Built for NVIDIA Nemotron 3 Ultra

Make the model's
thinking visible.

A minimal, powerful agent framework that surfaces every reasoning step, live tool-call construction, and self-correction from Nemotron 3 Ultra.

Watch the interactive demo Read the docs
Used in research demos & production agent tooling
nemotron-think hero: stylized transparent AI agent head with visible thought processes
enable_thinking: true
reasoning_budget: 8192
100%
of reasoning surfaced
2
lines to add a custom tool
0
extra API keys required
Replayable
traces for demos & evals
Why this exists

Every thought.
Every decision. Visible.

Most agents are black boxes. nemotron-think turns the black box into a glass box.

Visible reasoning stream

Nemotron 3 Ultra’s enable_thinking + reasoning deltas are rendered live in the terminal (green). You see the model think before it acts.

Live tool call streaming

Watch delta.tool_calls being assembled token-by-token (orange). The model’s argument construction is no longer hidden.

Rich, replayable traces

Every run becomes a self-contained JSON artifact with full reasoning, tool calls, and observations. Replay perfectly offline with nemotron-think replay.

Controllable reasoning budget

Dial --budget 8192 or --low-effort per run. Match compute to task difficulty without changing your agent code.

Pluggable tools, 5 lines each

Define a Python function + JSON schema. The agent handles streaming, execution, observation injection and trace capture automatically.

Gold for evals & fine-tuning

The AgentRun + per-step reasoning traces are perfect structured data for building reasoning datasets or running rigorous agent evaluations.

No API key required

Interactive trace replay

This is a real saved trace from the agent. Press play to watch the model reason and call tools exactly as it happened.

nemotron-think replay examples/traces/math_tool_example.json
nemotron-3-ultra
TASK
Step 0 / 2
— Waiting for playback —
reasoning
tool call deltas
final content
Learn more about traces
Under the hood

A tiny, transparent loop

  1. 01
    User sends task
    VisibleReasoningAgent.run(task, reasoning_budget=...)
  2. 02
    Stream with thinking enabled
    extra_body contains enable_thinking + reasoning_budget + tools
  3. 03
    Print + capture simultaneously
    Reasoning (green) → content → live tool call deltas (orange)
  4. 04
    Execute tools locally
    Results fed back as role:tool messages for the next turn
  5. 05
    Repeat or finish
    When model emits no tool_calls → we have the final answer
The key insight
Nemotron 3 Ultra doesn’t just return a final answer. It can stream its internal reasoning and build function call arguments incrementally. Most frameworks discard or hide this. nemotron-think captures and surfaces all of it by default.
ReasoningTrace
Raw chunks, tool acc state, extra_body used. Perfect for analysis.
AgentStep
One full turn: reasoning + tool calls + results + the trace.
AgentRun
Complete record. Serializable. The artifact you commit and share.
CLI (recommended for demos)
nemotron-think run "Research the best open-source
text-to-video models released in the last 6 months"

# With more thinking power
nemotron-think run "..." --budget 8192

# Perfect offline demos
nemotron-think replay traces/my_research.json
Install once → instant powerful demo harness
Python API (for building agents)
from nemotron_think import VisibleReasoningAgent

agent = VisibleReasoningAgent(max_steps=8, default_reasoning_budget=4096)

run = agent.run("Write and test a kth-largest function", 
                reasoning_budget=8192)

run.save("my_agent_run.json")
print(run.final_answer)
Every run.steps contains full reasoning

Default tools (pluggable & visible)

The model sees clean schemas. You see every decision to use them.

get_math_answer
Safe arithmetic evaluator. Supports +, −, ×, ÷, parentheses.
expression: string
get_current_time
Returns current UTC time as ISO 8601. Useful for time-aware tasks.
no parameters
python_exec
Execute Python snippets in a temp dir with timeout + basic guardrails. Great for algorithms & verification.
code: string, timeout?: number
web_search
DuckDuckGo-backed search (no key). Returns titles, links and snippets for research tasks.
query: string, max_results?: int
Learn how to add your own tools in 5 lines

Ready to make your agents trustworthy?

Clone the repo, set your NVIDIA key, and run your first visible reasoning agent in under two minutes.

Clone on GitHub Read the getting started guide