Your runtime, visible and operable.

Same runtime your application owns. A browser console for the development loop. Studio wraps the agents and pipelines your application owns, then connects the surfaces needed to understand them — playground, review, context, traces, and live runtime state. Prompt agents, inspect tool calls, and follow local workflows without leaving the browser.

Prompt real agents

Exercise the agents and model policies your application already owns from one browser workspace.

Human review

Inspect agent tool calls, resolve approval requests, and answer bounded questions during a Playground run.

See runtime state

Explore knowledge, evidence, memory, MCP tools, sandbox workspaces, processes, and logs.

Follow local workflows

Inspect pipeline graphs, sessions, traces, models, run output, and local history.

Register the object your app already owns.

Studio opens a local server around the exact agent instance your product uses. Its model, instructions, tools, middleware, memory, and policy do not move.

studio.ts
1import { Agent } from '@anvia/core'2import { Studio } from '@anvia/studio'34const agent = new Agent({5  id: 'support',6  name: 'Support',7  model,8  instructions: 'Answer clearly.',9})10new Studio([agent]).start({11  hostname: '127.0.0.1',12  port: 4021,13})

Read the package guide

Studio surrounds the runtime. It does not replace it.

A Playground run crosses the same model, tool, memory, and approval boundaries as a run in your product. That makes a local reproduction useful evidence, not a Studio-only approximation.

  • Same runtime object

    Studio registers the agent your product already constructs. There is no second definition to synchronize.

  • Same capabilities

    Models, tools, middleware, memory, instructions, and approval policy remain application code.

  • Inspection layer only

    The local console makes runtime relationships visible without becoming another application framework.

One run. Every clue keeps its identity.

A streamed tool call belongs to a message, that message belongs to a session, and the session can be followed into its trace, context, and replay.

  1. 01

    Prompt

    Start a run against the exact agent object registered by your application.

  2. 02

    Follow the stream

    Watch ordered text, tool arguments, results, usage, and errors arrive.

  3. 03

    Intervene

    Approve protected work, answer a human question, or cancel safely.

  4. 04

    Inspect

    Move from message to session, trace, memory, context, or live Sandbox state.

  5. 05

    Replay

    Correct an input or failed pipeline step and run the linked workflow again.

Follow the Studio workflow

Prompt, inspect, and intervene.

This illustrative preview uses a sample conversation to show messages, tool calls, and results together. Explore the full inspection views in the Studio guide.

http://127.0.0.1:4021/playground

Workspace/Chat/assistant

Y
You

Summarize failed checkout runs from the last hour.

Assistant

I’ll inspect recent traces and group the failures by cause.

trace.search Complete
status:error  since:1h
18 runs · 3 failure groups
Assistant

Most failures came from expired payment sessions (9), followed by inventory locks (5) and address validation (4).

Not one dashboard. Connected views.

Every surface answers a different debugging question while preserving the relationships between a run, its state, and the capabilities it touched.

Playground

Stream a real response, inspect mixed message parts, and handle approvals or human-input requests in context.

streamingapprovalsusage

Pipelines + tools

Read the graph, inspect inputs and outputs, replay a run, or invoke a selected tool handler with explicit arguments.

graphsschemasreplay

Sessions + traces

Follow a streamed call back to its message and session, then inspect the linked trace, timing, model, and token usage.

historyidentitymetrics

Context + memory

Inspect static knowledge, dynamic context and tools, retrieval evidence, and records exposed by the memory adapter.

knowledgeretrievalmemory

MCP + status

See connected MCP servers and tools alongside registered agents, models, storage status, record counts, and enabled capabilities.

serversmodelsstatus

Live Sandbox

Browse workspace files, running processes, exposed ports, and logs while the agent is still doing the work.

filesprocesseslogs

Disposable by default. Persistent when useful.

Studio owns its browser UI, local HTTP and streaming routes, inspectors, and optional local stores. Your application keeps credentials, product data, production policy, and deployment authority.

Register exact objects

Runtime authority

Agents, tools, credentials, model clients, memory adapters, and approval policy stay in application code.

loopback runtime

Local inspection

Browser views, HTTP routes, shared stream semantics, human input, replay, and direct development tools.

127.0.0.1dev datalocal only
optional history

  • In memory

    Fast disposable sessions and traces with no database left behind.

  • SQLite when needed

    Session, trace, pipeline, and run history can survive a local restart.

  • Read-only memory view

    Discover conversations when the agent's memory adapter exposes an inspector.

Studio does not include authentication, authorization, or TLS. Keep it on loopback with development credentials and test data. Read the security boundary

Debug the boundary, not just the output.

Three focused loops for the moments when a transcript or a terminal log cannot explain what the agent system actually did.

01

Test a suspended approval

Prompt an agent to request a protected action, inspect the proposed arguments, approve it, and verify the linked completion.

  1. prompt
  2. tool request
  3. approve
  4. resume
@anvia/studio@anvia/core
Test human-in-the-loop
02

Diagnose a pipeline

Open the graph, inspect the failed step and its recorded input and output, then rerun the saved input or start a corrected run.

  1. input
  2. failed step
  3. inspect
  4. rerun
@anvia/studio@anvia/core
Explore pipelines
03

Inspect a live Sandbox

Connect an agent run to the files it changed, the process it started, its exposed port, and the logs it produced.

  1. run
  2. workspace
  3. process
  4. logs
@anvia/studio@anvia/sandbox
Explore Sandbox

Reproduce in Studio. Operate with Lens.

Studio shortens the local development loop. Lens is an optional production companion for observability, evaluation, and release decisions; your agent can also use another observability backend.

Local · interactive · disposable

Studio

Prompt real agents, resolve interactive states, inspect context, and replay local workflows.

Durable · self-hosted · operational

Lens

Investigate production traces, evaluate datasets, compare releases, and enforce quality gates.

Continue to Lens

Same runtime identity, from local reproduction to production.

Start with a real inspection surface.

Add Studio to the project that owns your agents, start the local console, and inspect the registered runtime in your browser.

  1. 01

    Install Studio

    Add @anvia/studio to the server-side TypeScript project that owns your agents.

  2. 02

    Register the runtime

    Pass the same agent objects to Studio and add a local development script.

  3. 03

    Open the console

    Start the process, then open 127.0.0.1:4021/playground in your browser.

Read the Studio guide

Anvia
GitHubGet started