SandBase Harness: persistent agent sessions and sandbox operations over MCP

[denial123789]

SandBase Harness is now active in the official MCP Registry as
io.github.sandbaseai/sandbase-harness (v0.3.4).

It is a stdio bridge for a local or self-hosted agent runtime, rather than a
model provider. MCP clients can create durable sessions, wait for streamed
turns, inspect results and artifacts, and stop work while execution happens in
the runtime's local, Docker, Kubernetes, or self-hosted-worker sandbox.

Registry package:

{
  "mcpServers": {
    "sandbase-harness": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "MANAGED_AGENTS_URL",
        "-e", "MANAGED_AGENTS_API_KEY",
        "ghcr.io/sandbaseai/sandbase-harness-mcp:0.3.4"
      ],
      "env": {
        "MANAGED_AGENTS_URL": "http://host.docker.internal:3000",
        "MANAGED_AGENTS_API_KEY": "optional-when-runtime-auth-is-disabled"
      }
    }
  }
}

The fixed image exposes six tools:

• list_agents
• create_session
• run_session
• get_session
• list_artifacts
• stop_session

run_session waits for the streamed turn to reach a terminal state and returns
assembled text plus terminal-event metadata, so clients do not have to
reimplement the session event loop.

Trust boundaries are deliberately explicit:

• the bridge stores no credentials;
• its effective authority is the authority granted by
  MANAGED_AGENTS_API_KEY;
• session data and artifacts stay in the configured Harness workspace;
• the default local-process sandbox is for trusted development, while Docker
  or Kubernetes should be selected when a stronger isolation boundary is
  required.

The published OCI image completed an MCP initialize + tools/list handshake,
and the official registry record is currently active.

• Registry record: https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.sandbaseai%2Fsandbase-harness
• Source and local runtime quickstart: https://github.com/sandbaseai/sandbase-harness
• Fixed release: https://github.com/sandbaseai/sandbase-harness/releases/tag/v0.3.4
• DeepSeek Harness integration: https://github.com/sandbaseai/sandbase-harness/tree/v0.3.4/examples/deepseek-harness

Feedback is especially welcome on the stdio-to-durable-session boundary,
artifact ergonomics, and the minimum tool surface. Disclosure: I am affiliated
with SandBase, and AI assistance was used to prepare and verify this post.

[denial123789]

Companion update: if you need a lightweight client onboarding layer rather than the full persistent Harness runtime, SandBase CLI now has its own Registry entry and technical show-and-tell. It configures 25 AI client targets around a six-tool progressive MCP bridge for 2,000+ models and APIs.

• CLI showcase: SandBase CLI: one local MCP bridge for 25 AI client targets #1584
• Registry entry: https://registry.modelcontextprotocol.io/v0.1/servers/io.github.sandbaseai%2Fcli/versions/0.1.17
• Source: https://github.com/sandbaseai/cli