Nasiko
Nasiko is the OpenRuntime for AI agents, controlling what agents cost, touch, and run with model routing, policy enforcement, cost tracking, and observability.
Visit Nasiko-Labs/nasikoOverview
Nasiko is an open runtime for AI agents, coding harnesses, frameworks, and tools. It discovers already-running coding agents, attributes their spend by harness and model, routes LLM calls to chosen models, and enforces policy at the gateway while tracing each run.
Key Features
- TokenOps forecasts the cost of a run before it starts, monitors live spend by team, agent, and model, and halts or escalates runs that cross cost, loop, or retry boundaries.
- The LLM Router gives agents a base URL and short-lived identity token instead of a real API key, resolving provider, model, and key server-side so agents and logs never see the real key.
- Runtime Controls enforce policy at the gateway on each call, including tool and data restrictions, budget and retry ceilings, approvals for sensitive actions, a kill switch, and an audit trail.
- The Multi-Agent Harness composes workflows across existing models, providers, and frameworks, coordinating provider switching, fallback routing, state across steps, and traces for every handoff.
- The platform includes an agentic registry, identity and access mapping, tenant isolation, gateway routing, and observability for cost and traces.
- Nasiko discovers and manages coding agents such as Claude Code, Codex, OpenCode, and Cursor CLI, with session reporting and LLM routing available for supported harnesses.
Use Cases
- Attribute AI spend across coding agents, harnesses, models, and teams instead of reconciling a single provider invoice.
- Route Claude Code, Codex, or OpenCode LLM calls to a chosen provider and model, including serving Anthropic-format traffic with an OpenAI key.
- Enforce MCP tool and data access policies, approvals, budget limits, and kill switches for production agents.
- Compose and trace multi-agent workflows that span multiple model providers and frameworks while checking cost and policy in flight.
Getting Started
- Install the CLI from a checkout with cargo install --path cli/ --force, then run nasiko agents discover to see detected coding agents without changing their settings.
- Use the Docker-only quick start when you do not want to install Rust.
- Download the Nasiko desktop app for macOS, Windows, or Linux.
- For routing, create an LLM config and connect a supported harness, such as nasiko connect claude --config <llm-config-name>.
Deployment & Requirements
- Docker is sufficient for the Docker-only quick start; Rust is required for installing the CLI from source.
- Reporting and routing require an active control plane and valid login.
- AGENT_JWT_SECRET is required for local coding-agent routing, including custom LLM configs.
- CODING_AGENT_OTLP_ENDPOINT is needed for coding-agent telemetry export; when unset, receipts remain pending without an export worker.
Before You Adopt
- License: Other. Review its terms before using, modifying, or distributing the project.
- Connecting a coding agent changes per-user global settings and affects every Claude Code or OpenCode process on the machine.
- Local coding-agent routing requires AGENT_JWT_SECRET, including for custom LLM configs.