> ## Documentation Index > Fetch the complete documentation index at: https://docs.rkat.ai/llms.txt > Use this file to discover all available pages before exploring further. # Meta Forge > Design an agent team in a guided conversation, then build, try and install it as a Meta application on your own computer. *Mirrored from Meerkat Meta main at [26f5ae9da1fd](https://github.com/lukacf/meerkat/commit/26f5ae9da1fd3caac378f1cd847ce50f9ec5ac92) on 2026-11-03; Meta version 0.9.9 at that commit.* Meta Forge is the setup and application authoring tool in Meerkat Meta. You describe a useful job, answer one focused question at a time, and inspect the proposed team of agents and its behavior. Forge then installs the result on your computer, or gives you a package you can review and install yourself. > [!NOTE] > > Forge is a developer alpha. A running process or an attractive simulation does > not establish that your job works. Try a real task in the installed > application, correct it, and restart it before you rely on it. ## How Forge fits with Core, Meta and Connect | Part | Role | | - | - | | Meerkat Core | Runs agents and mobs. Owns agent, session, communication and workflow semantics. | | Meerkat Meta | Composes applications around Core with lasting identities, policy and the Console, and provides the SDKs and gateways. | | Meta Forge | Owns the desired design (the blueprint), reviewable proposals, compilation and packaging. | | Meerkat Connect | Optional integration setup. A Forge release package carries a selected Connect product, and an application built with it can offer **Connect setup** in its installed Console. | Forge runs on your own computer as a local web application. It is a single-user tool, not a multi-tenant public server, and it listens only on `127.0.0.1`. Forge is not required to use Meta. If you compose your own application with the Meta SDKs, start with the [Meta quickstart](https://docs.rkat.ai/meta/quickstart.md) instead. ## What you can build > [card] Agent team > > Your agents, their configured tools and the native Console. Forge compiles > the team directly and installs the published Meta release it pins. It adds > no custom application code. > [card] Custom application > > For outcomes that need custom screens or application code. A builder agent > writes the code and its tests, and an independent reviewer checks them > before you can install the result. ## The guided workflow 1. **Describe one job.** A rough sentence is enough. Forge asks about the audience, scale, interaction, reliability and deployment before it narrows down tools and connectors. 2. **Shape the proposal.** Forge's own setup team, a durable four-agent team (deployment architect, integration specialist, systems reviewer and visualization programmer), asks the next question and publishes the proposal. An animation shows the proposed agents and how they collaborate. 3. **Accept the plan.** **Use this plan** saves the design and its choices. It creates no infrastructure. 4. **Build.** Install the agent team directly, or build a custom application that the builder writes and an independent reviewer checks. Optional trials run the design or the built application with sample data. 5. **Install and use it.** Forge installs on the same computer when the plan's destination matches it, and checks startup and runtime health. You then try a real task through the application or its Console. 6. **Return.** Saved setups, installations and their history stay on your computer. Start, stop, change private settings and apply checked updates from the same saved setup. ## What Forge checks and what you still verify Each stage is a separate outcome: | Stage | What it tells you | | - | - | | Proposal animation | An illustration of the current design. It does not run the agents. | | Agent design trial | Real native collaboration with sample data, without live services or application adapters. | | Checked application build | The code passed the build and independent review. Behavior on the destination still needs testing. | | Built application trial | Actual application callbacks run with isolated sample data. It does not install the application or verify live integrations. | | Installation | Startup and runtime availability. A useful job remains to be demonstrated. | ## Current scope * Each release package names the one destination it supports. Forge 0.3.0 supports Linux x86\_64 with exactly glibc 2.39, and Apple Silicon macOS 13.5 or newer. Other platforms and cloud deployment are not qualified by these packages. * Generated applications support static durable identities, native flows and the Console, scoped tools, local notes and an optional knowledge service. * Authenticated end-user applications, inbound messaging adapters, automatic recurring triggers, data-driven rosters, ephemeral worker fleets and synchronized ingestion are application extensions. The builder can write them as ordinary code. * Forge installs on the computer it runs on. To run an agent team on another computer, download its package and install it there from a terminal. * Forge also contains guided installation on an existing Linux computer over SSH and on a Google Cloud VM, for ordinary application packages. It refuses applications built on a selected Meta and Toolkit runtime, which is what the release packages build, and a complete live cloud journey remains unverified in this alpha. ## Next steps > [card] [Quickstart](https://docs.rkat.ai/meta/forge/quickstart.md) > > Download and verify Forge, then install a first agent team. > [card] [Concepts](https://docs.rkat.ai/meta/forge/concepts.md) > > Setups, blueprints, builds, trials, installations and updates. > [card] [Reference](https://docs.rkat.ai/meta/forge/reference.md) > > Commands, settings, files, blueprint fields and limits. > [card] [Troubleshooting](https://docs.rkat.ai/meta/forge/troubleshooting.md) > > Messages Forge prints and what to do about them. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.