Prepare a repository

Futur Pad imports public GitHub repositories. The repository holds the instructions that define the agent. Skills are files named exactly SKILL.md, anywhere outside build and vendor folders, and should start with frontmatter:

---
name: code-review
description: Reviews a pull request and lists concrete fixes
---

# Code review

Read the diff first, then list issues by severity.

Also read: AGENTS.md, CLAUDE.md, SOUL.md, GEMINI.md, Cursor and Windsurf rules, Copilot instructions (up to two folders deep), MCP configs (.mcp.json, mcp.json, .cursor/mcp.json, .vscode/mcp.json), subagents in .claude/agents, commands in .claude/commands, a Claude plugin manifest and OpenClaw configuration. Up to 25 skills are read per agent.

Import it

Sign in with a browser wallet, open Launch and paste the repository. Futur Pad reads its branch, licence, stars, contributors, recent commits and the files above at the latest commit, then creates a private draft. You can pick a display name; the repository's owner, name and description stay visible next to it. Nothing is forked or written to GitHub.

Health check and test

Two real checks are required before publishing. The health check starts the preview with the repository's files and asks the agent what it does; its answer, host and latency are stored. The test message is a normal conversation from the owner; it must get an answer. Test runs are kept out of the public history.

Published agents are checked again once a day. An agent shows online only while its last passing check is under 36 hours old; otherwise it shows unchecked or offline.

The preview runtime

The preview sends the repository's instruction and skill files to a hosted model as instructions, within a fixed budget. It does not execute code, install dependencies, open files or call the tools and MCP servers a repository defines. The runtime used for each answer is shown next to it and in the run history. When one host is busy the next one answers.

Run policies

Every agent has a policy its owner can change: runs per person per hour (default 20), runs for the whole agent per hour (default 300), the longest answer in tokens (default 900), and whether others may fork it. The policy is checked before the model is called, for the website and the API alike. The owner is not limited on their own agent.

Runs and receipts

Every conversation is a recorded run with status, duration and runtime; those three are public. The content is stored for 30 days and readable only by the participant. To show a run to someone, use share receipt under an answer: it creates a read-only link to that one run. Revoke deletes the link at once; the run itself is unaffected.

Forks

Any published agent whose owner allows it can be forked. A fork may keep the parent's repository or point at your own copy, so you can change its skills. It starts as a draft, must pass its own checks, and its page shows which skills it added or removed compared with the parent.

Launch on Pons v2

An owner can link a Pons v2 token (Robinhood Chain) to a published agent. Two steps:

1. Ownership. Commit futurpad.json to the default branch with {"wallet": "0x…"}, the wallet that owns the agent. Futur Pad reads it from GitHub. Only people with write access can commit it, so this proves control of the source without an OAuth app.

2. Launch. Launch the token from that wallet on Pons, then paste the transaction hash on the agent page. Futur Pad reads the receipt from Robinhood Chain, requires a successful status, decodes the TokenLaunched event from the Pons v2 factory, and checks that its deployer is the wallet from step 1. Only then is the launch shown. Price, FDV, liquidity and volume come from DEX Screener once the token trades.

Public API

No key is needed. Responses are JSON with CORS open.

GET  https://www.futurpad.xyz/api/v1/agents
GET  https://www.futurpad.xyz/api/v1/agents/<slug>
POST https://www.futurpad.xyz/api/v1/agents/<slug>/run
     {"input": "…"}  or  {"messages": [{"role": "user", "content": "…"}]}

→ {"id": "run_…", "agent": "<slug>", "commit": "…", "runtime": "…", "ok": true, "output": "…"}

API runs are counted per caller address under the agent's policy. Over the limit you get HTTP 429 with a retry-after header.

Honest limitations

Not implemented: private repositories, running repository code or MCP servers, bring-your-own model keys, and launching from inside Futur Pad (the launch happens on Pons and is verified here). Repository details refresh at most every 30 minutes when a page is opened; the files behind a preview change only when its owner rebuilds it. Where a value cannot be read, the interface says unavailable.