Tool servers

A tool is a Python function the agent imports and calls in code, and the tools it may call come from tool servers. These two endpoints list the servers your account can use and the tools inside each one — the names you write into an agent’s tools.

The servers

curl -s "$SPLOX_BASE_URL/v2/tool-servers" \
  -H "Authorization: Bearer $SPLOX_API_KEY"

{
  "data": [
    {"id": "system:splox",   "name": "Splox",   "kind": "system", "description": "Splox platform tools: agent, harness, task management, skills library (skills_manage)"},
    {"id": "system:compute", "name": "Compute", "kind": "system", "description": "Compute: one unified, target-parameterized surface over three execution backends..."},
    {"id": "system:media",   "name": "Media",   "kind": "system", "description": "Read and analyze images, audio, video, and documents..."},
    {"id": "system:search",  "name": "Search",  "kind": "system", "description": "Web search the live internet and pull page contents back as text..."},
    {"id": "system:email",   "name": "Email",   "kind": "system", "description": "Your own email address: read the mail sent to you and send mail from it..."},
    {"id": "system:sms",     "name": "SMS",     "kind": "system", "description": "A real phone number you can receive SMS on..."},
    {"id": "system:memory",  "name": "Memory",  "kind": "system", "description": "Your memory at two distances: the sessions this person had before you, and the window you are carrying right now..."},
    {"id": "system:harness", "name": "Harness", "kind": "system", "description": "What a harness cannot read off its own checkout..."}
  ]
}

System servers come first, with stable system:* ids, and your own connected MCP servers follow with UUID ids and kind: "user". The descriptions are long because they are written for the agent that has to decide when to reach for the server — they are trimmed here.

The tools of one server

curl -s "$SPLOX_BASE_URL/v2/tool-servers/system:compute/tools" \
  -H "Authorization: Bearer $SPLOX_API_KEY"

{
  "data": [
    {"name": "compute_exec",       "description": "Run a shell command on the selected execution target..."},
    {"name": "compute_read_file",  "description": "Read a file from the selected execution target..."},
    {"name": "compute_write_file", "description": "Write a file (create or overwrite) on the selected execution target..."},
    {"name": "compute_edit",       "description": "Perform an exact string replacement in one file..."},
    {"name": "compute_grep",       "description": "Fast text search across files..."},
    {"name": "compute_shell",      "description": "Persistent interactive shell on the selected execution target..."},
    {"name": "compute_computer",   "description": "Drive the SANDBOX desktop..."},
    {"name": "compute_target",     "description": "Acquire and manage execution targets..."},
    {"name": "compute_download",   "description": "Pulls a file OUT of the selected execution target..."},
    {"name": "compute_view",       "description": "View an image file from the selected execution target..."},
    {"name": "compute_preview",    "description": "Expose a port from the sandbox target as a signed, time-limited public preview URL..."}
  ]
}

Eleven tools on system:compute; six on system:splox, including splox_ui, the one that raises an interaction; two on system:search, search and fetch.

What the names are for

A server id is what an agent’s tools entry names, and a # narrows it to particular tools:

researcher = agent(
    "Researcher",
    system_prompt="You are a research assistant.",
    model="claude-sonnet-4-6",
    provider="anthropic",
    tools=["system:compute", "system:search#search,fetch"],
)

Without the #, the agent gets every tool the server has. With it, it gets exactly the ones you list — and the names must be ones this endpoint returns, or the publish is refused.

System and user servers

They differ in where the answer comes from. A system server is listed in-process, so it is always available. A user server is queried over the wire, which means it can be down:

Answer	When
`404 not_found`	the server does not exist, or belongs to somebody else
`502 upstream_unavailable`	your MCP server is unreachable or failing

In the SDKs

servers = client.tool_servers.list()
compute = next(s for s in servers if s.id == "system:compute")
print([t.name for t in client.tool_servers.tools(compute.id)])

const servers = await client.toolServers.list();
const compute = servers.find((s) => s.id === "system:compute")!;
const names = (await client.toolServers.tools(compute.id)).map((t) => t.name);

servers, err := client.ToolServers.List(ctx)
tools, err := client.ToolServers.Tools(ctx, "system:compute")

​The servers

​The tools of one server

​What the names are for

​System and user servers

​In the SDKs

The servers

The tools of one server

What the names are for

System and user servers

In the SDKs