Cloudflare Code Mode (@cloudflare/codemode)

Cloudflare Code Mode (@cloudflare/codemode)

One-liner

Code Mode is an agent tool-use pattern (and Cloudflare SDK) where the model gets one code-execution tool, writes JS/TS against typed connector APIs, and runs that plan in an isolated sandbox—instead of round-tripping every MCP/tool call through the model.

Related: 2026-09-26 Cloud agent orchestrators in the wild · 2026-09-26 Herdr for v0 cloud orchestrator · Cloud agent orchestrator

---

Confidence / disambiguation

Primary (this note): Cloudflare Code Mode / npm @cloudflare/codemode — high confidence. Coined and productized by Cloudflare (Kenton Varda, Sunil Pai / @threepointone); documented in Agents SDK docs; independent parallel framing from Anthropic as “code execution with MCP.”

Other meanings (brief):

Name	What it is	Relevance to Noa
Anthropic code execution with MCP	Same core idea (tools as code APIs + sandbox); filesystem progressive disclosure	Sister pattern; cite alongside CF
Community “Code Mode MCP” meta-servers	DIY wrappers that collapse many MCPs into search + execute_code	Useful DIY precedent
VS Code / Copilot Agent Mode	IDE agent harness with MCP—not “Code Mode”	Naming collision only
OpenCode / other harness “codemode” support	Harnesses adopting the pattern (MCP resources as Code Mode tools, etc.)	Watch for client-side adoption
Unrelated “code mode” (editors, games, accessibility)	Generic UI labels	Ignore

---

What it is / who makes it

Code Mode = pattern: LLMs are better at writing code than at direct tool calling, because training data has vast real TypeScript and little synthetic tool-call markup. So: convert tools/MCP/OpenAPI into a typed SDK, give the model a single codemode({ code }) (or search + execute) surface, execute in a sandbox, return only what the code logs/returns.

Cloudflare’s implementation:

• Package: @cloudflare/codemode inside cloudflare/agents (MIT; marked experimental).
• Authors/public face: Kenton Varda (Workers / isolates / Dynamic Worker Loader), Sunil Pai (@threepointone) — Agents SDK implementation.
• First public framing: blog “Code Mode: the better way to use MCP” (2025-09-26). Follow-on: “Code Mode: give agents an entire API in 1,000 tokens” (2026-02-20) — full Cloudflare API MCP via server-side Code Mode.
• Anthropic independently described the same pattern (2025-11-04) and notes Cloudflare’s “Code Mode” naming.

---

How it works

Model-facing shape

1. Tools are not listed as dozens of separate function calls.
2. The model sees roughly one outer tool (codemode) whose input is { code: string }, or (large APIs) two tools: search + execute.
3. Inside the sandbox, connectors appear as globals (github.*, stripe.*, …) plus a platform SDK:
   • codemode.search(query) → ranked paths (methods/snippets), not full schemas
   • codemode.describe(path) → focused TypeScript docs
   • codemode.step / codemode.run → durable steps / saved snippets
4. Generated code composes calls, loops, filters, branches; only the final result (plus logs) returns to the model.

Runtime pieces (@cloudflare/codemode)

Piece	Role
Executor	Runs one code block once; no durable state. Default: DynamicWorkerExecutor (Worker Loader → fresh V8 isolate). Also: browser IframeSandboxExecutor.
Connectors	Host-side capabilities exposed into the sandbox: McpConnector, OpenApiConnector, AI SDK ToolSetConnector, custom CodemodeConnector. Credentials stay on the host; sandbox gets RPC bindings.
Durable runtime	createCodemodeRuntime() — SQLite-backed Durable Object facet: execution log, approvals, replay, rollback, snippets.

Isolation: fetch/connect blocked by default (globalOutbound: null). External I/O only via connectors. No API keys in sandbox.

Approvals: Methods marked requiresApproval pause the pass; approve/reject resume via replay of the same script (applied calls return recorded results). Optional revert enables compensation/rollback.

AI SDK path (stateless):

const executor = new DynamicWorkerExecutor({ loader: env.LOADER });
const codemode = createCodeTool({ tools: myTools, executor });
streamText({ model, messages, tools: { codemode } });

MCP server patterns:

• codeMcpServer() — wrap an existing MCP → one code tool
• openApiMcpServer() — large OpenAPI → fixed search + execute (~1k tokens for whole Cloudflare API in CF’s own MCP)

---

Why it matters / use cases

Problem it attacks: MCP tool catalogs and intermediate results eat the context window. CF claims ~99.9% input-token reduction vs naive per-endpoint MCP for the Cloudflare API (~1.17M → ~1k tokens); earlier Code Mode work cited ~80%+ savings and better multi-step results. Anthropic reports similar (~98.7% in a filesystem progressive-disclosure example).

When to use Code Mode vs direct tools (CF docs):

Prefer direct tools	Prefer Code Mode
Small, fixed tool set; simple one-shot calls	Composition, dependent calls, loops/branching
	Large catalogs / progressive discovery
	Filter/transform before results hit the model
	Reusable snippets / durable approvals

Use cases aligned with Noa (agents, Cursor, herdr, DIY):

• DIY orchestrators / personal MCP hubs — collapse Slack+GitHub+browser+internal APIs into search/execute without stuffing every schema into Cursor/Claude/Codex context (see Cloud agent orchestrator, herdr fleet tooling).
• Server-side Code Mode — agent clients need no special sandbox; MCP server runs JS in isolates (CF’s mcp.cloudflare.com pattern). Good when harnesses lag on client-side Code Mode.
• Harness adoption — X discussion (OpenCode, dax/@thdxr, others) frames Code Mode as something clients/harnesses should own so every MCP doesn’t ship its own sandbox.
• Security model — keys never enter generated code; bindings + approvals + rollback fit “agent may write side-effecting scripts” better than raw shell.

Not a cloud orchestrator

Code Mode is a tool-composition / context pattern inside an agent loop. It does not replace Herdr panes, Cursor Projects, or overnight PR factories—but it is highly complementary when those stacks grow large MCP surfaces. See 2026-09-26 Cloud agent orchestrators in the wild.

---

Alternatives / related concepts

Approach	Idea	Tradeoff
Direct MCP tool calling	One schema per tool in context	Simple; collapses at scale
Anthropic code execution with MCP	Tools as FS modules / code APIs	Same insight; different packaging
Dynamic tool search (e.g. Claude Code)	Retrieve relevant tools per turn	Still pays per matched schema
CLI / shell progressive disclosure	Agent runs --help, scripts	Broader attack surface; needs a shell
Programmatic tool calling (various SDKs)	Client-side code mode variants	Needs a sandbox in the agent host
MCP server portals (Cloudflare One)	Gate approved MCPs; Code Mode policies	Enterprise/gateway layer on top of the pattern

---

Sources

Docs & blogs

• Code Mode · Cloudflare Agents docs
• How Code Mode works
• AI SDK integration
• Durable Code Mode runtime
• Code Mode API reference
• Code Mode MCP server patterns
• Dynamic Workers — Code Mode example
• Code Mode: the better way to use MCP (2025-09-26, Kenton Varda & Sunil Pai)
• Code Mode: give agents an entire API in 1,000 tokens (2026-02-20, Matt Carey)
• Sandboxing AI agents, 100x faster (Dynamic Workers)
• Anthropic — Code execution with MCP (2025-11-04)

GitHub

• cloudflare/agents — packages/codemode
• Example: examples/codemode

X / Twitter (public posts via X API; not login-walled for these URLs)

Primary authors

• Kenton Varda — early teaser (2025-09-04 CST+8: ~23:18) — “everyone using MCP wrong” + unannounced Workers feature; later revealed as Code Mode
• Kenton Varda — Cloudflare OS thread (2026-08-05) — agents act via Code Mode snippets + Cap’n Web RPC / Gatekeepers (high-engagement primary thread)
• Sunil Pai (@threepointone) — “code mode!” (2026-09-17)
• Sunil Pai — “throw codemode at this” (2026-09-17)

Recent discussion / adoption

• dax (@thdxr) — “now that we have codemode you can add as many MCPs as you want without cost” (2026-09-11) — OpenCode / harness angle
• Nathan Flurry — CLI vs MCP; Code Mode / scripting tipped the scale (2026-09-10)
• Marwan Atef — Code Mode on MCP server portals: portal_codemode_search + portal_codemode_execute (2026-09-08)
• chebyte — CF API MCP as search+execute ~1k tokens (2026-09-06)
• OpenCode Changelog — MCP resources as Code Mode tools in v2.0.16 (2026-09-24)
• Yifan Xu — codemode + script as coding-agent endgame (2026-09-27)

X access note

Post URLs above resolved via Noa’s connected user-X search/get APIs (public post objects + url fields). Full Twitter web UI can still soft-gate browsing when logged out; these status links are the citable public permalinks.

---

Takeaways for Noa

1. Treat Code Mode as a design default for any DIY MCP hub or agent coordinator that might grow past ~dozens of tools—especially before stuffing more servers into Cursor/Claude/Codex.
2. Prefer server-side Code Mode (two tools, host sandbox) when the coding harness doesn’t own a good isolate yet; push client-side Code Mode when building on Cloudflare Agents / Dynamic Workers.
3. Complementary to Cloud agent orchestrator / Herdr: Herdr owns sessions & panes; Code Mode owns efficient tool composition inside a session.
4. Mark experimental: CF docs warn of breaking changes; durable approvals/replay are powerful but add ops complexity (replay divergence, connector lifetime hooks).