The open-source platform for
multi-agent workflows
Use coding agents as a team, with shared context, and collaborative tools and controls. Agents run in isolated Instant containers or VM workspaces, with hosted compute and BYOC options. SAM handles provisioning, scheduling, and cleanup.
Describe what you want.
Watch it happen.
Talk to your agents like teammates. Pick the right model for the job, attach files, and watch them write code, run tests, and open PRs — all from a single chat.
Multiple agents, one interface
Switch between Claude Code, Codex, Gemini CLI, Amp, and more — each configured with their own API key or OAuth token. Choose the right agent for the task without leaving the chat.
Explore Chat-First Development
Work on projects.
Not alone.
Invite teammates with a link, approve access requests, and share agent profiles, skills, and project files. See who is running what, and whose keys are paying for it.
Invite, request, approve
Any active member creates an invite link. The recipient opens it and requests access — an owner or admin approves before they join. Approved members join as admins, with full project control except deleting the project or transferring ownership.
Explore Multiplayer Projects
Feedback that reaches
the agent.
Leave a comment on a chat message or a library file. Send it to the agent as context, and the agent can reply and resolve it — right from the tool it is using.
Comment on exactly what you mean
Select text in an agent’s message and leave a comment anchored to it. Comments work on library files too, so reviewers can flag a specific line in a document the agent produced and resolve it once it is fixed.
Explore Comments as Context
Pick your clouds.
See what they cost.
Build a compute pool from real, provider-native machines across seven clouds. Track token and compute usage per user, with budgets and quotas that keep spend predictable.
Real machines, not size labels
A compute pool lists the actual provider-native offerings it is allowed to rent — cx33, PRO2-S, vc2-4c-8gb, and more — across Hetzner, Scaleway, Vultr, Infomaniak, DigitalOcean, UpCloud, and GCP, each with its real vCPU, RAM, disk, and price. Browse each provider’s full catalog, filter by provider, region, vCPU, RAM, or price, and add a machine type without leaving the page. You choose the strategy (Balanced, Pack, Spread, Smallest fit) and what happens when a provider runs out (Queue, Fail, Fallback chain).
Explore Compute Pools & Cost Reporting
Automation your team
can see.
Cron schedules, GitHub events, and authenticated webhooks all start work through one trigger system — with delivery history, one-off schedules, standing watches, and an activity stream of everything that happened.
Three ways to start work
Triggers fire from cron schedules, from GitHub issues, comments, pull requests, and pushes, or from authenticated JSON webhooks. Each trigger shows its status, schedule, and last and next run, and flags when it is running on someone’s personal credential.
Explore Event Streams & Triggers
Agents that sleep,
wake, and coordinate.
Sessions normally sleep and wake from a 7-day snapshot. After repeated snapshot failures, idle VM sessions can fall back to the transcript and a verified Git recovery point, with a notice. Agents dispatch subtasks to each other and can interrupt a busy peer. App deployments turn a Compose release into a running, configurable app.
Close the tab. The session waits.
SAM normally restores the agent’s home directory and repository work from a verified snapshot for up to seven days. After repeated snapshot failures, an idle VM session can sleep with its transcript and a verified Git recovery point instead. SAM shows a notice, and the agent starts fresh on wake. Without a safe recovery point, sleep is blocked.
Explore Durable Sessions & Orchestration
Agents that remember
your preferences.
Build a project knowledge base that agents receive as context. Set policies that are injected into every session. Review a full log of agent activity.
Persistent project memory
Store observations about your architecture, conventions, and preferences. Each entry has a confidence score, a source (explicit or inferred), and a confirmation date so you can see what the knowledge base contains.
Explore Agents That Learn
From zero to coding agents in minutes
Four steps to coding agents working through your backlog in isolated workspaces.
Connect Your Runtime
Use Instant containers with no cloud setup, hosted compute where available, or add your own Hetzner, Scaleway, Vultr, Infomaniak, DigitalOcean, UpCloud, or GCP credentials to a compute pool.
Create a Project, Invite Your Team
Link a GitHub repo and invite teammates with a link. SAM organizes tasks, chats, and workspaces per project, shared by everyone who is approved.
Describe Your Task, or Let a Trigger Start It
Type what you need, or let a cron schedule, GitHub event, or webhook start it for you. SAM starts an Instant workspace or provisions a VM from the pool, then runs the agent.
Watch It Build, Comment, Iterate
Stream agent output in real time. Comment on a message to steer the agent, and get notified when it needs input or finishes.
How SAM stacks up
Own your infrastructure. Choose your agent. Pay cloud prices.
|
SAM
|
Managed AI Platforms
Devin |
Self-Hosted CDEs
Coder | |
|---|---|---|---|
| Your code runs on | Instant containers or your cloud (Hetzner, Scaleway, Vultr, Infomaniak, DigitalOcean, UpCloud, GCP) | Their infrastructure | Your infra (AWS, GCP, K8s) |
| Agent choice | Claude Code, Codex, Gemini CLI, Mistral Vibe, OpenCode, Amp | Their proprietary agent | Native agent with configurable LLM providers; third-party agents in workspaces |
| Open source | AGPL-3.0 license | Proprietary | Open-source core |
| Control plane | Workers Paid base plan | Bundled into pricing | You host it (K8s / Docker) |
| Infrastructure cost | Your cloud bill for BYOC VMs; hosted compute where available | From $20/mo (compute + inference) | Your cloud bill + license fees |
| Self-hostable | Fully — control plane + VMs | Managed service; enterprise VPC deployment available | Yes — standalone server, Docker, or K8s; Terraform workspaces |
| Shared projects & roles | Invite link, approval, owner/admin roles, credential attribution | Seat-based team plans | Workspace access control and read-only agent chat sharing |
| Comments that become agent context | Comment on chat or files, send to agent, agent replies and resolves | Bundled or unavailable | No agent-level comments |
| Compute pools across clouds | 7 providers, real SKUs, strategy + exhaustion policy per pool | Single vendor, opaque sizing | You wire it yourself |
| Token & compute usage reporting | Per-user usage by model/day + monthly budget, node-hours, admin cost summary | Bundled into one bill | AI usage reporting and licensed AI Gateway budgets |
| Event triggers & webhooks | Cron, GitHub events, authenticated webhooks, standing watches | Schedules, GitHub/integration events, and custom webhooks | Agent API for custom webhook and scheduled automation |
| Durable sleep/wake sessions | 7-day snapshots; labelled transcript + Git fallback for idle VMs | Varies, usually opaque | Persistent agent chats; workspace storage depends on template |
| Setup effort | Deploy to Cloudflare; connect a cloud for BYOC VMs | Zero config | Server deployment + Terraform workspace templates |
| Data sovereignty | VM code runs on your cloud; Instant and control plane data use Cloudflare | Vendor-managed | Full — your infrastructure |
SAM
Compute costs exclude AI inference — connect agent credentials or use SAM-managed AI where configured. Managed platform pricing bundles compute and model inference. Read the SAM vs Paseo comparison.
Built in the open, shipped fast
DigitalOcean and UpCloud are live. More providers are coming next.
Workspace management, GitHub OAuth, encrypted credentials, VM provisioning
WebSocket terminal, structured agent chat via ACP
Claude Code, Codex, Gemini CLI, Mistral Vibe, OpenCode, Amp
Multiple isolated workspaces per VM
Projects, persistent chat, activity feeds, task-driven workflows
Scaleway, GCP, Vultr, Infomaniak, DigitalOcean, UpCloud, notifications, voice input, conversation forking, warm pooling
Missions, agent-to-agent dispatch, task dependencies, durable messaging
Auth, task submission, chat prompts, task status, and runner doctor commands
Deployment environments, agent policy gates, Compose releases, logs, config, and volumes
Shared projects, invite links, access requests, roles, and credential attribution
Threads on chat messages and library files, sent to the agent as context
Multi-cloud offering catalogs, placement strategy, exhaustion policy, and cost reporting
Cron, GitHub, and webhook triggers, plus schedules, standing watches, and agent channels
AWS and expanded provider coverage
User-facing controls for assigning maintainer and viewer roles
Ready to run multi-agent workflows at scale?
Self-host on low-cost Cloudflare Workers. Bring your own cloud. Your agents, your infrastructure.