🔍 Drishti (दृष्टि)

See what your agent thinks.

Drishti automatically captures, visualizes, and exports traces of AI agent execution. Add one decorator — see every LLM call with tokens, cost, and latency. Zero code changes to your agent logic.

from drishti import trace

@trace(name="my-agent")
def run_agent(query):
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": query}],
    )
    return response.choices[0].message.content

🔍 Drishti Trace — my-agent
├── ✅ [1] openai/gpt-4o-mini   312 tokens  $0.0001  124ms
└── ✅ [2] openai/gpt-4o        891 tokens  $0.0089  387ms

╭──────────── Summary ────────────╮
│ Total Tokens  1203              │
│ Total Cost    $0.0090 USD       │
│ Wall Time     511ms             │
│ LLM Calls     2                 │
│ Status        SUCCESS           │
╰─────────────────────────────────╯

✨ Features

• 🔌 Zero-config auto-detection — OpenAI, Anthropic, Groq, Ollama intercepted automatically
• 🌳 Rich terminal tree — every LLM call with tokens, cost, and latency at a glance
• 💾 JSON export — full traces saved to .drishti/traces/ for sharing, diffing, and replaying
• 🖥️ CLI tool — drishti version, drishti list, drishti view, drishti diff, drishti stats, drishti export, drishti replay, drishti clear
• 💰 Cost tracking — real-time pricing for 15+ models across 4 providers
• 🛡️ Budget guard — warn when cost exceeds a threshold
• ⚡ Async support — works with async def functions out of the box
• 🔒 Thread-safe — correct isolation for concurrent agents via thread-local + ContextVar
• 🪶 Zero overhead — pure passthrough when no @trace context is active

🚀 Quickstart

Install

pip install drishti-ai[openai]        # OpenAI support
# or
pip install drishti-ai[anthropic]     # Anthropic support
# or
pip install drishti-ai[all]           # All providers

Trace Your Agent

from drishti import trace
import openai

client = openai.OpenAI()

@trace(name="research-agent")
def research_agent(query: str) -> str:
    # Step 1: Generate search queries
    plan = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": "Generate 3 search queries for the topic."},
            {"role": "user", "content": query},
        ],
    )
    queries = plan.choices[0].message.content

    # Step 2: Synthesize answer with a stronger model
    answer = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": "Synthesize a comprehensive answer."},
            {"role": "user", "content": f"Topic: {query}\nQueries: {queries}"},
        ],
    )
    return answer.choices[0].message.content

result = research_agent("What is quantum computing?")

That's it. Drishti automatically:

1. Intercepts both LLM calls
2. Captures tokens, cost, latency, and full I/O
3. Renders a rich terminal tree
4. Exports the trace to .drishti/traces/ as JSON

📦 Installation

pip install drishti-ai              # Core only (no provider SDKs)
pip install drishti-ai[openai]      # + OpenAI SDK
pip install drishti-ai[anthropic]   # + Anthropic SDK
pip install drishti-ai[groq]        # + Groq SDK
pip install drishti-ai[ollama]      # + Ollama SDK
pip install drishti-ai[all]         # All providers

Requirements: Python 3.10+

🎯 Supported Providers

Provider not installed? Drishti keeps running and prints a one-time actionable warning with the install extra.

Unknown model? Cost defaults to $0.00. Trace still works perfectly.

🖥️ CLI

# Print installed version
drishti version

# List all saved traces
drishti list

# Replay a trace in the terminal
drishti view <file>          # by file path
drishti view <id-prefix>     # by trace ID prefix
drishti view <file> --full   # show full prompt/completion payloads

# Compare two traces
drishti diff <trace-a> <trace-b>

# Aggregate stats
drishti stats

# Export a trace as CSV
drishti export <trace> --format csv

# Replay the same LLM requests and compare deltas
drishti replay <trace>

# Delete all saved traces
drishti clear

Example output of drishti list:

📋 Saved Traces

  20260416_153042_research_agent.json  research-agent  success  1203 tokens  $0.0090
  20260416_152801_claude_agent.json    claude-agent    error    0 tokens     $0.0000

⚡ Async Support

Drishti auto-detects async functions and Just Works™:

from drishti import trace
import openai

client = openai.AsyncOpenAI()

@trace(name="async-agent")
async def async_agent(query: str) -> str:
    response = await client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": query}],
    )
    return response.choices[0].message.content

⚙️ Configuration

Per-call configuration

@trace(
    name="my-agent",      # Custom trace name (default: function name)
    budget_usd=0.05,      # Per-trace budget in USD
    on_exceed="warn",     # "warn" (default) or "abort"
    display=True,          # Print tree to terminal (default: True)
    export=True,           # Save JSON to disk (default: True)
)
def my_agent():
    ...

Config file

Create .drishti/config.toml in your project root:

[drishti]
display = true              # Print trace tree to terminal
export = true               # Save traces to disk
default_export_dir = ".drishti/traces"  # Preferred export dir (traces_dir still works)
budget_usd = 0.10           # Budget threshold
on_exceed = "warn"          # "warn" or "abort"
quiet = false               # Suppress terminal tree output
auto_open_on_error = false  # Auto-open trace output on errors
max_preview_chars = 220     # Prompt/completion preview truncation length
estimate_stream_tokens = true  # Use optional tiktoken estimation for streams

Decorator usage patterns

# Bare decorator — name defaults to function name
@trace
def my_agent():
    ...

# With custom name
@trace(name="research-agent")
def my_agent():
    ...

# Budget guard
@trace(budget_usd=0.05)
def expensive_agent():
    ...

# Hard abort once budget is exceeded mid-run
@trace(budget_usd=0.05, on_exceed="abort")
def budget_guarded_agent():
    ...

🛡️ Error Handling

Drishti follows one golden rule: never change the behavior of your agent code.

📄 JSON Export Format

Traces are saved to .drishti/traces/ as JSON:

{
  "trace_id": "a1b2c3d4-...",
  "name": "research-agent",
  "started_at": "2026-04-16T15:30:42.123456+00:00",
  "ended_at": "2026-04-16T15:30:42.634567+00:00",
  "status": "success",
  "summary": {
    "total_tokens": 1203,
    "total_cost_usd": 0.009,
    "total_latency_ms": 511.0,
    "span_count": 2
  },
  "spans": [
    {
      "span_id": "...",
      "step": 1,
      "name": "openai/gpt-4o-mini",
      "provider": "openai",
      "model": "gpt-4o-mini",
      "tokens": { "prompt": 45, "completion": 267, "total": 312 },
      "cost_usd": 0.0001,
      "latency_ms": 124.0,
      "status": "success"
    }
  ]
}

🧪 Development

# Clone and setup
git clone https://github.com/aarambh-darshan/drishti.git
cd drishti
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,all]"

# Run tests
pytest tests/ -v

# Run with coverage
pytest tests/ --cov=drishti --cov-report=term-missing

# Lint
ruff check drishti/ tests/
ruff format --check drishti/ tests/

🗺️ Roadmap

See ROADMAP.md for the full feature plan.

📐 Architecture

See ARCHITECTURE.md for the complete system design, including:

• System architecture diagram
• Data flow walkthrough
• Provider interception strategy
• Thread safety / async support design
• Error handling philosophy

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

📄 License

MIT — see LICENSE.

Drishti (दृष्टि) — See what your agent thinks.