 September 2026. Stars, forks, contributors and merged pull requests.
• Heym documentation: Agent node, Playwright node, Human-in-the-loop, Download and import, Evals.

Steps at a glance

1. Pick a template by the job, not the name. Write down the one job you want off your plate this week, such as answering policy questions or copying invoice totals. Pick the template whose input and output match that job, and watch its recorded run before you import it.
2. Import the JSON onto your canvas. Download the template JSON from its page on heym.run and drag it onto the Heym canvas. Every node, edge and setup note arrives in place, including the sticky notes that list what to configure first.
3. Attach your own credentials. Open each agent node and choose a model credential. Add the extra credentials the template names, such as Slack for human review, LlamaParse for invoices, or a Decision Model credential for the humanizer.
4. Run it once with the sample input. Run the workflow exactly as shipped before changing anything. If the sample run works, any later failure comes from your change, which makes debugging a question of one diff instead of guesswork.
5. Change one thing and run again. Swap the sample input for your real data, or edit the system instruction for your domain. Change one thing per run and compare the Execution Log and traces between runs.
6. Add a check before it touches customers. Before a template writes to customers or systems of record, add human review on the agent and a small eval suite for its prompt. Both are built in, and both take minutes to add.