Managed connectors for sandbox agents
MIOSA agents should build inside MIOSA sandboxes or computers. The local CLI or MCP client should orchestrate the work; the sandbox/computer should be where the agent writes files, installs packages, runs commands, previews apps, and produces artifacts.
Managed connectors give those in-VM agents tools without asking every client to bring vendor credentials.
Connectors, egress, and MCP
These three surfaces work together, but they are not interchangeable:
| Surface | Owns | Example |
|---|---|---|
| Connector | Identity and credential grant for a provider or tool | refero/design-research, anthropic/workspace-claude, a tenant OAuth install |
| Egress | Network policy and token brokering from the sandbox/computer to approved endpoints | Swap miosa-tok-* only when calling https://api.refero.design/mcp |
| MCP | Tool protocol the agent runtime speaks inside the sandbox, computer, editor, or hosted endpoint | Claude Code calls a Refero MCP server or the MIOSA hosted MCP server |
A connector says what account/tool is granted. Egress says where network traffic may go and when placeholders can be exchanged. MCP says how an agent discovers and invokes tools. A product may use all three in one run: attach a connector, enforce egress to the provider endpoint, and configure the agent runtime to call that provider through MCP.
Default flow
Create a sandbox with enough time for active agent work:
miosa create builder \
--template nextjs \
--ttl 1h \
--wait \
--json Attach Refero Design Research:
miosa connections add tools \
--name refero-design-research \
--url https://api.refero.design/mcp \
--header-env Authorization=REFERO_MCP_TOKEN \
--json Run the agent inside /workspace:
miosa prompt --sandbox <sandbox-id> \
--harness claude \
--json \
"Research product references, build the page, run tests, and leave files in /workspace" For direct tool use, the same sandbox can be edited with file and command commands:
miosa files cp ./page.jsx <sandbox-id>:/workspace/app/page.jsx --json
miosa exec <sandbox-id> --cwd /workspace --json -- npm run build
miosa wait <sandbox-id> --timeout 180s --json How the token works
The sandbox receives a brokered placeholder:
REFERO_MCP_TOKEN=miosa-tok-... The real provider token stays encrypted in MIOSA-controlled infrastructure. The egress proxy swaps the placeholder only for the provider endpoint it belongs to, for Refero:
https://api.refero.design/mcp That means:
- clients do not paste the Refero token,
- the token is not shown by
miosa connections list, miosa connections test refero-design-researchdoes not return the token,- sandbox code can still use normal environment-variable based MCP config.
MCP config inside the sandbox
Use the brokered env var in the agent’s MCP configuration:
[mcp_servers.refero]
enabled = true
url = "https://api.refero.design/mcp"
[mcp_servers.refero.http_headers]
Authorization = "Bearer ${REFERO_MCP_TOKEN}" Tenant-provided connectors
When MIOSA does not provide a connector, tenants can still bring their own:
printf '%s' "$ANTHROPIC_API_KEY" \
| miosa connections add models anthropic \
--label workspace-claude \
--key-stdin \
--json Which runtime should I use?
Agent coding, generated files, tests, package installs, live preview, and publish prep.
Persistent desktop or VM workflows, browser/GUI tasks, SSH, and long-running operator sessions.
Durable app containers on the workspace Docker Host host after the sandbox-built app is ready.