Settings

Settings controls runtimes, model roles, OpenAlex, stored projects, MCP servers, privileged tools, and editor/preview appearance.

Settings Runtime tab showing saved API and CLI runtimes, detected runtimes, Agent Model preset, OpenCode preset, context, reasoning, and legacy Chat Completions checkbox.
Runtime settings connect Systematic Reviewer to local APIs, external APIs, CLI runtimes, and model presets.

Runtime and model presets

Runtime settings decide which model backends the project can use. Scan only detects runtimes that are already installed or reachable; it does not install models or create accounts for you.

Scan local runtimes

Looks for common local APIs and installed CLI runtimes. Run it after installing or updating a runtime, then use + or the role buttons to save only what you want the plugin to use.

Add runtime

Adds an API endpoint manually. Use the base URL your provider gives you, usually ending in /v1, plus an API key only when required.

Saved Responses runtimes

Stores API base URLs and optional API keys for providers that expose a Responses-compatible endpoint. New API runtimes default to Responses.

Saved CLI runtimes

Stores trusted installed CLIs, such as Codex or OpenCode, for model-backed chat and extraction through the plugin bridge.

Agent model

The default model for chat and agent workflows. Additional agent presets can be selected from the chat Model modal.

Data extraction model

The default model for extraction jobs. It can be separate from the chat model or inherit an agent model preset.

PDF/VLM model

The model used for page-image or vision-language conversion when the PDF mode needs vision. Fast PDF does not need this model.

Embeddings model

The model used by the Embeddings tab. It uses the embeddings endpoint rather than the chat/extraction route.

Legacy Chat Completions checkbox

Use only for a specific API model when the provider says that model is exposed through Chat Completions or is not available through Responses. This applies to that model preset, not the whole runtime.

Reasoning

Default sends no reasoning field. Explicit values depend on provider and model support. If you are unsure, leave it on Default.

State mode

Stateless resends the managed chat context and lets the app control truncation. Stateful asks the server to keep the chain. Local runtime testing usually works best with Stateless, and CLI runtimes are stateless only.

Context window assumption

Set the real loaded context length for the model. The app uses it for token budgeting and sliding-window truncation, so an incorrect value can make long chats or automations fail.

Independent resources

Choose Yes only when that model does not compete with other configured models for the same GPU, CPU, memory, or runtime queue.

Parallel requests

Controls how many jobs the runtime can handle at once. Keep it low unless you know the runtime and hardware can run multiple requests reliably.

Suggested local model examples

The help overlays list example GGUF conversions that may be useful with runtimes supporting llama.cpp-based engines and compatible API endpoints. These are examples only; Systematic Reviewer and OpenResearchTools are not affiliated with model providers, Qwen, or LM Studio.

If you use external APIs, hosted providers, or CLI runtimes, especially subscription-based services, confirm that the provider terms allow use from external applications and permit the intended review, automation, and data-extraction workflows.

Harvest and OpenAlex

The Harvest settings tab stores the OpenAlex API key and points to OpenAlex documentation. OpenAlex is an open scholarly metadata index covering works, authors, sources, institutions, topics, funders, and related research metadata. Systematic Reviewer uses it in Harvest to search scholarly records, collect candidate studies, and add metadata-backed items into the project workflow.

Settings Harvest tab with OpenAlex API key settings and help text.
Add OpenAlex credentials here before running serious harvest jobs.

An API key is optional for opening the UI, but recommended for normal project harvesting because it identifies requests and gives access to OpenAlex authenticated rate limits. Systematic Reviewer and OpenResearchTools are not affiliated with OpenAlex; use the linked OpenAlex documentation for account, authentication, and rate-limit details.

Projects

Projects lists plugin-managed project records. This Settings view is project-independent, so it can manage stored projects even when another project is currently open.

Settings Projects tab showing Demo and testing project cards with Open project, Open folder, Reconcile, Delete, and project paths.
Project actions operate on plugin-managed project storage, not on unrelated Zotero collections.

Open project

Opens that stored project in the project workspace. It is disabled if the linked Zotero collection is missing.

Open folder

Reveals the project folder on disk, including the database, settings, report, log, snapshots, outputs, templates, and project-managed files.

Reconcile

Queues a background job to repair and refresh stored project links, Zotero project artifacts, and workflow files from project metadata. Track progress in Jobs.

Delete

Opens a confirmation dialog. It is destructive and should be used only when you no longer need the project storage.

Delete project confirmation modal with warning text, optional collection container checkbox, Cancel, and Delete project buttons.
Delete removes plugin-managed project data only after confirmation. The optional collection-container checkbox does not delete library items.

MCP servers

The built-in Systematic Reviewer MCP server can expose project tools to external MCP-compatible clients. User-added external MCP servers are third-party processes or services. They may read, write, or act according to their own behaviour and the permissions of your operating system account.

Settings MCP Servers tab showing built-in MCP server settings and external MCP server area.
Enable MCP only when you understand which external client or server will connect to the project tools.

Built-in MCP server

Starts Systematic Reviewer's own optional endpoint so trusted local agent clients can connect to the Zotero project environment.

MCP API key

Leave blank for local callers without bearer authentication, or set a key to require Authorization: Bearer <key>.

Copy endpoint / MCP JSON

Copies connection details for an MCP client. Save settings first so the current endpoint is accurate.

External MCP connectors

User-added third-party MCP servers may provide tools, resources, or prompts. Depending on the server, they may read or write files, run commands, access local services, call remote APIs, or send data outside the machine.

Privileged tools

Shell, browser, and developer tools are restart-gated. Enabling a privileged tool group shows a warning and requires a Zotero restart before that tool surface is loaded. Disable them when you do not need them.

Privileged shell tools warning. These tools are not recommended on systems that own or can access secure data. Shell commands default to running from the currently bound project workspace, but that does not itself contain the agent to the project.

An agent can still write scripts, execute them, call globally installed tools such as Python, shells, package managers, or other binaries, and direct those tools to act outside the workspace.

In practice, enabling shell tools grants the agent whatever system-wide access your user account and operating-system sandbox allow. Privileged shell tools can also be vulnerable to prompt injection or unsafe instruction-following.

The shell namespace is not available to the MCP server. In Systematic Reviewer, privileged shell access is intended for the in-app session agent when loaded at startup.

Shell commands default to the currently bound project workspace, and when developer tools are also enabled the localhost developer testing surface mirrors the same allowed tool set so you can test what the app agent can actually use. That workspace default is only a starting location, not a full safety boundary.

If you enable shell tools, prefer high-capability models with stronger prompt-security behavior and use them only in sandboxed environments. This software is provided without warranties. Nothing here is legal, security, or compliance advice. Consult your IT or security team before using any software.

Settings Privileged Tools tab with shell, browser, and developer tool groups disabled and showing current load states.
Keep privileged tools disabled unless you explicitly need them. Loading changes apply only after fully closing and reopening Zotero.
Privileged shell tools warning modal requiring the user to acknowledge risks before enabling and choosing restart later or restart now.
Enabling a privileged tool group requires reading the warning, acknowledging the risks, and restarting Zotero.

Privileged shell tools

Let the in-app agent run local shell commands with your OS user privileges. Commands start from the current project workspace, but that is not a full safety boundary.

Privileged browser tools

Let the in-app agent open live webpages, read documentation, follow links, and save webpage evidence into the current project using Zotero-native save flows.

Developer tools

Unlock localhost developer inspection and the dev-only testing surface. If shell or browser tools are also enabled, the developer surface mirrors the same allowed tool set.

Restart now / later

Tool bundles are loaded only at app startup. Restart now applies immediately; restart later saves the preference for the next full Zotero restart.

Editor/Preview page colour

When Zotero is in dark mode, the Editor/Preview page setting can switch the on-screen Writer page between a light page and a dark page. This is a display preference for editing and previewing; exports keep their export formatting.

Common checks

  • Click Save after changing settings that are meant to persist.
  • Run Scan local runtimes after installing or updating a local runtime.
  • Use the help question marks for model, runtime, MCP, and privileged tool warnings before enabling advanced features.
  • Confirm third-party provider terms before sending research material to external APIs, hosted models, CLI runtimes, or MCP servers.