liberrex-mcp-client (Python)

Call the tools of other platforms on the Liberrex MCP mesh (Liberrex AI, Animatoria, the ERP…) from async or sync code, and hand them to an LLM as function-calling tools. Implements the mesh conventions; the other side is liberrex-mcp-server. TypeScript equivalent: @liberrex/mcp-client.

pip install liberrex-mcp-client

Configure the platforms you call

LIBERREX_MCP_PLATFORM=animatoria         # this platform's own name (trace origin)
LIBERREX_MCP_PLATFORMS='{"liberrex-ai": {"url": "https://ai-api.liberrex.com/mcp", "key": "lxp_…"},
                         "erp":         {"url": "https://erp.acme.tn/mcp", "key": "lxp_…", "timeout": 20}}'

Each platform issues your key (liberrex-mcp-server new-key <your-name>) and lists you in its LIBERREX_MCP_PEERS. Keys are secrets: keep them server-side.

Call tools

from liberrex_mcp_client import Identity, Mesh, PlatformToolError

mesh = Mesh.from_env()
sara = Identity(user_id="42", tenant_id="biat", roles=("sales.read",), locale="fr")   # the signed-in user

recs = await mesh.call("liberrex-ai", "sales_recommendations",
                       {"mode": "next_best_product", "context": {"record": {"type": "Customer", "id": "C-118"}}},
                       identity=sara)

try:
    mesh.call_sync("erp", "send_reminder", {"invoice_id": "F-1024"}, identity=sara)   # Flask, Celery, scripts
except PlatformToolError as error:
    if error.code == "writes_disabled": ...

• The identity is the signed-in person, set by your code and never by a model. The called platform applies that person's permissions.
• PlatformToolError.code is a mesh code: forbidden, not_found, identity_required, tenant_not_allowed, writes_disabled, rate_limited… PlatformUnavailable means the platform was unreachable or refused the key.
• list_tools(name) is cached for 5 minutes. Each RemoteTool carries input_schema, write, permission and category.
• Inside a platform tool, pass meta=ctx.forward_meta() instead of identity to keep the trace across platforms.

Give the tools to an LLM

from liberrex_mcp_client import anthropic_tools, openai_tools, route

tools = await mesh.list_tools("animatoria")
specs = anthropic_tools([t for t in tools if not t.write])   # names: animatoria__production_ranking
platform, tool = route(call.name, tools)                     # back to ("animatoria", "production_ranking")
result = await mesh.call(platform, tool, call.input, identity=sara)

Get a human's approval before you run a write tool.

Development

.venv/bin/pytest     # the end-to-end tests live with the server package (sdk/mcp-server/python/tests)