> ## Documentation Index
> Fetch the complete documentation index at: https://koreai.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Arch MCP tools and skills

Arch MCP gives an MCP-capable coding assistant access to supported, permission-checked platform operations. The package supports project discovery, authoring, evaluation, debugging, and lifecycle work. To install it, see [Install and configure MCP server](https://koreai.mintlify.site/agent-platform/mcp-server.md). For working practices, see [Work with a coding assistant](https://koreai.mintlify.site/agent-platform/best-practices/work-with-coding-assistant.md).

The exact tools depend on the installed package version, the connected features, and your permissions. The live `arch://guidance/v1/manifest` is the current tool catalog.

## Tool catalog

For every tool, grouped by Build, Evaluate, Optimize, Debug, and Analyze, see [Build with Arch using MCP](https://koreai.mintlify.site/agent-platform/use-mcp-server.md). This page covers the guidance skills, guarded edits, and how each Studio task maps to a tool.

> [!WARNING]
>
> Arch MCP acts as the signed-in person, so project roles apply to every call. Some tools, such as `platform_agents`, combine read and write actions in one tool. Require approval for any tool that can change the platform, and auto-approve only tools that just read.

## Guidance skills

A skill is a set of reusable instructions that helps the assistant choose a safe, consistent procedure for a type of task. Skills don't add MCP tools or grant permissions. Install them with the guidance installer described in [Work with a coding assistant](https://koreai.mintlify.site/agent-platform/best-practices/work-with-coding-assistant.md#connect-a-coding-assistant-to-arch-mcp).

| Skill | What it helps with |
| - | - |
| `arch-agent-guide` | Connects the assistant to the project and routes requests to a specialist skill. Start a task here. |
| `arch-platform` | Plans, carries out, and verifies work through Arch MCP. |
| `arch-topology` | Chooses the agent structure, including flows, supervisors, and specialists. |
| `arch-tool-design` | Designs tool inputs, effects, confirmations, results, and error handling. |
| `arch-workflow-design` | Designs and troubleshoots workflows, mappings, timeouts, and failure paths. |
| `arch-voice-design` | Configures and evaluates pipeline and realtime voice experiences. |
| `arch-context-budget` | Manages prompts, history, memory, tool results, and context limits. |
| `arch-latency-cost` | Finds the sources of slow or expensive turns and suggests ways to reduce them. |
| `arch-knowledge-search` | Sets up and troubleshoots knowledge base retrieval. |
| `arch-experience-contract` | Defines success measures and latency or cost targets for an experience. |
| `arch-eval-catalog` | Builds evaluation scenarios, personas, evaluators, and run comparisons. |
| `arch-explain-session` | Explains a session trace and identifies where behavior first went wrong. |
| `arch-release-check` | Checks project readiness before publishing or deployment. |

Never edit the installed `arch-*` skill folders, because the installer manages them. To add your team's procedures, build a customer skill on top of them.

## How guarded edits work

`platform_agents:edit_dsl` uses the `expectedSourceHash` returned by the latest read of that agent. This compare-and-save guard refuses the edit if the definition changed after it was read, which protects a newer Studio or assistant edit from being overwritten. For a new draft, the tool accepts an explicit null hash.

After a stale-edit refusal, read the current definition, compare it with the intended change, reapply only the edits that are still needed, validate, and read the result back. The guard detects concurrent changes; it doesn't validate the agent's behavior. Use a whole-project import only for a whole-project package, not for a single agent edit.

## Where to find it

| Task | Studio | Arch MCP |
| - | - | - |
| Test while building | Agent → **Chat with Agent** → **Debug** | `debug_load_agent`, `debug_send_message` |
| Find and explain a session | **Project** → **Operate** → **Sessions** | `debug_session_history`, `debug_analyze_session` |
| What the model received | Session → **Traces** → LLM step → **Request** | `debug_traces` |
| Eval sets and runs | **Project** → **Evaluate** → **Evals** | `platform_eval_sets`, `platform_eval_runs` |
| Publish a version | Agent → **Version history** → **Publish Version** | `platform_versions` |
| Deploy, promote, roll back | **Project** → **Operate** → **Deployments** → **Environments** | `platform_deployments` |
| Channel binding | **Project** → **Operate** → **Deployments** → **Channels** | Studio only |
| Guardrails and PII | **Project** → **Govern** → **Guardrails**; **Project** → **Govern** → **PII** | `platform_guardrails` |
| Models and runtime settings | **Project** → **Settings** → **Models**; **Project** → **Settings** → **Runtime Config** | `platform_config` (models: read and set; runtime config: read) |
| Production approval policy | **Project** → **Settings** → **Governance** | Studio only |
| Production signals | **Project** → **Insights** | `platform_analytics` |
| Alerts | **Project** → **Operate** → **Alerts** | `platform_alerts` |


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.