Sim
TL;DR
Sim is an open-source, Apache-2.0 workspace for building, deploying and monitoring AI agents and workflows, maintained by Sim Studio. Agents are blocks on a visual canvas that can be routed, run in parallel, nested and paused for human approval, and it can call A2A agents and MCP servers. It suits teams building workflow-style agent automations.
Key facts
| Type | Platform |
|---|---|
| Languages / SDKs | TypeScript |
| License | Apache-2.0 |
| Pricing model | Open core |
| Orchestration pattern | Graph |
| GitHub stars | 29,755 (as of 2026-09-30) |
| GitHub forks | 3,852 |
| Last push | 2026-09-30 |
| Latest release | v0.9.6 |
| Repository | simstudioai/sim |
| Website | www.sim.ai |
| Documentation | docs.sim.ai |
| Last verified | 2026-09-30 |
Key features
- An agent is a workflow whose reasoning runs through one or more Agent blocks; each block takes a model and prompt and can be given tools, skills, knowledge bases, memory and structured output. (source)
- The Router block lets a model pick one of several paths by meaning, while Condition blocks branch on plain rules, so work can be sent to different agent blocks. (source)
- The Parallel block runs the block inside it concurrently, once per item or a fixed number of times, and the Loop block repeats work. (source)
- The Workflow block calls a child workflow as a step and waits for its result, so workflows can be composed like functions. (source)
- The Human in the Loop block pauses a run with no timeout until a person answers through an approval portal, the API or a webhook. (source)
- The A2A block sends messages to external A2A agents, tracks or cancels their tasks and fetches their Agent Cards. (source)
- Agents can use tools from external MCP servers, and the Sim MCP server exposes the Sim API to apps such as Claude Code and Codex. (source)
- The Pi Coding Agent block runs the Pi coding harness in an isolated sandbox to open or update pull requests or review code. (source)
Architecture and orchestration pattern
Pattern: Graph
Sim is a Next.js and Bun monorepo backed by PostgreSQL, with Socket.io for realtime, Trigger.dev for background jobs and E2B or isolated-vm for code execution, per the README tech stack. It can run as the hosted service at sim.ai, as a macOS desktop app, or self-hosted with an npx wizard that starts published Docker Compose images.
Multi-agent work is expressed as a workflow graph. Several Agent blocks are wired together with Router and Condition blocks for branching, Parallel and Loop containers for fan-out, and Workflow blocks for nesting; each block's output is read by later blocks by name. A separate Chat surface delegates work to built-in specialised subagents that appear as collapsible sections. Deployed workflows can be exposed through API, chat or MCP endpoints.
State is the saved workflow graph with immutable deployment versions, run logs, and workspace resources such as tables, files and knowledge bases. Agent memory is optional and keyed to a conversation; without it each run starts fresh, per the docs.
Human in the loop
The Human in the Loop block stops a run at a chosen point and shows the approver context in a portal, with notifications by Slack, Gmail, Microsoft Teams, SMS or webhook. The approver fills a resume form (for example approve or reject plus comments) that later blocks can read, and the run stays paused with no timeout. The same pause can be resumed through the API or a webhook. Pi Coding Agent blocks produce pull requests for a person to review and merge rather than committing to the default branch.
Harnesses it can drive
- Claude Code (evidence)
- Codex (evidence)
Protocols
| Protocol | Support | Note |
|---|---|---|
| MCP | Yes evidence | Both directions: the Sim MCP server (Streamable HTTP, mcp.sim.ai or the self-hosted /api/mcp) exposes the Sim API to MCP apps, and agents can use tools from external MCP servers; workflows can also be deployed as MCP endpoints. |
| A2A | Partial evidence | Client only as far as documented: the A2A block calls external A2A agents (send message, get task, cancel, Agent Card); no page describing Sim workflows served as A2A agents was found. |
| AG-UI | Unknown | Not found in llms-full.txt for docs.sim.ai or the README; GitHub code search for ag-ui matched only a competitor-comparison data file, and Sim is not in the AG-UI README integration list. |
Best for
- Building agent workflows on a visual canvas where several agent blocks are routed, run in parallel and nested as subworkflows.
- Approval-gated automations that pause for a person and resume from a portal, the API or a webhook. (shortlist)
- Self-hosting an agent workspace with a setup wizard and local models via Ollama or vLLM. (shortlist)
- Connecting an existing agent stack: call A2A agents and MCP servers from workflows, or drive Sim from Claude Code or Codex through its MCP server.
Not for
- Code-first teams that want an agent library inside their own process rather than a hosted or self-hosted application.
- Fully offline self-hosting without the Sim-managed Chat service, which needs a Chat API key.
- Autonomous supervisor-style agent teams; multi-agent structure is drawn as a workflow graph.
Quickstart
npx sim-setup # Self-host (Node.js 20+ and Docker required)
npx sim-setup # wizard: database, secrets, .env, Chat API key, Docker Compose
# open http://localhost:3000 when it finishes
# manage the install from its directory
npx sim-setup status
npx sim-setup logs
npx sim-setup update
# add optional capabilities later
npx sim-setup add llm
npx sim-setup add integration slack
# let Claude Code use your Sim workspace through Sim's MCP server (Sim Cloud URL)
claude mcp add --transport http sim https://mcp.sim.ai/mcp
Common pitfalls
- Self-hosting needs Node.js 20+ and Docker; the wizard creates a
sim/deployment directory and does not clone the repository. - Chat is a Sim-managed service: the wizard connects a Chat API key and opens your browser to sign in.
- Remote code execution in the tech stack uses E2B or isolated-vm; an open issue asks for a local Docker sandbox for self-hosted installs.
- A self-hosted MCP server URL is
https://<your-sim-host>/api/mcp; Claude's hosted connectors need a self-hosted Sim to be reachable from the internet. - MCP API keys do not expire until revoked; the docs prefer OAuth.
- Pi Coding Agent Create PR mode needs sandbox execution, your own provider API key and a GitHub token.
Pros
- Branching, parallel fan-out, loops and nested workflows are built-in blocks, so multi-agent structure needs no code. (source)
- The Human in the Loop block pauses a run without a timeout and can be resumed by portal, API or webhook. (source)
- It can call external A2A agents and MCP servers, and it exposes its own API as an MCP server. (source)
- Apache-2.0 licensed with a setup wizard that supports local models through Ollama and vLLM. (source)
- Deployed workflows are immutable versions that can be exposed as API, chat or MCP endpoints. (source)
Cons
- The README says Chat is a Sim-managed service and the self-host wizard connects a Chat API key. (source)
- An open issue reports that the v2 API rate-limits self-hosted installs by IP even with billing disabled. (source)
- An open feature request asks to replace E2B with a local Docker sandbox for self-hosted deployments. (source)
- The latest release is v0.9.6, so the project is pre-1.0 and releases arrive every few days. (source)
- Multi-agent behaviour is designed by drawing a workflow graph; the docs describe no autonomous agent-team or handoff primitive. (source)
Alternatives
FAQ
Does Sim support MCP and A2A?
MCP yes in both directions. A2A is documented as a client block that calls external A2A agents; no server side for serving Sim workflows over A2A was found.
Is Sim free?
The code is Apache-2.0 and can be self-hosted with npx sim-setup. Sim Studio also sells hosted plans (Free, Pro, Max and Enterprise on its pricing page), so the pricing model is open-core. Model provider costs are separate.
How does Sim do multi-agent work?
By composition on a canvas: several Agent blocks connected with Router, Condition, Parallel, Loop and Workflow blocks, plus a Human in the Loop block for approvals. Chat separately delegates to built-in subagents.
Can Claude Code or Codex use Sim?
Yes, as an MCP client: the docs show claude mcp add --transport http sim https://mcp.sim.ai/mcp and codex mcp add sim --url https://mcp.sim.ai/mcp. Sim does not launch those harnesses itself; its Pi Coding Agent block runs the Pi harness.
Sources
- Sim repository README
- Sim docs home
- Docs: Agents
- Docs: Router block
- Docs: Parallel block
- Docs: Workflow block
- Docs: Human in the Loop block
- Docs: A2A block
- Docs: Sim MCP
- Docs: Pi Coding Agent block
- Docs: Deploy workflows
- Sim pricing
- Issue 8452: v2 API rate limits self-hosted installs by IP
- Issue 4965: local Docker sandbox for self-hosted
- Release v0.9.6