fix(mcp): deconflict local and remote MCP servers

CLAUDE.md

@@ -107,11 +107,12 @@ When uploading Workers to Cloudflare's dispatch namespace, upload ALL files from
 
 ## Jack MCP (for AI Agents)
 
-**CRITICAL:** When Jack MCP is connected, always prefer `mcp__jack__*` tools over CLI commands or wrangler. MCP tools are cloud-aware and work with Jack Cloud mode where wrangler won't.
+When Jack MCP is connected, prefer `mcp__jack__*` tools for **cloud operations** (deploy, databases, logs, crons, domains, storage, vectorize) over CLI commands or wrangler. MCP tools are cloud-aware and work with Jack Cloud mode where wrangler won't.
 
-- Check your available tools for anything prefixed with `mcp__jack__`
-- Use those instead of `jack` CLI commands or `wrangler` commands
+- For deployments, databases, logs, and services: use `mcp__jack__*` tools
+- For file editing: always use built-in Read/Edit/Write tools, never remote MCP tools like `stage_file` or `read_deployed_file`
 - If a capability isn't available via MCP, ask the user to run it via CLI
+- The local MCP (`mcp__jack__*`) and remote MCP (`mcp__jack_cloud__*`) are mutually exclusive — never connect both simultaneously
 
 ## Deploy Mode: Managed vs BYO
 
@@ -238,3 +239,13 @@ Rules:
    - `findWranglerConfig()` in `wrangler-config.ts` — wrangler config path resolution (supports .jsonc, .json, .toml)
    - `parseJsonc()` in `jsonc.ts` — JSONC parsing (never use regex comment stripping)
    - `readProjectLink()` + `getDeployMode()` in `project-link.ts` — project link and deploy mode
+
+**After modifying shell integration (`shell-integration.ts`), regenerate the shell file for local testing.**
+
+`jack update` only regenerates `~/.config/jack/shell.sh` after a successful npm update — it won't pick up local dev changes. After editing `shell-integration.ts`, run:
+
+```bash
+bun -e "import { writeShellFile } from './apps/cli/src/lib/shell-integration.ts'; writeShellFile();"
+```
+
+Then open a new terminal (or `source ~/.zshrc`) to load the updated shell function. Without this, you'll be testing against the stale shell.sh from the last published version.

apps/cli/src/commands/mcp.ts

@@ -262,7 +262,7 @@ async function outputProjectContext(): Promise<void> {
 			lines.push("- Status: `mcp__jack__get_project_status` or `jack info`");
 			lines.push("");
 			lines.push(
-				"Prefer `mcp__jack__*` tools or `jack` CLI over raw `wrangler` commands for consistency.",
+				"For cloud operations, prefer `mcp__jack__*` tools or `jack` CLI over raw `wrangler` commands. For file editing, use built-in Read/Edit/Write tools.",
 			);
 		}
 
@@ -272,7 +272,7 @@ async function outputProjectContext(): Promise<void> {
 		);
 		lines.push("");
 		lines.push(
-			"**Always prefer `mcp__jack__*` tools over CLI commands or wrangler** — they are cloud-aware and work in all deploy modes.",
+			"**For cloud operations (deploy, databases, logs, services), prefer `mcp__jack__*` tools over CLI commands or wrangler** — they are cloud-aware and work in all deploy modes. For file editing, always use built-in Read/Edit/Write tools.",
 		);
 		sections.push(lines.join("\n"));
 

apps/cli/src/lib/mcp-config.ts

@@ -127,6 +127,45 @@ export function isAppInstalled(appId: string): boolean {
 	return existsSync(configDir);
 }
 
+/**
+ * Check if an MCP server entry looks like a remote Jack MCP connection.
+ */
+function isRemoteJackMcpEntry(entry: unknown): boolean {
+	if (!entry || typeof entry !== "object") return false;
+	const obj = entry as Record<string, unknown>;
+
+	if (typeof obj.url === "string" && obj.url.includes("getjack.org")) {
+		return true;
+	}
+
+	if (typeof obj.command === "string" && obj.command.includes("getjack.org")) {
+		return true;
+	}
+
+	return false;
+}
+
+/**
+ * Scan mcpServers for remote Jack MCP entries and remove them.
+ * Remote entries conflict with the local stdio-based MCP server.
+ */
+function removeRemoteJackMcpEntries(mcpServers: Record<string, unknown>): void {
+	const keysToRemove: string[] = [];
+
+	for (const [name, entry] of Object.entries(mcpServers)) {
+		if (name === "jack-cloud" || isRemoteJackMcpEntry(entry)) {
+			keysToRemove.push(name);
+		}
+	}
+
+	for (const key of keysToRemove) {
+		console.error(
+			`\n⚠ Removing remote Jack MCP server ("${key}") — it conflicts with the local server.\n  Local MCP covers all remote capabilities when you have filesystem access.\n  Remote MCP is only needed in terminal-less environments (claude.ai web).\n`,
+		);
+		delete mcpServers[key];
+	}
+}
+
 /**
  * Install MCP config to a single app
  * Reads existing config, merges jack server, writes back
@@ -163,6 +202,9 @@ export async function installMcpConfigToApp(appId: string): Promise<boolean> {
 	// Get or create mcpServers object
 	const mcpServers = (existingConfig[appConfig.key] as Record<string, unknown>) || {};
 
+	// Detect and remove remote Jack MCP server entries that conflict with local
+	removeRemoteJackMcpEntries(mcpServers);
+
 	// Add/update jack MCP server
 	mcpServers.jack = getJackMcpConfig();
 

apps/cli/src/mcp/resources/index.ts

@@ -119,11 +119,13 @@ This project is managed by jack.
 
 ## MCP Tools Available
 
-If connected, prefer \`mcp__jack__*\` tools over CLI:
+For cloud operations (deploy, databases, logs, services), prefer \`mcp__jack__*\` tools over CLI:
 - \`mcp__jack__deploy_project\` - Deploy changes
 - \`mcp__jack__execute_sql\` - Query databases
 - \`mcp__jack__get_project_status\` - Check status
 
+For file editing, always use built-in Read/Edit/Write tools — never remote MCP tools.
+
 ## Documentation
 
 Full docs: https://docs.getjack.org/llms-full.txt

apps/mcp-worker/CLAUDE.md

@@ -29,24 +29,36 @@ src/
     ├── deploy-template.ts # Deploy prebuilt template via control plane
     ├── endpoint-test.ts  # HTTP request to deployed endpoints
     ├── projects.ts       # list_projects, get_project_status
-    ├── source.ts         # list/read project files, update_file staging
+    ├── source.ts         # browse/read deployed files, stage_file staging
     ├── logs.ts           # get_logs
     ├── database.ts       # create_database, list_databases, execute_sql
     └── rollback.ts       # rollback_project
 ```
 
+## Local vs Remote MCP — Mutually Exclusive
+
+The local MCP (`jack mcp serve`, stdio) and remote MCP (`mcp.getjack.org`, HTTP) serve different environments and should **never be connected simultaneously**:
+
+| Environment | MCP Server | Why |
+|-------------|-----------|-----|
+| Claude Code (has terminal + local FS) | Local only (`jack`) | Agent has Read/Edit/Write for files, local MCP for cloud ops |
+| claude.ai web (no terminal) | Remote only (`jack-cloud`) | No local FS, must use stage_file + deploy for file changes |
+| Claude Desktop | One or the other | Depends on whether shell access is available |
+
+If both are connected, agents use remote file tools (`stage_file`, `read_deployed_file`) instead of local filesystem tools, causing partial deploys and failures. The `jack mcp install` command detects and removes remote MCP configs to prevent this.
+
 ## Tools (14 total)
 
 | Tool | Type | Description |
 |------|------|-------------|
 | `deploy` | write | Unified deploy: `files`, `template`, `changes`, or `staged` mode |
-| `update_file` | write | Stage a file change for later deployment via `deploy(staged=true)` |
-| `list_staged_changes` | read | List files staged via `update_file` pending deployment |
+| `stage_file` | write | Stage a file change for later deployment via `deploy(staged=true)` |
+| `list_staged_files` | read | List files staged via `stage_file` pending deployment |
 | `list_projects` | read | List all user's projects with URLs |
 | `get_project_status` | read | Deployment status, URL, resources for a project |
 | `test_endpoint` | read | HTTP request to a deployed endpoint, returns status + body |
-| `list_project_files` | read | File tree of deployed project's source |
-| `read_project_file` | read | Read single source file from deployed project |
+| `browse_deployed_source` | read | File tree of deployed project's source |
+| `read_deployed_file` | read | Read single source file from deployed project |
 | `get_logs` | read | Start log session and collect entries |
 | `create_database` | write | Create D1 database for a project (idempotent) |
 | `list_databases` | read | List D1 databases for a project |
@@ -64,8 +76,8 @@ This is the core workflow the tools enable:
    → Returns project_id + live URL
 
 2. User: "add a /forecast endpoint"
-   → LLM calls list_project_files(project_id) → sees current files
-   → LLM calls read_project_file(project_id, "src/index.ts") → gets source
+   → LLM calls browse_deployed_source(project_id) → sees current files
+   → LLM calls read_deployed_file(project_id, "src/index.ts") → gets source
    → LLM modifies the code
    → LLM calls deploy(changes: {"src/index.ts": "..."}, project_id) → redeploys
 
@@ -75,8 +87,8 @@ This is the core workflow the tools enable:
    → LLM calls test_endpoint(project_id, "/api/health") → verifies fix
 
 4. Large file update (>15KB):
-   → LLM calls update_file(project_id, "src/index.ts", content) → stages file
-   → LLM calls update_file(project_id, "src/styles.css", content) → stages another
+   → LLM calls stage_file(project_id, "src/index.ts", content) → stages file
+   → LLM calls stage_file(project_id, "src/styles.css", content) → stages another
    → LLM calls deploy(project_id, staged=true) → deploys all staged changes
 ```
 
@@ -106,7 +118,7 @@ The remote MCP (this worker) has 10 tools. The local MCP (`apps/cli/src/mcp/`) h
 | Category | Local MCP | Remote MCP | Notes |
 |----------|-----------|------------|-------|
 | Deploy/Projects | create_project, deploy_project, get_project_status, list_projects, rollback_project | deploy, list_projects, get_project_status, rollback_project | Remote merged create+deploy into `deploy` |
-| Source | — | list_project_files, read_project_file | Remote-only (new) |
+| Source | — | browse_deployed_source, read_deployed_file | Remote-only (new) |
 | Logs | start_log_session, tail_logs | get_logs | Remote merged into one |
 | Database | create_database, list_databases, execute_sql | create_database, list_databases, execute_sql | Aligned |
 | Vectorize | create/list/delete/get_vectorize_index | — | Not yet in remote |

apps/mcp-worker/src/server.ts

@@ -13,7 +13,7 @@ import type { Bindings } from "./types.ts";
 
 export function createMcpServer(token: string, env: Bindings): McpServer {
 	const server = new McpServer({
-		name: "jack",
+		name: "jack-cloud",
 		version: "1.0.0",
 	});
 
@@ -26,11 +26,11 @@ export function createMcpServer(token: string, env: Bindings): McpServer {
 - files: Full file set for a brand-new project. Pass all source files as { "path": "content" }. Only use this for the FIRST deploy of a new custom project.
 - template: Deploy a prebuilt template (hello, api, miniapp, nextjs, saas). Always creates a new project.
 - changes: Partial update to an existing project. Pass only changed/added files as { "path": "new content" } or { "path": null } to delete. Requires project_id.
-- staged: Deploy files previously staged via update_file. Set staged=true with project_id. Use this when files are too large to pass inline in a single changes call.
+- staged: Deploy files previously staged via stage_file. Set staged=true with project_id. Use this when files are too large to pass inline in a single changes call.
 
 IMPORTANT: To update an existing project, ALWAYS use changes mode with project_id. Do NOT use files mode for existing projects — it replaces all files and may create a duplicate project. If the user mentions an existing app, call list_projects first to find its project_id, then use changes.
 
-For large files (>15KB): Use update_file to stage files one at a time, then deploy(staged=true, project_id). This avoids output token limits that can truncate large inline content.`,
+For large files (>15KB): Use stage_file to stage files one at a time, then deploy(staged=true, project_id). This avoids output token limits that can truncate large inline content.`,
 		{
 			files: z
 				.record(z.string(), z.string())
@@ -52,7 +52,7 @@ For large files (>15KB): Use update_file to stage files one at a time, then depl
 				.boolean()
 				.optional()
 				.describe(
-					"Deploy files previously staged via update_file calls. Requires project_id. Use when files are too large for inline changes.",
+					"Deploy files previously staged via stage_file calls. Requires project_id. Use when files are too large for inline changes.",
 				),
 			project_id: z
 				.string()
@@ -73,8 +73,10 @@ For large files (>15KB): Use update_file to stage files one at a time, then depl
 	);
 
 	server.tool(
-		"update_file",
-		`Stage a file change for later deployment. Use this to build up changes across multiple calls, then deploy them all at once with deploy(staged=true).
+		"stage_file",
+		`Stage a file for cloud deployment via deploy(staged=true). Only use this in environments WITHOUT local filesystem access (e.g. claude.ai web, Claude Desktop without terminal).
+
+If you have built-in Read/Edit/Write tools available, do NOT use this tool — edit files locally and deploy with deploy_project or jack ship instead.
 
 Best for:
 - Large files that exceed output token limits when passed inline via changes
@@ -100,8 +102,8 @@ Pass content=null to mark a file for deletion.`,
 	);
 
 	server.tool(
-		"list_staged_changes",
-		"List files currently staged via update_file that haven't been deployed yet. Use to review pending changes before calling deploy(staged=true).",
+		"list_staged_files",
+		"List files currently staged via stage_file that haven't been deployed yet. Use to review pending changes before calling deploy(staged=true).",
 		{
 			project_id: z.string().describe("The project ID"),
 		},
@@ -188,8 +190,12 @@ Pass content=null to mark a file for deletion.`,
 	);
 
 	server.tool(
-		"list_project_files",
-		"List all source files in a deployed project. Use before read_project_file to see what files exist, or before deploying with changes to understand current project structure.",
+		"browse_deployed_source",
+		`List all source files in the DEPLOYED version of a project on Jack Cloud. Shows the file tree as it exists in production, not local files.
+
+If you have local filesystem access (e.g. Claude Code with Glob/LS tools), read the local project directory instead — it's faster and more accurate.
+
+Use before read_deployed_file to see what files exist, or before deploying with changes to understand current project structure.`,
 		{
 			project_id: z.string().describe("The project ID"),
 		},
@@ -199,11 +205,15 @@ Pass content=null to mark a file for deletion.`,
 	);
 
 	server.tool(
-		"read_project_file",
-		"Read the contents of a single source file from a deployed project. Use after list_project_files to inspect specific files before making changes with deploy(changes).",
+		"read_deployed_file",
+		`Read the contents of a single source file from the DEPLOYED version on Jack Cloud. Returns the file as it exists in production.
+
+If you have local filesystem access (e.g. Claude Code with the Read tool), read the local file instead — it's faster and always up-to-date with your working copy.
+
+Use after browse_deployed_source to inspect specific files before making changes with deploy(changes).`,
 		{
 			project_id: z.string().describe("The project ID"),
-			path: z.string().describe("File path from list_project_files (e.g. 'src/index.ts')"),
+			path: z.string().describe("File path from browse_deployed_source (e.g. 'src/index.ts')"),
 		},
 		async ({ project_id, path }) => {
 			return readProjectFile(client, project_id, path);

apps/mcp-worker/src/staging.ts

@@ -1,7 +1,7 @@
 /**
  * File staging for multi-call deploys.
  *
- * Uses KV to accumulate file changes across multiple update_file calls,
+ * Uses KV to accumulate file changes across multiple stage_file calls,
  * then deploy reads and clears the staged changes. This works around
  * LLM output token limits that prevent sending large files in a single
  * tool call.

apps/mcp-worker/src/tools/deploy-code.ts

@@ -177,7 +177,7 @@ export async function deploy(
 	const { files, template, changes, staged, project_id, project_name, compatibility_flags } =
 		params;
 
-	// Staged mode: deploy from accumulated update_file calls
+	// Staged mode: deploy from accumulated stage_file calls
 	if (staged) {
 		if (files || template || changes) {
 			return err(
@@ -202,7 +202,7 @@ export async function deploy(
 			return err(
 				"VALIDATION_ERROR",
 				"No staged changes found for this project.",
-				"Use update_file to stage file changes before deploying with staged=true.",
+				"Use stage_file to stage file changes before deploying with staged=true.",
 			);
 		}
 
@@ -252,7 +252,7 @@ export async function deploy(
 		return err(
 			"VALIDATION_ERROR",
 			"Exactly one of files, template, changes, or staged=true must be provided.",
-			"Pass files for a full deploy, template for a prebuilt app, changes for a partial update, or staged=true to deploy files from update_file calls.",
+			"Pass files for a full deploy, template for a prebuilt app, changes for a partial update, or staged=true to deploy files from stage_file calls.",
 		);
 	}
 	if (modeCount > 1) {

apps/mcp-worker/src/tools/source.ts

@@ -34,7 +34,7 @@ export async function readProjectFile(
 		return err(
 			"NOT_FOUND",
 			message,
-			"Check the file path. Use list_project_files to see available files.",
+			"Check the file path. Use browse_deployed_source to see available files.",
 		);
 	}
 }
@@ -93,7 +93,7 @@ export async function listStagedChanges(
 		return ok({
 			staged_files: 0,
 			files: [],
-			note: "No staged changes. Use update_file to stage file changes before deploying.",
+			note: "No staged changes. Use stage_file to stage file changes before deploying.",
 		});
 	}