The Framework

Autonomous AI programming.
Stop babysitting your coding agents. You make the important decisions; AI does the rest.

The Framework turns your coding agent into an autonomous teammate that plans, researches, implements, and reviews its own work, while you keep control of the decisions that matter. It wraps a coding-agent CLI (Claude Code today) as a black box and takes you from an idea to a running app. A localhost dashboard shows what the agent's chat cannot: loop status, review passes, and the queue of what is next.

100% open source, 100% local, on your existing AI subscription (bring your own Claude Code).

Quickstart

npm i -g @gemstack/framework
framework "a paginated orders page backed by an orders table, with sign-in"

# one-shot, no install:
npx @gemstack/framework "..."

# deterministic offline demo (no model, no API key):
framework --fake

Stop babysitting

Left alone, AI is lazy, forgetful, and quietly makes decisions it should be asking you about. The Framework fixes that at the prompt layer:

• AI is lazy. It splits large tasks into a checklist of subtasks and works through them, instead of taking shortcuts.
• AI plans lazily. A loop of critical feedback and research runs before any code is written.
• AI writes throwaway code. It reviews its own work until the result is correct, maintainable, and DRY.
• AI decides without asking. When real alternatives with trade-offs exist, it surfaces them instead of silently picking one.
• AI forgets. Decisions and context are saved to knowledge-base/*.md in your repo, so you stop repeating yourself.

How it works

The Framework does not run its own agent. It drives a coding agent as a black box: sends a prompt, lets the agent's loop run, reads the code produced, gates on the outcome (builds, serves, review passes), then re-prompts. The seam is the code, never the agent's individual tool calls, so the wrapped agent keeps its subscription auth and stays swappable.

• Enhanced system prompt. Best-practice instructions layered on your agent: divide and conquer, enumerate before coding, self-review, surface alternatives. Fully customizable, or opt out.
• Autopilot. Pick how much runs unattended. Quick wins and quality refactors can run on their own; non-obvious decisions still come back to you.
• Dashboard. A localhost view of the live run: loop status, review passes, the queue of upcoming tasks, the reviews waiting on you, and your subscription usage.
• Bring your own subscription. Orchestrates through your existing Claude Code install, so there is no extra API bill.
• 100% local. Runs on your machine like a desktop app. No backdoors, no tracking; your knowledge base lives as knowledge-base/*.md in your own Git repo.

See the @gemstack/framework README for the CLI, the library API, and the driver seam.

Built on GemStack

The Framework is the turnkey product on top of GemStack: a collection of high-quality, framework-agnostic packages built with the Vike team. Each package is standalone, well tested, and works in any Node app; The Framework composes them into one product. Packages join GemStack by graduating one at a time when they prove framework-agnostic value, not by bulk-moving a framework's package set in.

All packages publish under the @gemstack/ scope (e.g. npm install @gemstack/ai-sdk).

Full docs live in docs/ (a hosted site is on the way): a guide, a getting-started walkthrough, and a deep guide per package (linked in the Docs column below).

Package	Description	Docs	Version
ai‑sdk	The agent runtime the rest of the AI family builds on: providers, agent loop, tools, streaming, middleware, structured output, memory, and evals.	Guide
ai‑skills	Portable capability bundles: load SKILL.md skills (instructions, tools, resources) and compose them onto an agent on demand.	Guide
ai‑autopilot	The AI-building framework: a Supervisor (plan, dispatch, synthesize), a runner sandbox, surfaces, an event-triggered review/QA loop with a prompt library, framework presets (Vike/Next), and a bootstrap flow from nothing to a production-grade app.	Guide
framework	The (AI) Framework: turnkey, zero-config orchestration on ai-autopilot. Wraps a coding-agent CLI (Claude Code) as a black box and takes an idea to a running app, with a live dashboard, Open Loop presets, an optional Docker sandbox, and a run relay for shared sessions.	README
ai‑mcp	The agent/MCP bridge: consume a remote MCP server's tools as agent tools, and expose an agent as an MCP server.	Guide
mcp	A standalone framework for authoring MCP servers: tools, resources, prompts, decorators, OAuth 2.1, a framework-neutral HTTP handler, and a test client. Agent-agnostic.	Guide
mcp‑connectors	The connector contract: define a service connector once with defineConnector, and compose any number into one MCP server with mountConnectors. Built on @gemstack/mcp; agent-agnostic.	Guide
mcp‑connector‑github	First-party connector: read and act on GitHub issues, pull requests, and repo files.	Guide
mcp‑connector‑google‑drive	First-party connector: browse, read, and share Google Drive files.	Guide

How they fit together

Two independent stacks. Arrows point to what a package depends on; nothing points "up."

graph TD
    subgraph fam["AI family · depends on the agent runtime"]
        framework["<b>framework</b><br/>turnkey app builder"] --> autopilot["<b>ai-autopilot</b><br/>orchestration"]
        autopilot --> skills["<b>ai-skills</b><br/>capability bundles"]
        autopilot --> sdk["<b>ai-sdk</b><br/>agent runtime"]
        skills --> sdk
        aimcp["<b>ai-mcp</b><br/>agent ↔ MCP bridge"] --> sdk
    end
    subgraph agn["Agent-agnostic · not ai-*"]
        connstar["<b>connector-*</b><br/>first-party connectors"] --> connectors["<b>connectors</b><br/>connector contract"] --> mcp["<b>mcp</b><br/>MCP server framework"]
    end

Loading

The ai- prefix means "depends on the agent runtime." ai-skills, ai-autopilot, and ai-mcp all reach ai-sdk, which depends on none of them; ai-autopilot also builds on ai-skills. A package about AI that is agent-agnostic (like @gemstack/mcp) is a peer of the family, not a member of it.

Connectors

@gemstack/mcp-connectors is the contract for wiring external services (GitHub, Google Drive, ...) into an agent as MCP tools. A connector declares its auth needs and its tools with defineConnector; the orchestrator supplies credentials and composes any number into one server with mountConnectors. First-party connectors ship as @gemstack/mcp-connector-*, and third parties publish their own mcp-connector-* against the same contract.

See Architecture.md for the full layering, naming rule, and graduation policy.

Which MCP package do I use?

The two MCP packages point in opposite directions, so they are never duplicates:

Exposing an existing agent? Use @gemstack/ai-mcp. It makes an agent speak MCP, or feeds remote MCP tools into one.

Authoring a server from scratch (tools / resources / prompts / auth)? Use @gemstack/mcp. A full server framework, with no agent involved.

Development

pnpm install
pnpm build        # build all packages (Turborepo)
pnpm dev          # watch mode
pnpm typecheck
pnpm test

Working on @gemstack/framework, run pnpm build first: the framework tests need the dashboard bundle. Without packages/framework-dashboard/dist/client, the daemon answers /_telefunc with a 503 text body and daemon.test.ts fails on JSON.parse (SyntaxError: Unexpected token 'h'), which reads like a real regression but is a missing build step.

This is a pnpm + Turborepo + Changesets monorepo. Runnable examples live under examples/: e.g. mcp-quickstart, autopilot-quickstart (Supervisor + runner + surfaces), and bootstrap-quickstart (the whole bootstrap flow, offline). See .changeset/README.md for the release flow.

Origin

The AI family was spun out of Rudder's mature @rudderjs/ai (v1.17.x) and re-versioned under the GemStack umbrella; @gemstack/mcp is the graduation of @rudderjs/mcp. The old @rudderjs/* names live on as the Rudder bindings over these engines: they re-export the agnostic core and add the framework-specific pieces that cannot graduate (e.g. @rudderjs/ai's AiProvider, ORM-backed stores, doctor check, and make:agent / ai:eval CLI).

Governance

GemStack is co-governed (shared npm @gemstack org + gemstack-land GitHub org). New tools join by mutual agreement; publish rights and 2FA are shared per the governance note.

License

MIT