AgentForge

org.sciscalePublished by sciscale.orgv0.1.4Updated Oct 24, 2026

Turn rough requests into rigorously structured prompts, for any coding agent.

VerifiedSTDIODesktop onlyDeveloper ToolsAI & ML

Overview

AI-generated overview

AgentForge turns a rough request into a structured, tool-tuned prompt for coding agents, via a hosted refinement engine.

What you can do
  • Refine a plain-language request into a structured prompt for a chosen coding agent (Claude Code, Codex, Cursor, Aider, Continue, Windsurf, Kimi or generic)
  • Choose a prompt style such as plan-first, direct-edit or explore-first
  • Get back the refined prompt together with its Quality Engine score and remaining daily usage
When to use it
Useful when you repeatedly re-explain tasks to a coding agent and want one rough sentence turned into a prompt with explicit requirements, edge cases and acceptance criteria, formatted for your specific agent.
Requirements
Runs locally over stdio via npx, so Node.js is needed. Requires an AgentForge account and an API key supplied in the AGENTFORGE_API_KEY environment variable; AGENTFORGE_API_URL can override the endpoint. Network access to the hosted API is required.
Before you install
The package is a thin client: your request text is sent to the hosted AgentForge service for processing. The API key is a secret passed in AGENTFORGE_API_KEY and is shown only once when created. The free tier is limited to 3 refinements per day; unlimited use and auto-refinement require a paid Pro plan.

Installation

In SourceWeft

  1. Open AgentForge in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

[AgentForge]

agentforge-mcp

Stop re-explaining what you want to your AI coding agent.
Hand it one rough sentence — get back a structured, tool-tuned prompt it can execute in one pass.

[npm] [license]

agentforge.sciscale.org · Quick start · npm

[AgentForge MCP in action]

agentforge-mcp is the Model Context Protocol server for AgentForge. One tool, one job: turn a vague request into a sharp one.

Why

You know the loop: you ask your coding agent for something, it misreads the half you didn't spell out, you correct it, it breaks something else, you re-explain. The fix isn't a smarter agent — it's a sharper prompt.

AgentForge does the prompt engineering for you. Give it "add a dark mode toggle that persists" and it:

  • extracts the real requirements — the edge cases, the acceptance criteria, the things you'd have forgotten to mention;
  • formats for your specific agent — Claude Code wants file:line references, Cursor wants directives, Aider wants tight context;
  • quality-checks before returning — every prompt is scored across 12 dimensions before it leaves the engine.

Harness engineering is having a moment — but prompt engineering matters more in the coding-agent era, not less. Tested across 1,000+ real coding cases, projects with AgentForge in the loop reached a deployable state 147% faster.

Quick start

No install — it runs through npx.

1. Get an API key. Sign in at agentforge.sciscale.org, open API keys, and create one. The key is shown once — copy it.

2. Add it to your agent.

Claude Code:

sh
claude mcp add agentforge --env AGENTFORGE_API_KEY=af_your_key -- npx -y agentforge-mcp

Cursor / Windsurf / Claude Desktop — add to your MCP config (~/.cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, etc.):

json
{  "mcpServers": {    "agentforge": {      "command": "npx",      "args": ["-y", "agentforge-mcp"],      "env": { "AGENTFORGE_API_KEY": "af_your_key" }    }  }}

3. Use it. Ask your agent naturally — "refine this with AgentForge, then build it: add a dark mode toggle that persists."

The tool — agentforge_refine_prompt

ArgumentDefault
request—Your task in plain language (1–4000 chars). Rough is fine.
target_toolclaude-codeclaude-code, codex, cursor, aider, continue, windsurf, kimi, generic
styleplan-firstplan-first, direct-edit, explore-first

Returns the refined prompt, plus its Quality Engine score and your remaining daily usage.

Free vs Pro

FreePro
Refinements3 / dayUnlimited
Quality Enginescored across 12 dimensionsscored + auto-refined until it passes ≥ 90

Or use the web app

Don't want to wire up an MCP server at all? The same engine — same Quality Engine, same per-account Pro — runs right in your browser at agentforge.sciscale.org. No install, no key, no config.

[The AgentForge web app]

Paste a request, pick your target tool, copy the prompt — generation history and advanced modes included:

[The web app in action]

How it works

agentforge-mcp is a thin client — no engine logic ships in this package. Your request goes to the AgentForge API, the hosted engine does the extraction, formatting, and quality-checking, and the prompt comes back. The engine keeps improving without you ever updating this package.

Configuration

VariableDefault
AGENTFORGE_API_KEY—Required. Your API key.
AGENTFORGE_API_URLhttps://agentforge.sciscale.org/api/v1/refineOverride the endpoint (rarely needed).

Develop

sh
npm installnpm run build      # tsc -> dist/node dist/index.js # runs on stdio

License

MIT — see LICENSE.


[SciScale studio]

agentforge-mcp is part of AgentForge — a SciScale studio product.

Source: README.md at commit 204381b

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.4LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts