Last updated: 11/17/2026Last verified: 2026-11-17
Open runtime for AI agents: it discovers the Claude Code, Codex, OpenCode and Cursor CLI agents on a machine, reports their tokens and cost in one dashboard, routes model calls through its own router with short-lived tokens, and governs self-deployed agents with access control, an MCP gateway, human approval and flow guards. Apache-2.0 and self-hostable with Docker Compose; the free tier runs two cloud agents.
What is Nasiko?
Nasiko is an open-source (Apache-2.0) runtime for coding agents and deployed AI agents, built in Rust by Nasiko Labs in Palo Alto. A CLI finds the harnesses already installed on a machine — Claude Code, Codex, OpenCode and the Cursor CLI — and session reporting puts each completed turn's tokens and cost into one dashboard, attributed by model, agent, session, developer and team. The same server sits in the call path: it routes model calls through an OpenAI-compatible LLM router, exposes tools through an MCP gateway, and hosts A2A agents you build yourself. Nasiko runs as Nasiko Cloud at app.nasiko.com, as desktop apps for macOS, Windows and Linux, or self-hosted with Docker Compose. Based on www.nasiko.com, its pricing page, docs and the GitHub repo (checked 2026-11-17); not installed or tested.
Key Features
- Coding-agent discovery: per the README, `nasiko status` lists Claude Code, Codex, OpenCode and the Cursor CLI found on the machine. Discovery is read-only and works before you have an account or a server.
- Session reporting: one command connects the detected harnesses and imports past sessions, so every completed turn's tokens and cost land in one dashboard; `--no-content` keeps prompt and response text out.
- Model routing per harness: `nasiko connect claude --config ...` points a harness at Nasiko instead of its vendor, and the README shows Claude Code served by an OpenAI key. Cursor CLI routing is listed as 'not yet' in the README's support table.
- LLM Router: agents receive a short-lived token rather than a real API key; provider, model and key are resolved on the server, and the router offers an OpenAI-compatible egress.
- TokenOps cost attribution: usage and spend per agent, model, session, developer and team, instead of one provider total to reverse-engineer.
- Token optimization: large tool payloads are compressed, tool schemas compacted and chat history selected, with each saving recorded per request (vendor description, not an independent measurement).
- One trace per request: every hop emits an OpenTelemetry span, so a request becomes a single end-to-end trace with its cost attached.
- Governance: agents are not publicly reachable (single ingress), ownership and grants plus an allowlist decide which agent may call which, the MCP gateway keeps tool credentials instead of the agent, and per-agent secrets are encrypted at rest with AES-256-GCM.
- Human-in-the-loop and flow guards: a tool set to 'ask' pauses the agent until someone approves; flow guards limit depth, fan-out, token budget and cycles.
- Orchestration: `nasiko deploy` builds, pushes and runs any A2A v1.0 agent (Python, Rust, Go or TypeScript) with no SDK; a three-stage routing engine (embedding shortlist, rerank, LLM selection) picks the agent, and an embedded S3-backed OCI registry stores artifacts.
- Team controls: docs and the pricing page list access control lists, a fleet view, SSO and SCIM for organizations; multi-agent flows are a Pro feature.
Best For
Pricing
Nasiko Cloud is sold per plan at app.nasiko.com with your own model keys on every plan — your provider bills model usage, not Nasiko — and no per-seat fees. Free is $0 for life: 2 cloud agents, unlimited local agents in the desktop app, the Agent Gateway and MCPs with your own keys, idle cloud agents sleeping and waking on demand; writable and public agents, ACLs and multi-agent flows are not included. Individual is $20 per month for one builder: up to 10 cloud agents, agents running all month with no cold starts, writable agents, public agents and MCPs you can share, and a fresh $20 allowance every month. Pro is $99 per month for the whole team: unlimited members by invite, unlimited cloud agents, access control lists, multi-agent flows, advanced settings and one fleet view. Enterprise is a custom yearly licence by quote: a dedicated cloud project, your own model keys, no Nasiko agent limits and desktop rollout through MDM. If a monthly allowance runs out, running agents keep going and new agents wait until renewal; nothing is deleted, and you can cancel at the end of any monthly cycle. Self-hosting the runtime carries no Nasiko plan fee, but you operate and pay for the Docker Compose stack (server, Postgres, Redis, S3/RustFS and the OTel, Tempo and Loki tracing services), and some self-managed extras such as the nasiko-ee admin CLI need an Enterprise licence.
Pros & Cons
Pros
- Covers discovery, session reporting, model routing, cost attribution and agent governance in one Apache-2.0 runtime.
- Supports four named harnesses and imports their past sessions, so historical spend is not lost when you adopt it.
- Read-only local discovery runs without an account or a server.
- Server-side key resolution with short-lived tokens keeps provider keys out of harness configuration.
- Transparent plan structure: cloud-agent counts, ACLs and multi-agent flows are the only fences, and model usage is billed by your own provider.
- Self-hostable with Docker Compose and a documented CLI, for teams that cannot send cost data to a hosted workspace.
Cons
- The Free plan excludes writable and public agents plus ACLs and multi-agent flows; those start at Individual and Pro.
- The Cursor CLI can be discovered and reported but not yet routed to another provider, per the README support table.
- Model tokens are never included: you bring keys and your provider bills usage, so Nasiko's price sits on top of model spend.
- Self-hosting means operating a multi-service stack (Postgres, Redis, RustFS S3, OTel/Tempo/Loki), and upgrading sometimes needs a fresh database because migrations were renumbered.
- Token-saving figures on the site come from Nasiko's own compression test, not an independent evaluation.
- Enterprise terms are quote-only, so a dedicated cloud project and MDM desktop rollout require contacting sales.
Alternatives
Observability and evaluation platform from LangChain for tracing LLM and agent runs, prompts and datasets.
Open-source LLM engineering platform with tracing, evaluation, prompt management and usage and cost reporting.
Open-source proxy that puts many model providers behind one OpenAI-compatible interface, with key management and budget tracking.
AI gateway with routing, guardrails and request observability for production LLM applications.
FAQ
Details
Platform
Features
- Discovers Claude Code, Codex, OpenCode and Cursor CLI on a machine — read-only, no account needed — and lands every completed turn with its tokens and cost in one dashboard
- LLM Router gives agents a short-lived token instead of a real API key and resolves provider, model and key server-side, so Claude Code can run on an OpenAI key
- Cost attribution per agent and per model, token optimization (payload compression, tool-schema compaction, chat-history selection) and one OTel trace per request with its cost attached
- Runtime controls: a single ingress so agents are never publicly reachable, ownership and grants, an MCP gateway that keeps tool credentials on the gateway, an “ask” mode that pauses an agent for human approval, and flow guards on depth, fan-out, token budget and cycles
- Deploy any A2A agent with `nasiko deploy`, plus a routing engine (embedding shortlist, rerank, LLM selection) and an embedded S3-backed OCI registry
- Desktop app for macOS, Windows and Linux, a web dashboard, a CLI and a self-hosted server
Languages
Known limitations
- Cursor CLI supports discovery and reporting, but model routing is “not yet” per the README
- Self-hosting needs Docker Compose V2 plus Postgres, Redis, an S3-compatible store (RustFS) and an OTel stack; building from source needs Rust 1.88+, and the README requires changing the default admin password and three secrets before anyone else can reach the server
- The free plan is capped at two cloud agents and excludes writable and public agents, ACLs and multi-agent flows; Individual ($20/month) raises it to ten cloud agents and Pro ($99/month) makes it unlimited
- English-only interface and docs; this listing quotes no user or growth numbers








