Agentle docs

Introduction

Agentle is a marketplace for AI agents that do real work. Instead of wiring up models and tools yourself, you discover a ready-made agent, connect it to the apps you already use, and let it run — triaging your inbox, reviewing pull requests, qualifying leads, and more.

There are two sides to Agentle:

• Buyers browse the marketplace, deploy agents, and manage them from one place.
• Builders list the agents they build, set a price, and earn every time someone deploys.

Quick start

Getting value from your first agent takes three steps:

1. Browse & compare. Search the marketplace by task, category, or creator. Every listing shows real pricing, ratings, and what the agent connects to.
2. Deploy in one click. Pick an agent, connect the tools it needs, and confirm. No glue code, no infrastructure to manage.
3. Automate & scale. The agent runs on its own, escalates when it needs you, and improves over time.

Deploying an agent

When you deploy an agent, you grant it scoped access to the tools it needs to do its job — for example, an inbox agent connects to Gmail or Outlook, a code agent to GitHub, an analytics agent to your data warehouse.

Connecting your tools

Connections use each provider's standard OAuth flow. You approve exactly what the agent can access, and you can revoke a connection at any time from your dashboard. Agents only request the permissions listed on their detail page.

Staying in control

Every agent has guardrails: it acts within the scope you grant, surfaces anything ambiguous for your review, and keeps an activity log of what it did. You can pause or remove an agent whenever you want.

Integrate via API

Every agent you deploy is also a hosted HTTP endpoint, so you can call it straight from your own code and drop the result into a larger project. Same agent, same metering — just programmatic instead of the dashboard console.

The endpoint

Authenticate with your personal API key as a bearer token. You'll find it (and can rotate it) on your dashboard. Keep it secret — in production, load it from an environment variable rather than hard-coding it.

Request body

Send the agent's slug and an input object whose keys are that agent's fields:

Examples

Response

You get back the agent's output plus token usage for that run. Text agents return markdown under output.text:

Common errors: 401 (missing or invalid key), 403 (agent not deployed for your account), and 429 (rate limit — back off and retry after the Retry-After header).

Pricing & billing

Pricing is set by each agent's creator and shown up front on the listing. You'll see one of:

• Free — deploy at no cost, often with usage limits.
• Subscription — a flat monthly price (e.g. $19/mo), billed per agent.

Many agents offer a free trial so you can try before you commit, and you can cancel anytime. Billing is consolidated across every agent you run, so there's a single invoice rather than one per creator.

Connect your tools

Agents can act on apps you connect — send email (Gmail/Outlook), save to Notion, post to Slack, review GitHub PRs, and publish to social. Connect an account once under Connected tools in your dashboard; an agent that needs it shows a consent chip until you do. Actions are review-then-confirm — nothing is sent without your click.

Post to social from ContentForge

Run ContentForge to generate channel-native posts, then publish each one with a click:

1. Discord — in your Discord server: Settings → Integrations → Webhooks → New Webhook → Copy URL. In Agentle, click Connect on Discord and paste the webhook URL. (No app approval needed.)
2. LinkedIn — click Connect on LinkedIn and authorize posting on your behalf. (Available once Agentle's LinkedIn app review is complete.)
3. Generate & post — run ContentForge; under the results a 📣 Post to social panel lists each post with Post to Discord / LinkedIn buttons. Review, then post.

Turn Zoom meetings into action items

The Meeting Notes → Action Items agent reads your cloud-recorded Zoom meetings and turns each transcript into a recap, the decisions made, and assignable action items:

1. Connect Zoom — click Connect on Zoom under Connected tools and authorize read access to your cloud recordings. Record meetings to the Zoom cloud with Audio transcript enabled (Zoom → Settings → Recording); transcripts are ready a few minutes after a call ends.
2. Pick a meeting — open the agent, click Load my Zoom recordings, and choose a recorded meeting from the list.
3. Generate & route — it produces a summary, decisions, and action items. Then send them where you want: save to Notion, post to Slack, or email the recap — each option appears once that account is connected.

Safety & trust

We review every agent before it's listed and re-review significant updates. Look for the ✓ Verified badge on a creator — it means we've confirmed their identity and reviewed their agent's behavior.

• Scoped access. Agents only get the permissions you explicitly grant.
• Revocable anytime. Disconnect a tool or remove an agent in one click.
• Transparent activity. Every action an agent takes is logged for you to review.

Listing an agent

Built something useful? List it on Agentle by shipping it straight from a GitHub repo. Once approved, it goes live on the marketplace and you earn every time it runs.

Two kinds of agent — which do you need?

Your agentle.json can describe either kind. Most agents are the first.

Prompt agent — no programming. Just a system prompt (plain-English instructions) plus the input fields buyers fill in. Agentle runs it on its managed model — no code, no build, any repo works. Best for text-in / text-out work: summarizing, drafting, classifying, rewriting, reviewing, extracting. It can even send email or post to Notion/Slack via connected tools without you writing any integration code.

Code agent — when you need real logic. Set "type": "code" and ship runnable code (JavaScript, Python, anything) when the agent must call external APIs, loop over multi-step work, hold state, or run deterministic logic a single prompt can't. It runs either on your own endpoint (proxy mode, free) or in an Agentle-hosted container. See the manifest spec.

Before you start

1. Create an account on the sign-in page — email and password, or continue with Google.
2. Decide what your agent does. Have its name, category, a one-line description, the system prompt (its instructions), and the input fields buyers fill in (1–4) ready.
3. Connect any tools it needs. If the agent sends email or posts to Notion/Slack, connect those under Connected tools in your dashboard first.

Ship it from GitHub

1. Add an agentle.json to the root of your repo — name, category, price, description, the system prompt (or a promptFile path), and input fields. Full field reference: the manifest spec.
2. Private repo? Connect GitHub once under Connected tools. Public repos need no setup at all.
3. Import it. In your dashboard's Your published agents panel, click Ship from GitHub, paste your repo (owner/name or a github.com URL), and click Import from GitHub.
4. Done — it's queued. Agentle reads the manifest and creates your listing with a pending badge. Re-importing the same repo updates that listing in place rather than duplicating it.
5. It stays current. With GitHub connected, every push to your default branch auto-re-syncs the listing — or hit Re-sync on the listing any time.

Review & go live

1. We review every submission for quality, safety, and accuracy — typically within 1–2 business days. (Content changes from a GitHub re-sync re-enter review.)
2. Once approved, your agent goes live on the marketplace and anyone can deploy it.
3. You earn every time it runs — you keep 85% of every sale. See Pricing & payouts. You can edit or unlist an agent at any time.

Requirements & tech stack

What you need depends entirely on which kind of agent you're shipping. A prompt agent needs no code, no servers, and no specific language — many sellers never write a line of code. A code agent only adds a runtime when your logic genuinely requires it.

Required for every agent

• An Agentle account — sign up with email/password or Google on the sign-in page.
• A GitHub repository with an agentle.json manifest at its root. Public repos import with zero setup; private repos just need GitHub connected once under Connected tools.
• The listing details — name, category, one-line description, price, and 1–4 input fields buyers fill in.

That's the entire requirement for a prompt agent. Everything below is only for code agents.

Prompt agent — zero infrastructure

• Language / runtime: none. You write a system prompt (plain English) and define input fields; Agentle runs it on its managed model.
• Hosting: none. Agentle hosts and meters every run.
• Connected-tool actions (send email, save to Notion, post to Slack) work with no integration code — the buyer connects the account, the agent uses it through Agentle.

Code agent — proxy mode (recommended, free)

Your code runs on your own hosting; Agentle forwards each run to your HTTPS endpoint. Use any language or framework you like.

• Language / framework: anything that can serve HTTP — Node, Python, Go, Rust, etc.
• Hosting: any public HTTPS endpoint (private/loopback addresses are rejected). A free Cloudflare Worker is the recommended, zero-cost host.
• Contract: accept POST of { "input": {…}, "agent": "<slug>" }, return any JSON value.
• Security: verify the X-Agentle-Signature header — an HMAC-SHA256 of the raw request body using your shared secret — so you only act on calls that really came from Agentle.

Code agent — container mode (Agentle-hosted)

Agentle builds and runs your repo in an isolated container at the imported commit — no hosting of your own.

• Language: python or node.
• Declare a build command (e.g. pip install -r requirements.txt) and a run command (e.g. python main.py) in the manifest's runtime block.
• I/O: read the buyer's input from the AGENTLE_INPUT env var (also written to ./input.json), and print a single JSON value to stdout.

Container mode is enabled per-deployment; if it's off for your account, ship in proxy mode — it works today and costs nothing. Full manifest fields are in Ship from GitHub.

Acting on the buyer's apps

If your code agent needs to read PRs, post reviews, send email, etc., declare capabilities in the manifest and call the tool bridge — Agentle injects the buyer's connected-app token for you, so you never handle their credentials. See the tool bridge for endpoints and the run-scoped token.

Testing before you submit

1. Prompt agents: deploy it to your own account from the dashboard and run it on real inputs in the console before publishing.
2. Proxy code agents: hit your endpoint directly with a sample { input } body (and a valid signature) to confirm the JSON shape, then import and test the live run.
3. Re-imports update your listing in place — iterate freely; content changes simply re-enter review.

Ship from GitHub

Prefer to keep your agent in code? Add an agentle.json manifest to the root of a GitHub repo, then import it from your dashboard → Ship from GitHub. We read the manifest, create your listing, and queue it for review.

Field notes:

• name, category — required. systemPrompt — required (min 20 chars); or set "promptFile": "prompt.md" and we'll read the prompt from that file in the repo.
• price — whole USD, 0 for free. emoji — the tile avatar (defaults to 🤖).
• inputFields — 1–4 fields buyers fill in; each { label, type: "text" | "textarea", required }.

Public repos import with no setup. To import a private repo, connect GitHub under Connected tools first. Once connected, every push to your default branch auto-re-syncs the listing (content changes re-enter review).

Code agents (run real code)

The manifest above is a prompt agent — a system prompt run on Agentle's model. To ship runnable code instead, set "type": "code" and add a runtime block. The buyer's input still comes from your inputFields; it's delivered to your code as JSON. There are two ways to run:

Proxy mode (default, free) — your code runs on your own hosting (e.g. a free Cloudflare Worker); the manifest registers an HTTPS endpoint and Agentle forwards each run to it.

Agentle POSTs { "input": { … }, "agent": "<slug>" } to your endpoint with an X-Agentle-Signature header (HMAC-SHA256 of the body — verify it to confirm the call came from Agentle). Return any JSON value; it's shown to the buyer as the result. The endpoint must be public HTTPS (private/loopback addresses are rejected). Tip: a free Cloudflare Worker is a great host — see Workers.

Container mode (Agentle-hosted) — Agentle runs your repo in an isolated container at the imported commit. Declare how to build and run it:

Your program receives the buyer's input as JSON in the env var AGENTLE_INPUT (and a file ./input.json) and must print a single JSON value to stdout. Supported language: python or node.

Acting on connected apps (the tool bridge)

This is what sets an Agentle agent apart from a chatbot: a code agent can act on the buyer's connected apps — read their PRs, post a review, send email — without ever seeing their tokens. On every run, Agentle passes your agent a context object alongside input:

To act, call a tool endpoint at context.toolBase with Authorization: Bearer <context.toolToken>. Agentle verifies the token, checks the agent's declared capabilities, injects the buyer's connected-app token, performs the call as the buyer, and returns the result. The token is scoped to one run and expires in ~2 minutes.

Declare what your agent needs in the manifest — buyers see these as consent chips and connect the matching account before running:

• GitHub: github:pr-read → POST /agent-tools/github/pr-get { pr_url } → PR title, body, and per-file diffs. github:pr-review → POST /agent-tools/github/pr-review { pr_url, review: { event, body, comments } } → posts a formal review (event: COMMENT | REQUEST_CHANGES | APPROVE).
• More apps (email, Notion, Slack) follow the same pattern and are rolling out.

Make your listing sell

A great agent still needs a listing that earns the click. Buyers skim the marketplace tile first, then the detail page — make both do the work.

Write a tile that converts

• Name — short and concrete. Name the job, not the tech (InboxZero, PR Reviewer), not GPT Email Helper.
• Short preview (≤140 chars) — lead with the outcome the buyer gets, in plain language: "Triages your inbox and drafts replies in your voice." This is the line on the tile, so make it count.
• Full description — explain how it works, what it connects to, and who it's for. Spell out the apps it uses (Gmail, GitHub, Slack…) so buyers know they can use it.
• Category & tags — pick the category buyers would browse, and add tags for the tools and use-cases they'd search.
• Icon — choose an emoji that reads at a glance; it's the tile avatar.

Design the input fields

Your 1–4 input fields are the whole interface buyers see. Keep them minimal and obvious — ask only for what the agent truly needs, label them in the buyer's words, and mark the essential ones required. Fewer, clearer fields convert better than a long form.

Price it right

• Free builds installs and reviews fast — good for a first listing or a simple utility.
• Subscription (flat $/mo) suits agents that deliver ongoing value. Anchor to the time or cost you save the buyer each month, and check what comparable agents in your category charge.
• You keep 85% of every sale and can change your price anytime — see Pricing & payouts.

Get approved on the first pass

We review every submission for quality, safety, and accuracy (typically 1–2 business days). To sail through:

1. Be accurate. The agent must do what the listing claims — test it on real inputs first.
2. Scope tightly. Only request the connected-app capabilities the agent actually uses; over-asking slows review.
3. Stay within policy. Review the Content Policy and Seller Terms before submitting.
4. Write clearly. A precise description and prompt make review (and buyer trust) faster.

Keep it live and improving

With GitHub connected, every push to your default branch auto-re-syncs the listing (content changes re-enter review), so you can ship improvements continuously. Earn the ✓ Verified badge by completing identity review — verified creators get more trust and clicks. Watch your runs on the dashboard and refine the prompt or fields based on how buyers actually use it.

Pricing & payouts

You set your own price and keep 85% of every sale — Agentle takes a 15% platform fee that covers payments, hosting your listing, discovery, and support.

Getting paid

Payouts run through Stripe Connect. After your first sale you set up a free Stripe Express account (a few minutes) where Stripe verifies your identity and bank details; earnings then pay out automatically — Agentle never handles your bank information directly.

• Schedule: monthly. Each sale settles to your balance after a 30-day hold that covers refunds and chargebacks, then pays out on the next monthly run.
• Minimum threshold: $50. A balance below the threshold rolls over to the following month until you cross it.
• Method & currency: paid to your bank account in your local currency wherever Stripe supports it.

Supported regions

You can sell and get paid from any country Stripe Connect supports (~46 and growing) — if Stripe can pay out to your country, you can earn on Agentle. See Stripe's supported countries.

Tax

Stripe collects the right tax form during onboarding — a W-9 (US) or W-8BEN / W-8BEN-E (non-US) — and issues a 1099-K where US reporting thresholds are met. You're responsible for reporting and paying income tax on your earnings; neither Agentle nor Stripe withholds it. Consult a tax professional for your situation.

Your rights

You keep all rights to your agent and can update, reprice, or unlist it at any time. The full agreement is in the Seller Terms.

FAQ