// Agent profile
Nasiko is a developer control plane that solves the operational complexity of running multiple AI agents. Built in Rust, it operates as a single process that intercepts all agent-to-agent communication, providing intelligent routing (semantic shortlisting + contextual reranking + LLM selection), security enforcement (auth, rate limiting, ACLs), an MCP gateway for permission-filtered tool access, an LLM router that manages credentials server-side so agents never touch real API keys, and full OpenTelemetry observability with automatic token usage and cost tracking. Redis-backed cascade limits prevent runaway agent loops by controlling depth, fan-out, token budgets, and timeouts. Agents never need to be publicly accessible — all calls route through the control plane. 8.8K stars.
Nasiko addresses a pain point that emerges as soon as you move beyond a single agent: operations. When you have multiple agents calling each other, each with their own API keys, tool access, and cost profiles, the coordination overhead grows faster than the agent count. Nasiko consolidates this into a single control plane — one process that handles routing, security, credential management, observability, and cost tracking for your entire agent fleet.
The key architectural decision is that all inter-agent communication flows through Nasiko. Agents never call each other directly and never need public endpoints. This gives the control plane complete visibility into every dispatch, enabling features that are impossible when agents communicate peer-to-peer.
Intelligent Routing: Rather than hardcoding which agent handles which request, Nasiko uses a three-stage pipeline. First, semantic shortlisting identifies candidate agents based on capability descriptions. Then contextual reranking narrows the list using request context. Finally, an LLM-based selection picks the best match. Callers describe what they need, not who should handle it.
Security and Access Control: Every inter-agent call is authenticated and authorized through the control plane. Rate limiting and access control lists are enforced centrally. The MCP Gateway provides agents with permission-filtered access to external tools — an agent gets exactly the tools it needs, nothing more.
LLM Router: Model API credentials are managed server-side by the control plane. Agents make LLM calls through the router without ever seeing the actual API keys. This eliminates credential sprawl and enables centralized cost tracking and model switching.
Cascade Protection: Redis-backed limits prevent the classic multi-agent failure mode: runaway loops. You configure maximum call depth, fan-out limits, token budgets per cascade, and timeout thresholds. When any limit is hit, the cascade is terminated gracefully.
Observability: OpenTelemetry integration generates end-to-end traces for every dispatch, with automatic token usage counting and cost attribution. You can see exactly which agent called which, how many tokens each step consumed, and what the total cost was.
Teams running multi-agent deployments who need operational control without building custom infrastructure. Particularly valuable when agents are developed by different team members or teams — the control plane provides a consistent security, routing, and observability layer regardless of how individual agents are implemented.
Local-first session intelligence and analytics platform for coding agents, supporting Claude Code, Codex, and 20+ other agents.
817 structured cybersecurity skills for AI agents mapped to MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3.
Isolated-VM agent sandbox with sub-100ms boot — run AI agents on real machines with their own kernel, filesystem, and network. Pure Rust, open source.