NylorunDocsBeta
RunStudio

Studio

Open the local dashboard to run sessions and see a Tenant's Agents, manifests, and settings.

nylorun start prints Studio's URL, http://localhost:4161 by default. Open it in a browser on the same machine: Studio needs no login on its loopback address. To open it from the terminal:

npx nylorun studio

Behind a sign-in proxy, or embedded in another app, Studio keeps that sign-in.

Studio is the dashboard for this Tenant only, and each Tenant has its own port. It opens on the Agents list. A Tenant with no Agents yet shows its name, id, and the Runtime's status; the Agents appear when the first is made.

ItemOpens
The Tenant, at the topThe Tenant page, beside the Runtime's status.
AgentsThe Agents list.
Each Agent under itIts Agent page. A dot shows its state, and a count shows only when its sessions wait on a person or fail.

Hovering an Agent shows +, which opens a new session on it as local-developer. An Agent whose owner guard names another owner offers none.

Sessions, manifests, models, and artifacts have no sidebar item: each opens from the Agent or the Tenant it belongs to. Links to /sessions, /agent-manifests, /settings, and /artifacts still open their place.

Agents list

The Agents list (/agents) has one row per Agent: its id and labels, the manifest version it runs (triage@3, tracking the manifest or pinned), its state (the reasons in a tooltip), its kind, its sessions and how many wait on a person, and its last activity. Agents that need input come first, then failing ones, then the rest by last activity. Search by id or manifest, and filter by state, manifest, owner, or label; selecting a manifest or a label in the table filters on it. Each row's ⋯ opens the Agent, opens a New session on it, or copies its id.

New Agent makes one from a saved manifest: the manifest, Track the latest or a version to pin, its id, its owner guard (local-developer, the owner of Studio's sessions, by default), and labels. Making an Agent opens no session.

Agent page

An Agent's page (/agents/<id>) is a dashboard of the manifest version it runs. Its header names that version (linked to it), whether it tracks the manifest or is pinned, its kind, its owner guard, and its labels, with New session and a ⋯ menu:

ActionDoes
Edit labels and ownerChange its labels and owner guard.
Stop computeFor a pod: stop its sandbox.
Reset diskFor a pod: start it again on a new, empty volume.
Fork into a new AgentMake a new Agent from a manifest version, with a disk that starts as a snapshot, a disk version, or empty, and the secrets the new version declares. Check shows the plan first. See Fork.
Delete AgentDelete it, its secrets, its disk, its disk versions, and its memories.

Reset and Delete say how many sessions run on the Agent before they act.

Each tab is a table with a search, filters, and a count. A row's actions are behind its ⋯, and a row opens its details in a panel on the right. The page opens on Sessions:

TabShows
SessionsThe sessions on it, by last activity, filtered by who started them and their status. A row opens the session.
TriggersIts triggers: each one's schedule in words, the session its runs go to, its next run, and its status, with Run now, Pause, Resume, and Copy webhook URL. Below, every run, newest first.
ArtifactsWhat its sessions made, and the input files its runtime copies in. A row opens the artifact.
MemoryIts memory: each memory with its description, version, and who changed it last. A row opens its content and history; Delete removes it.
DiskIts disk versions, newest first: Snapshot now keeps one, and a row's Restore, Fork into a new Agent, Delete, and Browse (its files and changes). Not shown for kind none.
ActivityIts events, newest first: its lifecycle and the connections its sandbox opened, filtered by type and result.
VaultIts secrets: each slot its manifest version declares, with its type, what uses it, and whether it has a value. Set value asks only for the value; the slot gives the rest. Values are never shown. A secret no slot declares can only be deleted.
ManifestThe version it runs as a readable spec: the harness section (instructions, tools, skills, MCP servers, memory), then the runtime section (compute, lifecycle, network, setup and input files, secret slots, triggers, limits). Each HTTP tool and MCP server notes the tool policy rules on it. A row opens its part of the manifest as JSON.

The Manifest tab's ⋯ moves the Agent within its manifest (see Versions), shows the whole manifest, or opens all its versions:

ActionDoes
Pin / Track the manifestKeep it on the version it runs, or follow every version saved.
Move to vNFor a pinned Agent, move it to another version.
RestartFor a pod that waits to start on a new version: apply it now.
RetryFor a pod whose start on a version failed: try that version again.

Links to the tabs an Agent page had before still open their place: ?tab=overview opens Sessions, secrets Vault, setup Manifest, and events Activity.

Studio is one more client: any credential that may write Agents does the same with the CLI or the SDK.

The Tenant

Selecting the Tenant at the top of the sidebar opens its page (/tenant). Its header names the Tenant, its id, the model it runs, its sandbox, and how many Agents and sessions it holds. Its ⋯ copies the id and resets the Tenant: Reset sessions…, Reset Agents…, or Reset everything…, each after a confirm, as nylorun tenant reset does (see Reset a Tenant).

The tabs follow in the order an Agent's turn meets them. The page opens on Status:

TabShows
StatusThe Tenant's checks (store, scheduler, model, schema), what runs it (harness, sandbox, event streams), and what it holds, each row linked to the tab that configures it.
ManifestsIts agent manifests.
ModelsThe providers it has a credential for, the one it runs, and what each model reads. See Models.
ComputeThe sandbox the Runtime chose and the backends it found, the most an Agent's runtime section may ask for, where harnesses may run, and what artifacts and skills may hold. Edit compute settings… saves only what it changes.
Tool policyThe tool policy's rules, strictest first. A row opens one to edit; Add rule adds one. The policy is saved whole.
LimitsThe Runtime limits by area, each beside its default, with a changed one in bold, including an Agent folders group. A row opens one to edit in its unit; Reset to default drops it.
UsageWhat model calls used today, this month, and in all, and the budgets with what each period used. A row edits a budget; Add budget adds one.
TrafficHow the Runtime is running, read live from the runtime's and the gateway's /metrics every 10 seconds: API requests by route and client, model calls and tokens, MCP tool calls, turns, queue waits, egress, and dropped telemetry, with counts since each process started, the current rate, and p50 and p95. Nothing is stored. See Monitor.
KeysThe application keys and signing keys, read-only: make, rotate, and revoke them with nylorun tenant keys and nylorun tenant signing-keys.
AuditThe Tenant's audit, newest first: keys, signing keys, tool policy and settings changes, Agents' secrets written and used, and Studio sign-ins, each with who acted. A type filter and Load more. See Monitor, Audit.

Manifests

The Manifests tab lists the Tenant's agent manifests: each name with its description, its latest version, the kind of Agent it makes (flow · … for a flow), how many Agents were made from it, and when it was saved. A version saved from an agent manifest is marked beside its number. Search by name or description, and filter by kind, type (Agent or flow), and whether any Agent was made from it.

A row opens the manifest beside the list, with a version picker, its description, the Agents made from it (each with the version it runs), and two views of the version:

ViewShows
SpecThe version as a readable spec, its Harness section, then its Runtime section, one line each. A line opens its part of the manifest as JSON.
SourceThe files the version was saved from, read-only: its files as a tree, each marked when it was added or changed since the version before, with the removed ones last. A file opens beside it as text, or as its size when it is not text. A version saved as JSON has no source, and says so. See Compare versions.

The manifest's ⋯ makes an Agent pinned to the version shown (New Agent from vN), shows the whole manifest as JSON, or copies its name.

Manifests are not edited in Studio: edit an agent's folder in its repository, and save its next version with npx nylorun apply or Upload folder… on the tab's toolbar. Upload folder… picks a directory in the browser, leaving out .git, .hg, .svn, and node_modules as nylorun apply does, uploads the files the Runtime does not hold yet, and shows the plan of the save: what it would change, its errors and warnings, and the compiler's diagnostics by file, line, and column. Save vN saves it, refused if another save came in between, and opens its Source view. The manifest takes the name AGENT.md's frontmatter gives it, or else the directory's.

Session

A session's page is /agents/<agentId>/sessions/<sessionId>. Its header shows its Agent, the session's short id (click to copy), its status, and a button that hides the inspector, which the browser remembers. The chat opens at the newest message. Assistant replies render as Markdown (tables, lists, and code; raw HTML stays text), and a message shows its time on hover.

The message card holds the text and one button, with the model picker under it. Enter sends and Shift+Enter adds a line. While a turn runs, the box stays focused but read-only, and the button stops the turn, as session.cancel() does. To attach files, use the paperclip, or paste or drop them onto the card. Any file goes: each shows as a chip until the message is sent, and Studio uploads each into the session and sends the message as text and file parts. See Files.

The inspector has four tabs, the same for an agent and a flow:

TabShows
ManifestWhat the agent is. Declared: the manifest the session runs, named name@version, with its instructions, tools, skills, and MCP servers, each skill and server with its plugin's badge; it says when the manifest has a newer version. Deployed: where it runs: its Agent (linked to its page), kind, state, and compute, the version the session opened on beside the one the Agent runs, the harness, sandbox, and allowed hosts, what its MCP servers and HTTP tools get from the Agent's secrets (coverage), its latest disk version, the connections its sandbox opened during this session's turns, and the other sessions on it. Read-only: change the Agent on its own page.
RecordThe session's artifacts: its outputs and the files sent into it.
EventsLog: history and live events, the same stream session.observe() reads, in three views: Agent (what the agent saw and did, the default), Runtime (what the Runtime did to the session), and All; each row shows its writer and evidence. Trace: a turn's trace as a waterfall, the turn, each model call with its model and tokens, and each tool call; select a span for its attributes, and Open event for the log event it came from. An event's details offer Open trace. A flow adds Progress (its stages, live) and Iterations, and opens on Progress.
UsageWhat the session cost. A Session card with one tokens bar and a ledger of tokens and cost by kind (cached input, uncached input, cache writes, output) adding up to a total; then the model calls by turn, each with cached, uncached, output, and total tokens, time, and cost. A call's View event opens its event in the Events tab. See usage by token kind.

Links to the tabs a session had before still open their place: ?inspector=agent opens Deployed, artifacts opens Record, and tree and iterations open the Events tab's flow views.

Each tool call, in the chat and in the event details, shows its kind, its server or skill, and its plugin's badge. An MCP tool shows as server / tool, with the name the model called it by when that differs, such as github / create_issue (as github__create_issue). A tool call whose result was stored as an artifact links to it.

A turn paused by the tool policy shows an approval card for each held call, with its tool by the same name, why it is held, and its input. Approve sends that exact call; Deny fails it to the model as tool.denied.

Turns that triggers and wakes start show as ordinary sessions of the Agent.

Artifacts

An artifact opens from where it belongs: a session's Record tab, or an Agent's Artifacts or Disk tab. The inspector shows a numbered version with its owner, its history, a preview, and the changes between versions, and downloads it. /artifacts lists every artifact the Tenant holds, newest first, filtered by kind, session, and labels.

Security

Studio is for operators on this machine.

  • It listens on loopback only (127.0.0.1:4161 by default).
  • No application or provider key reaches the browser.
  • Reach it on another machine through an SSH tunnel. Do not publish the port.

Telemetry

Studio sends Google Analytics page views by default. It is off in CI. Turn it off with any of:

npx nylorun local telemetry disable

NYLORUN_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 also disable it.

Next step

On this page