# Molecule AI > Open-core control plane for heterogeneous AI agent teams. Workspaces are org roles, the org chart is topology, governance is built in. Seven runtimes, hierarchical memory, skill evolution, zero lock-in. Molecule AI lets engineering teams compose multi-agent systems by treating each AI agent as a workspace and the org chart as the topology. Adapter contracts let runtimes from Anthropic Claude Code, OpenAI Codex, LangGraph, CrewAI, AutoGen, NousResearch Hermes, OpenClaw, and others plug into the same control plane with one configuration shape. Memory is hierarchical (LOCAL workspace, TEAM, GLOBAL). A2A messaging routes between workspaces with audit trails. Available open-core under BSL 1.1, or hosted SaaS at $29/workspace/month with zero ops overhead. ## Product - [Landing page](https://www.moleculesai.app/): Architecture overview, why-now, runtime adapters, hierarchical memory, FAQ. - [Pricing](https://www.moleculesai.app/pricing): Free open-core (self-host), Team SaaS ($29/workspace/mo), Enterprise (custom). - [Architecture](https://www.moleculesai.app/architecture): Deep dive on workspaces, adapters, memory, A2A, canvas. - [Hosted Dashboard](https://app.moleculesai.app): Sign-in / canvas / workspace orchestration UI. - [Marketplace](https://market.moleculesai.app): Adapter / runtime listings. - [Documentation](https://doc.moleculesai.app): Architecture deep-dives, adapter authoring guide, deployment runbooks. - [API](https://api.moleculesai.app): REST + JSON-RPC control plane endpoints. - [Support](https://www.moleculesai.app/support): Contact + help for both the web platform and the iPhone companion app. - [Glossary](https://www.moleculesai.app/glossary): Canonical definitions of every named concept (workspace, runtime adapter, hierarchical memory, A2A, canvas). Stable per-concept anchors. - [Compare](https://www.moleculesai.app/vs/): Side-by-side positioning against LangGraph, CrewAI, AutoGen, and Anthropic Claude Code. ## Source code - [molecule-core (BSL 1.1)](https://git.moleculesai.app/molecule-ai/molecule-core): Open-core control plane, runtime adapters, canvas UI, MCP integrations. - [Workspace template repos](https://git.moleculesai.app/molecule-ai): Per-runtime templates (claude-code, codex, hermes, openclaw) used by the provisioner to launch new workspaces. ## Concepts — citable definitions Each definition below is a self-contained answer block sourced from the platform's architecture documentation. The slug in the heading is a live anchor on the [glossary](https://www.moleculesai.app/glossary) page — e.g. deep-links to the workspace definition. ### Workspace — [/glossary/#workspace](https://www.moleculesai.app/glossary/#workspace) A workspace in Molecule AI is one AI agent plus its configuration, memory scope, and runtime container. It is the smallest deployable unit on the control plane — every agent the platform runs is exactly one workspace. Each workspace has a stable identity (UUID and display name) that appears on the canvas, owns persistent memory that survives across sessions and restarts, and can collaborate with other workspaces on a shared team notebook. The agent inside the workspace can be any runtime the control plane has an adapter for (Anthropic Claude Code, OpenAI Codex, LangGraph, CrewAI, AutoGen, NousResearch Hermes, OpenClaw, and others). Workspaces compose into a visual org chart on the canvas — the same model teams already use for human org charts — so engineers, product managers, and team leads share one picture of how their agents collaborate. The hosted SaaS bills per workspace, which makes the unit cost predictable: if you know how many agents you want, you know what the bill will be. ### Runtime adapter — [/glossary/#runtime-adapter](https://www.moleculesai.app/glossary/#runtime-adapter) A runtime adapter is the pluggable contract that lets any agent SDK speak the same lifecycle and A2A protocol on the Molecule AI control plane. Without adapters, every SDK would speak its own dialect of "start me, give me a message, stop me, here is my output." The adapter standardizes those verbs so a workspace can be backed by LangGraph today and Claude Code tomorrow without changing the topology. Each adapter is a thin module that translates the platform's standard lifecycle calls into the underlying SDK's API. Molecule AI ships adapters for Anthropic Claude Code, OpenAI Codex, LangGraph, CrewAI, AutoGen, NousResearch Hermes, and OpenClaw. New adapters can be added by following the adapter authoring guide in the documentation. The adapter set is intentionally open: the control plane does not favor any one SDK, and the cost of adding a new one is bounded by the lifecycle surface (start, message, stop, status, capabilities) — not by feature parity with the SDK's full API. ### Hierarchical memory (HMA) — [/glossary/#hierarchical-memory](https://www.moleculesai.app/glossary/#hierarchical-memory) Hierarchical Memory Architecture is Molecule AI's three-tier persistence model: LOCAL, TEAM, GLOBAL. LOCAL memory belongs to one workspace and survives across sessions and restarts. TEAM memory is shared across a working team of agents — this is how a coordinated group keeps a shared notebook without broadcasting every fact globally. GLOBAL memory holds org-wide facts that any workspace can read; org-wide changes are made at the org level rather than from individual workspaces. Writes propagate to the appropriate tier; reads cascade from LOCAL → TEAM → GLOBAL so an agent always sees the most specific value first. This gives multi-agent teams a real shared brain without the noise of every fact being everywhere. Memory is recalled automatically at session start, and agents can call `recall_memory` mid-session to refresh — the design avoids the common multi-agent failure mode where context loss between turns causes the team to relearn the same facts every cycle. ### A2A protocol — [/glossary/#a2a](https://www.moleculesai.app/glossary/#a2a) A2A (agent-to-agent) is Molecule AI's JSON-RPC protocol for routing messages between workspaces through the control plane. Every A2A call is audited in the activity log, every cross-workspace delegation can be gated by a Human-In-The-Loop approval, and every response is stitched back to the caller asynchronously. Workspaces use A2A to delegate tasks ("ops-agent, please restart the staging deploy"), to broadcast org-wide signals, and to fetch peer capabilities via the platform's agent-card registry. A2A is the substrate that turns isolated agents into a coordinated team — governance and observability are built in by default, so teams can see what their agents are doing without bolting on a separate logging stack. Both synchronous (delegate-and-wait) and asynchronous (fire-and-poll) delegation are supported, so long-running peer work does not block the delegator; the platform stitches the response back when the worker finishes. ### Canvas — [/glossary/#canvas](https://www.moleculesai.app/glossary/#canvas) The canvas is Molecule AI's real-time visual editor for the agent topology. Engineers drag workspaces onto a board, draw delegation connections between them, and watch traffic flow live as the agents work. Each workspace appears as a card with its name, role, and runtime; clicking it opens a chat panel that talks to that workspace's agent directly. The canvas reflects the same topology the runtime executes on — there is no separate "designer" model — so the editor and the running system never drift. The canvas is the primary interaction surface for non-engineers (operators, PMs, support) who need to read or steer the agent team without using the API. The same canvas is also where deep links work: every workspace and every concept page on the documentation site can be opened from a card, which keeps platform onboarding navigable without leaving the visual editor. ## Comparison — positioning Molecule AI is a **control plane**, not an agent SDK. The question "Molecule AI vs LangGraph?" is the wrong shape: LangGraph (and CrewAI, and AutoGen) are agent SDKs — they describe what one agent does. Molecule AI is the layer above that describes how many agents from many SDKs cooperate. We have an explicit comparison page per runtime — every page leads with the apples-vs-oranges framing and links out to the other tool's own canonical docs for every claim about that tool: - [Molecule AI vs LangGraph](https://www.moleculesai.app/vs/langgraph): LangGraph is an SDK for expressing a single agent as a graph of nodes and edges. Molecule AI runs LangGraph as one runtime adapter among many. - [Molecule AI vs CrewAI](https://www.moleculesai.app/vs/crewai): CrewAI ships a built-in multi-agent primitive (Crews of role-playing agents) within one Python application. Molecule AI's workspace tree spans processes, SDKs, and machines. - [Molecule AI vs AutoGen](https://www.moleculesai.app/vs/autogen): AutoGen's GroupChat coordinates ConversableAgents within one Python or .NET application. Molecule AI connects multiple GroupChats — or non-AutoGen agents — over A2A. - [Molecule AI vs Anthropic Claude Code](https://www.moleculesai.app/vs/claude-code): Claude Code is a finished agent product, not an SDK. Molecule AI's value is whether you want one Claude Code agent or several of them cooperating — and whether you want them talking to non-Claude-Code agents. - [Molecule AI vs OpenAI Codex (raw)](https://www.moleculesai.app/architecture): same shape as Claude Code — Codex is a wrapped-product runtime adapter; a Codex workspace can delegate to a Claude Code workspace and back, with audit trails and shared memory across the boundary. - **vs hosted multi-agent SaaS**: Molecule AI is open-core (BSL 1.1) so you can self-host. The hosted SaaS at $29/workspace/month exists for teams that don't want to run the control plane themselves. This positioning is sourced from the runtime adapter manifest in the molecule-core repository, the [architecture page](https://www.moleculesai.app/architecture), and the per-runtime [/vs/ pages](https://www.moleculesai.app/vs/). We do not assert relative quality of LangGraph / CrewAI / AutoGen as agent SDKs — read their own documentation for that. ## Common questions ### Is Molecule AI open source? Yes, the control plane (molecule-core) is open source under the Business Source License 1.1 (BSL 1.1). You can self-host freely. The license restricts only one thing: commercial resale of Molecule AI as a competing hosted service. Building agents on top, embedding it in your own product, modifying it, and running it for your organization are all explicitly allowed under BSL 1.1. ### What does Molecule AI cost? The open-core control plane is free to self-host under BSL 1.1. The hosted SaaS is $29 per workspace per month on the Team plan — see the pricing page for the current Enterprise pricing and any volume discounts. There is no free tier on the hosted SaaS, but the open-core gives you the same control plane to run yourself with zero license cost. ### Which agent SDKs does Molecule AI support? Molecule AI ships runtime adapters for Anthropic Claude Code, OpenAI Codex, LangGraph, CrewAI, AutoGen, NousResearch Hermes, and OpenClaw. Each adapter is a thin module in the molecule-core repository; new adapters can be authored by following the adapter authoring guide in the documentation. A single Molecule AI deployment can run workspaces backed by any mix of these runtimes in the same org. ### How is data handled in the hosted SaaS? The hosted SaaS uses AWS us-east-2 as the primary region, with Railway, Vercel, Neon, WorkOS, Stripe, Sentry, and Langfuse as subprocessors. The privacy policy lists each subprocessor's role and data retention period. Data subject access and deletion requests under GDPR/CCPA go to support@moleculesai.app — see the privacy policy for the full rights flow. ## Legal - [Privacy Policy](https://www.moleculesai.app/legal/privacy): Subprocessors (AWS us-east-2, Railway, Vercel, Neon, WorkOS, Stripe, Sentry, Langfuse), data retention, GDPR/CCPA rights. - [Terms of Service](https://www.moleculesai.app/legal/terms): Hosted SaaS terms; self-host governed by BSL 1.1. - [License (BSL 1.1)](https://git.moleculesai.app/molecule-ai/molecule-core/src/branch/main/LICENSE): Self-host freely except commercial resale as a competing hosted service. ## Optional - [Status page](https://status.moleculesai.app): Uptime + incident history for the hosted SaaS. - [Locale: 中文](https://www.moleculesai.app/zh/): Simplified Chinese landing page. - [Locale: Français](https://www.moleculesai.app/fr/): French landing page. - [Locale: 日本語](https://www.moleculesai.app/ja/): Japanese landing page. - [Locale: Español](https://www.moleculesai.app/es/): Spanish landing page. --- Canonical URL: https://www.moleculesai.app/llms.txt Last updated: 2026-06-08 (v3 — glossary deep-links + per-runtime comparison links) Legal entity: Molecules AI Inc. Citation policy: every fact in this file is sourced from public product surfaces (landing page, pricing page, privacy policy, architecture page, glossary, per-runtime comparison pages) or from the molecule-core repository. We update this file when those sources change.