Perforce Agentic Gateway

Add an MCP server here — never edit MCP config files by hand; PAG writes the config and starts the server. For a stdio server give alias, command and args, e.g. {"alias": "time", "command": "uvx", "args": ["mcp-server-time"]} for a PyPI package or {"alias": "time", "command": "npx", "args": ["-y", "<npm-package>"]} for npm; url for a remote HTTP server; image for a container. Use this for servers not in a registry — for registry servers use pag__enable_server.

Update inputs, environment variables, arguments, tool filtering, or the pinned version for a server. Deep-merge semantics on the nested inputs tree: omitted keys are left unchanged, null values delete keys. To apply an available update, pass version with the new value. Changes that affect execution trigger a server restart.

Disable a server for this project. The server stops but its configuration is retained.

Scan the project directory for technology signals and get recommendations for MCP servers to enable. Call when the user asks what servers would help this project, or when a task needs a capability no enabled server provides.

Enable an MCP server from the registry for this project. Before calling, use pag__search_catalog to learn the server's inputs. Collect non-secret input values (and per-variable overrides for templated inputs) from the user and pass them under the inputs parameter. Writes to project config and starts the server. Blocks until the server reaches a stable state (running or failed). If secrets are missing, returns pag_missing_secrets with a browser URL for secret entry.

Execute a JavaScript program (the code parameter) that calls downstream MCP tools and returns structured results. Use pag__explore first to find tool names and schemas — calls with guessed arguments fail. Compose calls in one complete program per goal. Example — call a tool, return its result: const r = await callTool("klocwork__get_issues", { project: "myapp" }); const text = r[0].text; return r.structuredContent ?? (text.startsWith("{") || text.startsWith("[") ? JSON.parse(text) : text); callTool(qualifiedName, args) is async and returns content items — most tools' result is r[0].text; when the tool declares an outputSchema, the pre-parsed typed object is attached as r.structuredContent. Prefer it, and return it explicitly: returning the array directly does not carry the property (JSON.stringify ignores array properties). Not every tool replies in JSON. Some — PAG's own ui_builder and management tools among them — reply in prose meant to be shown to the user, and r[0].text is then the whole result: read it and return it as the string it is. JSON.parse throws on prose, and an uncaught exception loses the whole program including the tool call that produced it; parse only when the text starts with '{' or '['. On failure callTool throws an Error whose .message is prefixed with the qualified tool (e.g. "klocwork__get_issues: <reason>"); .server and .tool carry the alias and qualified name. Use try/catch. console.log() writes debug output, returned under "logs". Return a value to send results back (JSON-serialized). Async/await is supported. Constraints: plain JavaScript only. No TypeScript, no imports. Not available in this sandbox: require, setTimeout, setInterval, eval, globalThis, process, fetch, URL, TextEncoder, btoa, atob, crypto, performance, console.error, console.warn.

Query the catalog map (alias to tools) via the code parameter (JavaScript); read-only — no callTool here. Tool definitions stay server-side until your code returns part of them. Find tools and their schemas before calling them through pag__execute. Example — find tools by keyword, return what you need: const tools = Object.values(catalog).flatMap(s => s.tools); const hits = fuzzy("issues", tools, "description"); return hits.map(h => ({ qualifiedName: h.item.qualifiedName, description: h.item.description, inputSchema: h.item.inputSchema })); Catalog shape (TypeScript notation) — catalog is a map of server alias to { tools: [...] }; each tool has const catalog: Record<string, { tools: Array<{ name: string; qualifiedName: string; description: string; inputSchema: object; outputSchema?: object; }> }> name is unprefixed; qualifiedName is the "alias__toolname" form callTool takes in pag__execute. outputSchema, when present, describes what callTool returns: the structuredContent object and the JSON in content[0].text. Absent means undeclared. Also available: function fuzzy(query: string, items: any[], key?: string): Array<{item: any; score: number}> approximate matching, sorted by score; key picks the property to match when items are objects. function console.log(...args: unknown[]): void — debug output, returned under "logs". Return a value to send results back, and return only the entries you need — the whole catalog can be large. Not available in this sandbox: require, setTimeout, setInterval, eval, globalThis, process, fetch, URL, TextEncoder, btoa, atob, crypto, performance, console.error, console.warn, callTool.

Report PAG's own token savings — how many tokens code mode kept out of this client's context. Use it when the user asks about PAG's savings or token counts. The figures are PAG's gateway accounting of pag__execute processing and progressive disclosure: execute savings break down with group_by, disclosure savings are aggregate-only. Returns percent reduction and absolute token counts; pct is signed, and negative means code mode added more context than it kept out. status no_activity means no code-mode activity in the range yet. An execute figure's basis is measured, or apportioned when its net was split proportionally across the entities a call touched.

Restart a server, regardless of its current state. Use when the user requests a restart or when the server is in a stopped/error state.

Search configured MCP registries for servers. Omit query to list all available servers. Returns server metadata, required inputs, and whether each server is already enabled.

Retrieve recent stderr output from a server process. Use to diagnose why a server failed to start or is misbehaving.

Show server health status. With no arguments, returns a summary of all servers. With a server name, returns detailed status including configuration and recent history.