Conway
Anthropic’s unannounced internal always-on agent project, exposed in the early-2026 Claude Code source leak (the same packaging-error publication that produced the 12 Agent Primitives analysis). Not on Anthropic’s roadmap page. Discovered and analyzed by Nate B Jones in his April 2026 deep-dive on the leaked source.
Speculative / leaked source. Everything on this page is from the leak — not Anthropic’s published roadmap, not a launch, not a confirmed product name. The factual structure (sidebar UI, three sections,
.cnw.zipformat, automatic triggers) is treated as primary-source-from-leak at the same credibility tier as the 12 Agent Primitives. The trajectory framing is editorial. This page exists so the wiki has a place to file the next data point when Conway is announced or appears in another source.
What it is, from the leak
Conway is described as an entire agentic environment, not a chat window — a standalone sidebar inside the Claude interface that opens a dedicated page tied to a Conway instance. It has three core areas:
| Area | What it contains |
|---|---|
| Search | Cross-source search across the agent’s accumulated context |
| Chat | Conversational interface to the running Conway instance |
| System | The interesting one — see below |
The System section has three subsections:
- Extensions — install add-ons that extend Conway’s capabilities. Custom tools, interface panels, “ways for the agent to understand new kinds of information.” Packaged as
.cnw.zipfiles. App-store pattern. - Connectors and Tools — shows what other services are plugged in. Includes a toggle that lets Claude and Chrome connect directly into the Conway instance (browser control as a first-class connector).
- Automatic Triggers — public web addresses that outside services can ping to wake the agent up. Per-service allow toggles. This is the cron / webhook primitive Managed Agents is missing at launch.
Persistent memory across sessions is implied throughout the leaked structure — Conway “knows your organization in a way that nothing else does” because it accumulates context over time.
The .cnw.zip extension format — the structural detail
Conway extensions ship as .cnw.zip packages and contain custom interface panels, information handlers, and tools. Critically: they sit on top of MCP but are not portable MCP tools. They only work inside Conway.
The pattern Nate identifies: MCP is the open foundation; Conway’s extension ecosystem is the proprietary layer on top. Anthropic gets the credibility of publishing an open standard (MCP) and the commercial advantage of building the valuable tooling in a format that runs in their environment.
This is the Google Play Services / iPhone App Store pattern applied to AI agents:
| Open layer | Proprietary layer |
|---|---|
| Linux / AOSP | Google Play Services (Maps, Payments, Push, Play Store) |
| Web standards | iOS App Store |
| MCP | .cnw.zip extensions inside Conway |
Developers face the same choice mobile devs faced in 2008–9: build for the open standard (portable, no built-in distribution) or build for the platform store (locked-in, but featured placement and millions of pre-installed users).
Conway alongside Managed Agents
Anthropic now has two hosted-agent surfaces in the wiki, one announced and one leaked:
| Managed Agents (announced 2026-04) | Conway (leaked, unannounced) | |
|---|---|---|
| Status | Generally available | Internal; not on roadmap |
| UI | Web console at console.anthropic.com | Sidebar inside Claude interface |
| Extension format | MCP servers + skills | .cnw.zip (proprietary) |
| Triggers | None at launch — API call only | Automatic public web triggers, per-service toggles |
| Browser control | No | Yes (Chrome connector) |
| Memory across sessions | Stateless (private preview) | Persistent (implied) |
| Positioning | Onramp / “build your first agent” | Always-on / “knows your organization” |
The structural read: Managed Agents is the announced version of Anthropic’s hosted-agent bet, deliberately scoped down to a Claude.ai-user onramp. Conway is the unannounced version aimed at the always-on layer — the thing that locks in behavioral context. They are not redundant; they are two stages of the same platform play.
The 90-day Anthropic platform play (Nate’s framing)
Conway is the capstone, not a standalone product. Nate lines up Anthropic’s last quarter as a coordinated arc:
- Claude Code channels — message Claude Code from Discord/Telegram + task-completion notifications. Neutralizes OpenClaw inside Anthropic’s own surface.
- Claude Co-work — non-technical surface for the 95% of enterprise employees who aren’t engineers. Adoption reportedly outpaced Claude Code at the same stage.
- Claude Marketplace — enterprise procurement layer (GitLab, Harvey, Snowflake buyable through Anthropic; counts against existing spend commitments; no commission yet).
- $100M Claude partner network + Accenture training 30,000 professionals; Deloitte/Cognizant/Infosys as anchor SIs.
- Third-party tool ban — January quiet block, February ToS revision, recent enforcement against OpenClaw first; pay-per-use rates 10–50× subscription cover.
- Managed Agents (April) — first hosted-agent runtime. Onramp tier.
- Conway (leaked, unannounced) — always-on capstone; the Active Directory equivalent.
“Anthropic is speedrunning Microsoft. Model provider to developer tool to enterprise platform to agent operating system in 15 months. Microsoft took 15 years.”
The four-step playbook (Nate’s framing)
The OpenClaw / Steinberger / ban timeline exposes a repeatable pattern for how Anthropic absorbs community-built primitives:
- Copy — build the first-party version of what the community built (OpenClaw → Claude Code channels → Conway)
- Subsidize — make the first-party version free or cheap inside the subscription
- Block — make the third-party version expensive (10–50× pay-per-use) or technically impossible
- Lock the format — ship a proprietary extension format (
.cnw.zip) that ensures the ecosystem builds for the platform’s surface, not the open one
We’re at step 1 on Conway publicly. Steps 2–4 are visible in the rest of the leak.
What to watch
This page exists to be updated as Conway moves from leak → roadmap → preview → launch. Watch for:
- Conway naming on a real Anthropic surface — first appearance in docs, a roadmap page, or a launch post. Confirms the leak.
.cnw.zipbecoming a real, documented format — if Anthropic publishes an extension SDK, this is the “build for the App Store” inflection point for tool developers.- Automatic triggers shipping — would close the gap Managed Agents left and make trigger.dev less load-bearing for Anthropic-native always-on workflows.
- Browser control as a first-class Anthropic feature — would compete directly with Stagehand / OpenClaw / Computer Use.
- OpenAI’s and Google’s equivalent moves — Nate predicts both ship Conway-class always-on agents with similar proprietary extension layers in the next few months. If/when this happens, the Era 3 persistence wars become the wiki’s central thread.
- Any portability commitment — if Anthropic publishes a behavioral-context export format or a portability commitment before Conway launches, that’s the rare counterexample to the four-step playbook.
See Also
- anthropic — vendor; the 90-day platform play arc
- managed-agents — announced hosted-agent sibling
- behavioral-lock-in — the conceptual frame Conway anchors
- OpenBrain — the user-owned memory pattern that’s the structural counter
- mcp — the open standard Conway sits on top of with a proprietary layer
- claude-code — the leak source
- agentic-harness-primitives — Nate’s earlier analysis of the same leak
- openclaw — the playbook target
- loop — the local proactivity primitive Conway is the hosted version of
- Source: Nate B Jones — I Analyzed 512,000 Lines of Leaked Code