Tenuo for NVIDIA NeMo Agent Toolkit

Provider-owned Agent Toolkit middleware that checks a Tenuo warrant before a function runs. Unauthorized calls stop before call_next, so protected function code never receives them.

This package is the in-process Agent Toolkit integration. The OpenShell supervisor middleware at the repository root is the independent enforcement point for MCP traffic that leaves the sandbox. Use both: the plugin catches calls early, with readable errors, and the middleware enforces at the network boundary.

Install

Install the published plugin into the same environment as Agent Toolkit. PyPI is the recommended installation path:

pip install nemo-agent-toolkit-tenuo
nat info components

The package metadata selects a compatible Agent Toolkit range. Current source is tested with Agent Toolkit 1.8 and 1.9. Use the checkout path below only when testing support that has not reached the latest PyPI release.

To install from a checkout instead, run this from the repository root:

uv sync --locked --project python/nemo-agent-toolkit-tenuo --extra test
uv run --locked --project python/nemo-agent-toolkit-tenuo nat info components

For an existing environment, an editable or regular local install also works:

python -m pip install ./python/nemo-agent-toolkit-tenuo

The package uses NVIDIA's third-party plugin convention:

For a runnable ReAct workflow with an approval step, see the Agent Toolkit example. That example protects MCP calls with the signing proxy. This plugin is the path for native Agent Toolkit functions; the integration guide explains how the two compose.

Bind task authority

Bind authority in application-controlled task scope. Do not accept a warrant or holder key from model-generated function arguments.

from nat.plugins.tenuo import authority

with authority(warrant.bind(holder)):
    result = await workflow.ainvoke(input)

holder is the task's signing key and must match the leaf holder named by the warrant. For a local-only smoke test, application code can mint both with the public Tenuo API:

from tenuo import SigningKey, Warrant

holder = SigningKey.generate()
warrant = Warrant.mint_builder().tool("read_logs").ttl(300).mint(holder)

In production, bind the short-lived warrant issued for the authenticated task or end user. Do not mint broad authority inside the agent process. See Production issuance.

The binding follows Python context propagation, including async tasks created inside the block. The middleware does not write MCP _meta.tenuo; use Tenuo's SecureMCPClient when the same authority must cross an MCP boundary.

That authority is not tied to Agent Toolkit. The same warrant can be verified again by the OpenShell middleware, an MCP server, an A2A worker, or another supported agent runtime, so delegation does not require translating authority into a new framework-specific policy.

Configure middleware

middleware:
  task_authorization:
    _type: tenuo
    trusted_roots:
      - "<64 hex characters of a 32-byte issuer public key>"

functions:
  read_logs:
    _type: read_logs
    middleware: [task_authorization]

trusted_roots selects accepted issuers and cannot be empty. A denied call raises AuthorizationDenied; a call needing a signed approval raises ApprovalRequired, which subclasses it. The middleware is not final, so later middleware continues after an allow.

Compatibility and verification

Tested with Python 3.11–3.13, nvidia-nat-core 1.8.x and 1.9.x, and Tenuo 0.3.x. The locked environment uses Agent Toolkit 1.9. Run its tests and distribution build with:

uv run --locked --project python/nemo-agent-toolkit-tenuo --extra test \
  pytest python/nemo-agent-toolkit-tenuo/tests -q
uv build --project python/nemo-agent-toolkit-tenuo

To run the same tests and nat info components discovery against Agent Toolkit 1.8 in a fresh environment:

ci/nat-compat.sh 1.8

See the repository security policy before reporting a vulnerability and release guide for the publication gate.

License

Apache-2.0.