MCP Studio settings

Every setting in MCP Studio: what it does, its values and its default. They reach every server you run, hosted or installed, when you press Publish. Until then they're drafts.

For the settings of the server process itself, which the person running it chooses, see configuration. Those can narrow what you publish here, but never widen it.

Each tool

Set per tool in the tools list. They follow the endpoint, not the tool's name, so a rename in SDK Studio keeps them.

When an agent calls itchoicedefault: Runs on its own for reads, Asks first for the rest

Value	Label	The agent
`auto`	Runs on its own	Calls it without asking anyone
`confirm`	Asks first	Asks the person at the client to approve each call
`off`	Off	Never sees the tool

A read is GET, HEAD or OPTIONS. See permissions and limits. CLI: octri mcp permissions.

Only adds dataon or offdefault: off

Tells clients a tool that isn't a read never deletes or overwrites, so it isn't marked destructive.

Customize for agentsname, description, tags and fields

A tool's name (letters, digits, _ and -, up to 64), its description (up to 4,000 characters), up to 10 tags, and per argument: hide it, pin its value, or rewrite its description (up to 1,000 characters). See customizing tools.

Listing the tools

How agents see the toolschoicedefault: Automatic

Value	Label	Lists
`auto`	Automatic	Every tool, or search mode once there are more than the threshold below
`all`	Every tool	One tool per operation
`search`	Search	Three tools that find, describe and call the operations
`code`	Code	Two tools for scripts against your SDK. Needs code mode, below

See search and code mode.

Switch to search abovenumberdefault: 40

How many operation tools Automatic lists before it switches to search, from 1 to 500.

Turn on code modeon or offdefault: off

Lets agents write short scripts against your SDK with describe_sdk and run_code. On plans that include it.

Code mode languageslistdefault: TypeScript and Python

Which SDKs scripts can use: typescript, python or both. The first is the default when a script names none.

Safety

See permissions and limits.

Cap writes per sessionnumberdefault: no cap

Writes one agent session can make, from 1 to 10,000. A write is any call that isn't a read, unless you mark a tool otherwise. The page suggests 50.

Cap calls per minute per toolnumberdefault: no cap

Calls each tool takes per minute, from 1 to 6,000. The page suggests 60.

Per-tool limitsper tooldefault: none

For one tool: its own calls per minute, and whether its calls count as writes. A POST that only searches can count as a read.

Keep an audit log of writeson or offdefault: on

Records every write an agent attempts, approved or not, with its arguments redacted. On plans that include it. CLI: octri mcp audit.

The hosted server enforces the caps across every session, group and installation. A server someone runs themselves enforces them in its own process, as best it can.

What agents are told

See agent experience.

Warn agents about outageson or offdefault: on

When Octri monitoring sees an endpoint down, its tools say so, and a call that fails during the outage is told to retry later.

Tell agents what changedon or offdefault: on

Tools that changed since the agent last connected carry a note, what_changed gives the detail, and a call to a renamed or removed tool is told where it went.

Error guidance by statusmessage per statusdefault: none

What an agent is told when a call fails with a status: a code such as 404, or 4xx and 5xx for the rest. Up to 20 statuses and 600 characters each.

More tools

Custom tools, prompts and resourceslist

Chained tools that make several calls in order, prompt templates, and resources such as your spec or chosen guides. See custom tools.

Make a workflow from every published guideon or offdefault: off

Off makes workflow tools only from the guides you choose. On makes one from every published guide Octri can read as a workflow. See workflows.

Turn on practice modeon or offdefault: off

Serves a no-sign-in practice server that answers every call with sample data, and lets anyone running the server ask for the same with --practice. See practice mode.

Servers and access

Serverslist of groupsdefault: none

Tool groups: smaller servers cut from the same tools, each with its own address. Up to 20, each with a name (up to 60 characters), a slug, a description (up to 500), the tags and tools it serves, and whether it includes the documentation tools or only tools that read.

Sign-inchoicedefault: credential headers

How people sign in to the hosted server: credentials sent as headers, your own OAuth authorization server, an OAuth proxy to your identity provider, or a page where each person enters their API key. In a sign-in mode, Still accept credential headers keeps the header route open too, and is on by default. See hosted.

Hostnamedomaindefault: none

Your own address for the hosted server, such as mcp.acme.com. See custom domain. CLI: octri mcp domain.

Offer the pluginon or offdefault: off

Publishes a Claude Code plugin for your server, with your prompts as slash commands. On plans that include it. See packages. CLI: octri mcp plugin.

Configuration: the server process's own settings.
Tool reference: every built-in tool.
octri mcp: the MCP commands of the CLI.