Models, permissions, skills, and MCP

© 2026 Quadropic Pvt Ltd. Proprietary and confidential. Use restrictions.

Think in model lanes#

Syntheo assigns models by purpose. The composer exposes Standard, Hunter, Design, Browser, and up to two custom presets; Settings also contains the internal Light lane.

With Standard selected, the harness automatically routes Ask to Light and Architect/Debug to Hunter. An explicitly selected non-Standard preset remains selected for the action.

For subscription-backed access, connect an eligible Claude Pro/Max account through Anthropic OAuth, a ChatGPT Plus/Pro account through OpenAI Codex OAuth, or a Copilot account through GitHub Copilot OAuth. See Use existing AI subscriptions.

Configure assignments#

Open Settings → AI Provider → Model Assignments.

For each lane:

1. Select a provider.
2. Select a model.
3. Set thinking level if the model advertises reasoning support.
4. Configure the provider credential on its first occurrence.

Thinking levels range from Minimal through X High. Increase thinking when the problem is ambiguous, cross-cutting, or correctness-sensitive. Lower it for mechanical edits and quick answers. A higher level can cost more and take longer; it does not fix missing evidence or a vague brief.

Create custom presets#

You can add up to two custom mode presets. A custom preset starts from Hunter’s current assignment, then keeps its own concrete provider/model/thinking configuration.

Useful custom presets:

• “Security review” on a strong reasoning model;
• “Fast refactor” on a low-latency code model;
• “Vision QA” on a vision-capable model;
• “Docs” on a model with strong writing quality.

Do not create a preset for a one-off task. Select Hunter or adjust the action instead.

Connect custom endpoints#

Use Add provider for a local model or compatible proxy. Pick the correct API dialect and enter truthful capability metadata. The context meter and image/reasoning controls depend on those values.

For local endpoints:

• verify the server is running before launching a task;
• use the endpoint’s real model ID, not only its marketing label;
• set the context window conservatively;
• do not mark vision or reasoning support unless the endpoint accepts those inputs;
• use an environment-variable name when the server requires a key.

Global behavior settings#

Open Settings → Behaviour.

You can set:

• Verbosity — Default, Detail, or Point;
• Personality — Default, Relax, or Strict;
• Additional Instructions — stable preferences applied to agent behavior;
• a ready-made behavior preset.

Good global instructions are stable personal preferences:

When changing code, state assumptions and finish with the exact validation run.
Prefer TypeScript examples. Keep final summaries under 12 lines unless risk
requires more detail.

Poor global instructions are project-specific or temporary:

Always edit the billing reducer and run the migration from issue 421.

Put project rules in the repository and task-specific rules in the prompt. Conflicting global instructions add noise to every run.

Tool permissions#

Open Settings → Permissions to see every permission-controlled built-in and installed MCP tool.

Modes:

• Ask — pause and request a decision;
• Allow — execute without prompting.

The settings screen lists the tools whose permission mode you can configure.

Trust in layers#

Start with:

• read/search/navigation tools: Allow if their scope is clear;
• file writes and shell: Ask;
• browser/network: Ask;
• Git or deployment tools: Ask;
• newly installed MCP tools: Ask.

After observing several safe calls, you may allow a tool globally. Evaluate the tool’s maximum capability, not the harmless example you first saw.

For example, a generic shell tool can delete files, access local credentials available to the process, and contact programs on the machine. Its risk is not limited to npm test.

Per-call decisions#

The approval card shows tool arguments. Choose:

• Allow Once for a single known operation;
• This Turn for repeated calls during the active run;
• Deny when the target, scope, or consequence is wrong.

If you deny, give the agent a path forward:

Do not run the full deployment command. Use the local dry-run validator instead.

Skills#

A Skill is a reusable instruction package. It can contain procedures, references, scripts, and assets.

Install a Skill when:

• the same specialized workflow recurs;
• a task requires precise organization-specific steps;
• the model needs a tool-specific procedure;
• you want a reviewable, versioned operating playbook.

Do not use a Skill as a dumping ground for a project’s entire documentation.

Add a Skill#

Open Settings → Marketplace → Skills → Add skill.

Options:

• preview and install from a skills.sh page, repository, Git source, or local path;
• create a private local Skill with a name, description, tags, and instructions.

Review the preview’s selected skill path and files before installing. Update source-installed Skills when their upstream procedure changes.

Invocation mode#

Model invoked: The Skill is automatically discoverable in normal agent prompts. Use this for broadly relevant, trusted Skills with clear descriptions.

User invoked: The Skill stays out of normal prompts until explicitly selected with @. Use this for niche, sensitive, expensive, or easily misapplied workflows.

Prefer user invocation when in doubt. It reduces prompt noise and prevents a specialized procedure from influencing unrelated tasks.

MCP servers#

MCP servers add tools from another process or service.

Supported connection styles:

• stdio — start a local command with arguments and environment;
• HTTP — connect to an HTTP MCP endpoint;
• SSE — connect to an SSE endpoint.

To add one:

1. Open Settings → Marketplace → MCP → Add MCP.
2. Choose the transport.
3. Enter the command/arguments/environment or URL.
4. Choose Model invoked or User invoked.
5. Add the server.
6. Refresh it to discover tools.
7. Inspect tool names, descriptions, and any refresh error.
8. Review those tools in Settings → Permissions.

For stdio environment entries, use one KEY=value per line. Treat those values as secrets even if the UI stores sensitive fields securely.

MCP invocation mode#

Model invoked makes the server’s tools discoverable automatically. Use it only when the tools are broadly relevant and their descriptions are unambiguous.

User invoked exposes a tool only for a turn where you mention it with @. This is the right default for:

• production services;
• ticketing or communication systems;
• write-capable cloud tools;
• expensive queries;
• tools with overlapping names.

MCP safety checklist#

• Trust the server source and publisher.
• Understand where stdio commands execute.
• Use HTTPS/SSE over trusted endpoints.
• Keep credentials scoped and revocable.
• Inspect discovered schemas.
• Leave write/destructive tools on Ask.
• Disable or remove servers you no longer use.
• Refresh after upgrading the server so the cached tool catalog stays accurate.

Skill versus MCP versus prompt#

Skills tell the agent how to work. MCP tools give it something it can do. Many integrations need both: a tool server plus a Skill explaining the safe workflow.