Observe token usage. Control agent budgets. Route between providers. Understand where your tokens go in multi-agent systems.
A tool for developers building AI workflows with multiple agents, each with independent providers and budgets.
Narria is an orchestration layer that sits between your workflow and AI providers. It manages agent lifecycle, tracks token usage, enforces budgets, and routes requests to different providers. It's designed to be transparent, inspectable, and owned by the developer.
Manages multiple agents, coordinates their execution, tracks state, handles task dependencies, and returns structured results.
Records every API call: input tokens, output tokens, cost. Attributes tokens to specific tasks, steps, agents, and operations.
Sets and enforces spending limits at workflow, task, and agent levels. Stops execution if limits are reached.
Multi-agent AI systems have fundamental observability and control challenges that existing tools don't address.
Most frameworks run all agents through the same provider and model. There's no way to route different tasks to different providers or set per-agent budgets.
You know the total cost after execution. But where did tokens actually go? Which agent spent the most? Which operation is expensive?
You can set a total workflow budget, but not limits per agent or task. One expensive agent can consume the entire budget.
When a task costs 100 tokens, you don't know the breakdown: how much on context, how much on retries, how much on tool calls.
Define agents, roles, and dependencies. Specify which provider and model each agent should use, and set individual agent budgets.
Abstraction layer for different LLM providers. Currently: OpenRouter. In progress: OpenAI, Anthropic. Easy to extend.
Runs tasks, handles provider calls, tracks responses. Supports retries, fallbacks, and timeout handling.
Calculates token costs in real-time. Enforces spending limits. Stops tasks if budget is exceeded. Supports granular limits.
Records execution traces. Captures all API calls, responses, token counts, latency, and errors for every step.
Processes traces to generate reports. Breaks down token usage by task, agent, operation, and provider.
Narria processes execution traces to understand token distribution. Every trace records: API calls, tokens used, latency, errors, and which agent/task made the call.
| Task | Provider | Tokens | Cost |
|---|---|---|---|
| Code analysis | Anthropic | 98,240 | $1.47 |
| Code generation | OpenAI | 87,120 | $1.23 |
| Code review | OpenRouter | 60,272 | $0.72 |
Narria abstracts away provider-specific differences. Each agent is configured with a provider and model independently.
You define a workflow with agents. Each agent has a role, a target provider, a model, and a budget.
Narria executes tasks by routing them to the correct agent and provider. It tracks every API call and response.
Every response is parsed. Input tokens, output tokens, and cost are recorded. Budgets are checked in real-time.
After execution, traces are analyzed. You get detailed reports of where tokens were spent and why.
Narria is built in the open. The code is public, the design is transparent, and the development is community-driven. Anyone can review, contribute, or fork.
Narria is an early-stage project. If you're working on AI agent orchestration, cost observability, or have infrastructure needs, reach out.