Cursor vs Zed (2026)

Breadth across classes is met, and the ACP host role is the distinguishing property: Zed does not only integrate tools, it hosts other vendors' agents as first-class citizens in its own panel, which is a different kind of integration surface from a connector catalog.

The spawn_agent tool documents a coordinated handoff: a subagent takes a delegated task with its own context and the parent reviews what it returns.

Three native platforms plus Terminal Threads, remote development and multiplayer sessions is real surface breadth. What holds it below Full: every path is developer-initiated inside the editor. No event trigger such as a pull request or commit, no schedule, and no inbound chat or ticketing channel is documented. Zed is an editor, so this is a property of the form rather than a gap.

Context is assembled per run from a local Instant Grep index, Explore subagents and @ mentions, and Cursor states it stores no codebase embeddings.

Mention-based context plus agent-initiated codebase search is real grounding, and multi-core indexing makes it fast, but grounding is assembled per request rather than held as a structural representation. No repository index product, dependency graph or cross-file reasoning surface is documented, and a third-party FAQ observes that project-wide context relies on what the developer explicitly passes to the panel. An external agent connected over ACP may bring its own indexing, which belongs to that agent rather than to Zed.

Past agent transcripts can be searched and attached with @Chats, but no memory layer with a stated scope and lifetime is documented.

Checkpoints are a genuine state mechanism and restorable thread history is durable across sessions, which is more than most editors offer, but nothing accumulates learned context: Skills and Instructions are configuration the customer writes, not memory the agent forms. One limit worth knowing: the docs state that checkpoint, history restore and token display availability varies by external agent, so the capability is not uniform across the ACP roster.

Run Modes set when the agent stops for approval on shell, MCP and fetch calls, and a team configuration in the dashboard overrides user and project settings.

Tool permissions include a documented confirm step: a global or per-tool confirm default, always_confirm patterns that prompt case by case, and auto-deny rules that hold even when the default is allow. After a run, changes are accepted or rejected hunk by hunk, and checkpoints revert any edit batch.

SOC 2 is in progress on a first-party docs page, and SSO, SAML and SCIM are stated as planned. Zed tells a buyer plainly what it has not built yet, and the shipped Business admin layer covers the control side.

Hooks fire on every step of the agent loop and can feed the buyer's own audit log, while the Admin API audit log covers logins and settings changes rather than agent actions.

Following the agent live over the same multiplayer infrastructure that powers human collaboration is a distinctive visibility mechanism, and the public agent metrics page is unusual transparency. What holds it below Full: these show what is happening now and what it cost rather than retaining a per-action trace of what the agent did and why. Thread history preserves the conversation, not an execution record, and its availability varies by external agent.

Self-Hosted Machines run Cloud Agent tool execution on customer-managed machines while Cursor keeps the agent loop, inference and planning, and Enterprise can enroll in US only residency.

Residency is structural rather than offered: the editor is Apache 2.0 and runs entirely on the developer's machine, so there is no vendor-side service holding code at all unless the customer opts into hosted models. Zed's own phrasing on local models, that nothing leaves your machine, is unusually direct. Remote development inverts it usefully, letting the codebase stay on a server the customer controls while only the interface runs locally.

Reviewed plugins bundling rules, skills, agents, MCP servers and hooks install from the Cursor Marketplace, with private team marketplaces on Teams and Enterprise.

The ACP Registry is a one-step install catalog of ready external agents, which is the adoptable set this axis asks for.

Three different modes are documented with their own pages, hosted, local and bring-your-own-key, and Zed is explicit that local models mean nothing leaves the machine and that BYOK uses the customer's existing accounts and budgets. Named models are current, and community pull requests adding providers land continuously, which is a structural advantage of the open-source model.

Zed authored the protocol other vendors adopt: the Agent Client Protocol is Zed's own open standard for connecting any agent to any client. Publishing the interface an ecosystem integrates through is the strongest form of this axis, a larger claim than shipping an API for one's own product, and it is the point to weigh against tools whose extensibility rests on a REST endpoint.

Bugbot and a verifier subagent test the customer's code, but no harness evaluates the agent's own behavior.

Native debugging through the Debug Adapter Protocol is a real first-class capability, and diagnostics feed the agent as context it can act on. What holds it below Full: the agent running tests and the developer debugging verify the customer's code, not agent behavior, and no evals, benchmarking or regression surface for agents is exposed. Zed's editor performance benchmarks and public agent usage metrics are Zed's own measurement rather than a harness the customer can use.

The agent drives a browser pane inside Cursor to navigate, click, type, scroll and take screenshots, with approval required by default and an admin origin allowlist.

A terminal, a debug adapter and a tool call are all programmatic interfaces, so Terminal Threads running command line agents, native debugging and agent tool use in the workspace are not computer use. zed.dev/ai, the agent panel docs and the MCP docs show no browser tool, screenshot capability or GUI automation. Because Zed hosts external agents through ACP, an agent such as Claude Agent may bring its own browser tooling into the panel, but that capability belongs to the guest agent rather than to Zed.