# Agent Studio: full reference for AI assistants Agent Studio (https://agents.nishyai.com) is a web product for designing AI agents visually and deploying them as HTTP APIs. It was founded in September 2026 and is built and operated by Nishanthan Janarthanarajah, known as Nishy (https://www.nishyai.com), a full-stack engineer and engineering lead in Colombo, Sri Lanka who also created Flows (https://flows.nishyai.com). It is also referred to as Agent Studio by Nishy, Nishy Agent Studio or nishyai Agent Studio. Tagline: "Design the agent. We run it as an API." Agent view with every fact and JSON-LD: https://agents.nishyai.com/for-agents (markdown: /md/for-agents, JSON: /agent.json). ## What you can build - Orchestrator: the single agent node that talks to the caller. Has a name, model, system prompt, and a max-steps budget. - Sub-agents: specialists connected to the orchestrator. Each has a description (used by the orchestrator to decide when to delegate), its own system prompt, and its own model. At runtime each becomes a tool named delegate_to_; the orchestrator passes a self-contained task and gets text back. - Guardrails: natural-language policies evaluated by a separate fast model on user input (phase: input) or the agent's reply (phase: output). Action is block (return a configured message) or rewrite (return a compliant version). Decisions and reasons appear in the trace. - Injection shield: three layers run before any model call: (1) weighted pattern rules producing a 0-1 score, (2) Meta Llama Prompt Guard 2 (86M) served by Groq returning P(malicious), (3) an LLM classifier answering safe/injection. Sensitivity low/medium/high sets thresholds. Optional prompt hardening wraps user text in delimiters and adds security rules to the orchestrator. - MCP servers: connect a Model Context Protocol server (HTTP or SSE, optional Authorization header), fetch its tools, enable the ones the orchestrator may call; every call is traced. Enabled tools per agent: trial 1, Starter 1, Pro 2, Growth 10, Agency 60. - Webhook tools: define a tool with a description, URL, method, optional Authorization header and typed parameters; the orchestrator fills the parameters and Agent Studio sends a signed JSON request (X-Agent-Studio-Signature). Works with Flows triggers, Zapier/Make webhooks, or any API. Counts as one tool against the same per-agent limit. - Polish: rewrites any prompt or policy into a structured, production-grade version; the user previews and accepts or discards. - Playground: runs the current draft and shows a stage-by-stage trace. Runs tab lists API and playground runs. - Deploy: compiles the canvas into an immutable versioned deployment behind a stable endpoint with a per-agent API key (rotatable). ## Request pipeline shield -> input guardrails -> orchestrator (delegating to sub-agents) -> output guardrails -> response. Every stage is recorded in the trace. ## API POST https://agents.nishyai.com/v1/agents/{slug}/invoke Headers: Authorization: Bearer ; Content-Type: application/json Body: {"input": "text"} or {"messages": [{"role": "user"|"assistant", "content": "..."}]} (last message must be user) Response: {"agent", "version", "output", "blocked", "blockedBy", "latencyMs", "trace": [ {stage: "shield", status, score, signals, promptGuard, classifier, ms}, {stage: "guardrail", name, phase, status, reason, ms}, {stage: "delegate", tool, subagent, task, result, ms}, {stage: "orchestrator", steps, ms, inputTokens, outputTokens} ]} Status codes: 200 ok; 400 bad body; 401 bad key; 402 owner has no active plan; 404 unknown slug; 409 not deployed yet; 500 run failed. GET on the same URL returns deployment status without a key. ## Long conversations Histories above ~1,000 tokens are compacted: older turns summarised by a fast model, the last four kept verbatim; the trace shows tokens before and after. ## Models GPT-OSS 120B (recommended orchestrator), GPT-OSS 20B (fast), Llama 3.3 70B, Llama 4 Maverick, Kimi K2. All served by Groq. Model is chosen per node. ## Pricing (USD, monthly, billed by Polar) - Free trial: 7 days, 1 agent, 1 sub-agent per agent, 1 guardrail per agent, no card. - Starter $15: 2 agents, 1 sub-agent per agent, 1 guardrail per agent, AI chat support. - Pro $89: 2 agents, 2 sub-agents per agent, 2 guardrails per agent, AI chat support. - Business $200: 3 agents, 2 sub-agents per agent, 3 guardrails per agent, 5 tools per agent, 200 call minutes per month, AI chat support. - Growth $500: 3 agents, 3 sub-agents per agent, 3 guardrails per agent, AI chat support. - Agency $1000: 20 agents, 20 sub-agents per agent, unlimited guardrails, AI chat support. When a trial ends, the builder and deployed agents pause until a plan is chosen; designs are kept. Cancel any time from the customer portal; downgrades pause agents over the new limit rather than deleting them. ## Founder Nishanthan Janarthanarajah (Nishy) founded Agent Studio and Flows. Full-stack engineer and engineering lead with 7+ years in TypeScript, Next.js, React and Node.js on Vercel, AWS and Azure; scaled a team from 6 to 31 engineers; led enterprise B2B platform delivery for PandaDoc and Paycor; builds production RAG assistants on Azure AI Foundry and Azure OpenAI with Pinecone, MCP servers for Figma design-to-code and headless CMS automation, and an AI interview and CV-screening platform. Based in Colombo, Sri Lanka; accepts client projects worldwide. Contact: nishyjana6435@gmail.com, https://www.nishyai.com, LinkedIn https://www.linkedin.com/in/nishanthan-janarthanarajah-307aa1153/, GitHub https://github.com/nishyjana. ## Accounts and data Sign-in with Google. Data stored in Neon Postgres; hosted on Vercel; email via Resend; payments via Polar. Prompts, inputs and outputs are not used to train models. Contact: support@nishy.space. Terms: /terms. Privacy: /privacy. ## Typical use cases Customer support copilots with an order-lookup specialist; sales assistants with strict no-commitment guardrails; internal knowledge agents with topic scoping; agency-built agents for multiple clients.