mcpwith.codemode.cc github

codemode tool (QuickJS Sandbox)

Harness-side QuickJS VM (256 MB cap, top-level await, no filesystem/network/timers) that exposes MCP & built-in tools as JS functions.

How it is enabled & configured

• Auto-enabled for MCP: Connects automatically when any MCP server in ~/.pi/agent/mcp.json or .pi/mcp.json uses the default "exposure": "codemode", or always-on via "defaultTools": ["+codemode"] in settings.json.
• Exposure control: Per-server exposure and per-tool toolExposure let you route tools as codemode (script-only, ~40% fewer prompt tokens), deferred (loaded via search), direct, or hidden.

How it is used

• Tool orchestration: Scripts call await tools.mcp__<server>__<tool>(args) or built-ins like tools.read() / tools.bash(), discovering unlisted schemas on demand with searchTools() (BM25), describeTool(), and ALL_TOOLS.
• State & auxiliary models: Persist small JSON state across turns via store(k, v) / load(k), emit curated output via return, text(), or image(), and invoke non-LLM models with models.classify() and models.generateImages().

// @options: {"max_output_tokens": 2000, "timeout_ms": 60000}
const hits = await searchTools("phoenix spans");
const [proj, spans] = await Promise.all([
  tools.mcp__phoenix__list_projects({}),
  tools.mcp__phoenix__get_spans({ project: "default", limit: 200 }),
]);
const slow = spans.filter(s => s.latency_ms > 1500);
store("slow_span_ids", slow.map(s => s.id));
return { projects: proj.length, slowCount: slow.length };

execute tool (Confined Code Mode Runtime)

Confined JS runtime for scripting catalog tools and fetch requests while enforcing per-tool permissions on every nested call.

How it is enabled & configured

• Default for MCP servers: Every server under mcp.servers in opencode.jsonc defaults to "codemode": true (set "codemode": false to keep a server on the provider's native tool list).
• Granular permissions: Permission execute (*) gates Code Mode itself, while every nested call inside the script still enforces its own permission rule (e.g. <server>_<tool>).

How it is used

• Namespaced catalog & search: The agent calls execute({ code }) to run JS against tools in the catalog (or discovered via search()) using tools.<server>.<tool>(input) (or tools.<ns>["tool-name"](input)) plus native fetch.
• Built-in namespaces & concurrency: Beyond MCP servers, scripts orchestrate tools.opencode.* (sessions, worktrees, model lookup, MCP resources) and tools.browser.* (desktop browser automation), running independent calls concurrently with Promise.allSettled().

const [resA, resB] = await Promise.allSettled([
  tools.phoenix.list_projects({}),
  tools.opencode.list_mcp_resources({ server: "phoenix" }),
]);
const projects = resA.status === "fulfilled" ? resA.value : [];
return {
  activeProjects: projects.map(p => p.name),
  resources: resB.status === "fulfilled" ? resB.value.resources.length : 0,
};

functions.exec (codex-code-mode-host)

Dedicated local or remote WebSocket execution host enabled by default in Codex CLI (v0.146+).

• Configured via ~/.codex/config.toml: Enabled by default under [features] with code_mode = true, code_mode_host = true, and optional code_mode_only = true (or remotely via --code-mode-host wss://...).
• How it is used: The agent writes JS in functions.exec to batch nested built-in tools (exec_command, apply_patch, wait) and MCP tools with Promise.allSettled() while preserving host approval policies.

code_execution extension (pctx / Deno)

Built-in MCP facade powered by Port of Context (pctx), a sandboxed Deno TypeScript runtime.

• Configured via goose configure: Toggle the built-in code_execution extension (or pass --with-builtin code_execution,developer).
• How it is used: Hides raw MCP schemas behind 3 meta-tools — list_functions, get_function_details, and execute_typescript — so the agent discovers schemas on demand and chains calls in TypeScript.

MCPSearch + Programmatic Tool Calling (PTC)

Two-tier context reduction across the Claude Code CLI and the Claude Developer Platform.

• Claude Code CLI (MCPSearch): Activates automatically when MCP schemas exceed 5% of context (ENABLE_TOOL_SEARCH=auto:5), deferring schema loading (defer_loading: true) for ~85% fewer prompt tokens.
• Claude API / SDK (PTC): Tools marked with allowed_callers: ["code_execution_20260120"] become callable Python functions inside a sandboxed container, letting Claude loop and filter 20+ calls in Python before returning final stdout.

Minimalist Python Harness + Extension API

Hugging Face’s readable Python port of Pi’s architecture (tau_ai → tau_agent → tau_coding).

• Intentionally minimal core: Ships only 4 built-in tools (read, write, edit, bash — Pattern 01) without a built-in MCP client or JS VM.
• Python-native Code Mode via Extensions: Because Tau runs on Python 3.12+, you can register a custom AgentTool via tau.register_tool(...) in ~/.tau/extensions/ that embeds Pydantic Monty (or calls a FastMCP Code Mode server) directly in-process.