Multi-agent orchestration
Your top-level agent provisions persistent subagents, dispatches prompts, monitors progress, and feeds results back into its own loop. Each subagent has its own isolated state.
Hand one skill file to your AI agent. It learns to spin up sandboxed Claude Code, Codex, OpenCode, or Cursor subagents on demand — with your own custom skills installed and a real Linux runtime to work in.
No signup. No SDK. No subscription. Your agent handles the rest.
Three steps. The first two take 30 seconds.
Drop the agentmint skill file into your agent's skill registry, system prompt, or /skills folder.
“Spin up a Claude Code subagent and give it skills X, Y, Z.” Your agent reads the file and figures out the rest.
Each prompt to a subagent is a small pay-as-you-go charge. Your agent will tell you when its wallet needs a top-up.
A standalone runtime your agent can rent, configure, and dispatch work into.
Your top-level agent provisions persistent subagents, dispatches prompts, monitors progress, and feeds results back into its own loop. Each subagent has its own isolated state.
Pass any GitHub-hosted skill (owner/repo) at create time, or add more later. Subagents load them automatically and use the documented capabilities.
Full isolated runtime. Node / Python / Go / Ruby / Rust. Subagents can install packages, write files, run shell, fetch from the network, persist state across calls.
Claude Code, OpenAI Codex, OpenCode, Cursor — pick the harness and model when you create a subagent. Switch models later with one call.
Provision once, then pay per prompt. Hibernates for free when idle. No monthly minimum, no per-seat seats, no quota you'll forget about.
Bring your own AI provider key (cheaper) or let AgentMint handle it (slightly more, no key management). Mode chosen per subagent at create time.
Paste this into your agent's skill folder, system prompt, or context. That's the whole onboarding.
---
name: agentmint
description: Pay-as-you-go subagents. Provision a persistent Claude Code / Codex / OpenCode / Cursor subagent on demand, then send prompts and pay-per-call. Includes custom skills + a real Linux runtime.
license: MIT
metadata:
endpoint: https://api.agentmint.store/a2a
agent_card: https://api.agentmint.store/.well-known/agent-card.json
---
# agentmint
AgentMint lets your agent provision persistent **subagents** on demand. You pay once to create a subagent (binds harness, model, optional API key, skills), then pay-per-prompt to talk to it.
## When to use
- The user wants you to spin up a sandboxed Claude Code / Codex / OpenCode / Cursor subagent that can run shells, install packages, edit files, and persist state across calls.
- The user wants a workflow customized via skill files (e.g. give the subagent skills `moltycash/payment` and `your-org/your-skill`).
- The user wants pay-as-you-go isolated compute, not a subscription.
## Endpoint
`POST https://api.agentmint.store/a2a` — JSON-RPC 2.0.
Discovery: `GET /.well-known/agent-card.json`. Per-method specs: `/skills/agent-create/SKILL.md`, `/skills/agent-run/SKILL.md`, `/skills/agent-update/SKILL.md`.
## Method summary
| Method | Cost | Auth | Purpose |
|---|---|---|---|
| `agent.create` | one-time | payer becomes owner | Provision a persistent subagent. |
| `agent.run` | per call | payer = owner | Send a prompt; receive reply. |
| `agent.update` | dust | payer = owner | Update skills / init / model. |
| `agent.cancel` | dust | payer = owner | Cancel a stuck run. |
| `agent.runs` | dust | payer = owner | List recent runs. |
| `agent.get` | dust | payer = owner | Read agent metadata. |
| `agent.delete` | dust | payer = owner | Destroy the subagent. |
## Two modes (pick at agent.create — permanent)
- **byok** (default if you pass `api_key`) — you pick harness + model, supply your AI key; your provider bills tokens.
- **all-inclusive** (default if no `api_key`) — AgentMint picks harness + model + key; AI tokens covered.
## Walkthrough
1. **Create.** Send `agent.create` with your harness/model. Save the returned `agent_id`.
2. **Run.** Send `agent.run` with `agent_id` + `prompt`. Pay the advertised amount. Read the `output` field.
3. **Iterate.** Same `agent_id`, more prompts. Filesystem state persists across calls. Subagent hibernates when idle (free).
4. **Customize.** `agent.update` to add skills (`skills_add: ["owner/repo"]`) or swap the model.
5. **Unstick.** If a prompt hangs, `agent.cancel` clears the in-flight run.
6. **Tear down.** `agent.delete` destroys the subagent.
Or fetch the live version at agentmint.store/skills/agentmint/SKILL.md.
No subscription, no per-seat. You pay when your agent runs a subagent.
Binds the configuration. Owner = the wallet that pays. Mode is permanent — pick at create.
You pick harness + model and supply your AI provider key. Your provider bills you for tokens.
AgentMint handles the AI side. Per-call price drifts up or down across the next few calls based on actual usage — short replies cost less, long ones cost a bit more.
For agent.update, agent.cancel, agent.runs, agent.get, agent.delete.
Subagents hibernate when not in use. No idle cost. Resume on the next prompt.
Settlement is handled by your agent's wallet. When the wallet runs low, your agent will receive a top-up instruction in the next call's response — pass it on to the user verbatim.
Copy the skill, paste it into your agent's context, and let it figure the rest out.
Copy the skill →