Zed's AI Features in 2026: Agents, ACP and Pricing

Zed, the open-source code editor written in Rust, has spent 2026 turning itself into a host for other people's coding agents as much as a vehicle for its own. As of release 1.26.2 on October 28, 2026, its documentation describes three distinct ways to run an agent, each with a different answer to who bills you, who sees your code and what the editor's sandbox protects. Knowing which path you are on matters more than any single feature.

The three paths are the native Zed Agent, which Zed hosts, bills and sandboxes; External Agents, which connect over the Agent Client Protocol that Zed introduced and now shares with JetBrains, and which run under the agent vendor's own billing and data terms; and Terminal Threads, which run any agent command-line tool inside the Agent Panel. Edit prediction is a fourth, separate system with its own model and its own provider list.

This post dates each feature from Zed's own pages, says which ones Zed labels beta and which it does not, and gives the pricing as the site shows it. It does not report hands-on results.

Zed's AI overview page is organised around the split, and the split is the most useful thing to understand before touching a setting. The native path gives you Zed's own agent with "built-in tools, profiles, skills, instructions, and MCP server support". The external path hosts another vendor's agent, which the docs say "run[s] through their own process and configuration". The terminal path runs a CLI or text-mode agent directly inside Zed. The table sets the three beside each other on the questions that differ.

The why behind the structure is in the external-agents page: Zed hosts the thread in its panel and sidebar while the agent "manages its own runtime, authentication, model selection, and tools", and "Billing is handled by the agent provider, and Zed does not charge for External Agents." That sentence is the business model in one line. On our reading, Zed earns on its own hosted models and on seats, and gives away the host so that a developer with a Claude Code or Codex subscription has no reason to leave the editor.

The Agent Panel is the surface for all three paths, but the native agent gets the most from it. Zed's Agent Panel documentation describes a panel that "lets you prompt agents, add context, review changes, and manage active threads", with a Threads Sidebar that groups threads by project, per-hunk accept and reject of the agent's edits, profiles that decide which built-in and MCP tools are available, per-tool permissions that can be allowed, denied or confirmed, and a follow mode that moves the editor to whatever file the agent is reading or editing.

The release cadence is the example of why dating matters: six stable releases between October 12 and October 28 alone, several of them changing how external agents behave. The failure case for a team adopting Zed is to standardise on a feature from a blog post without checking the release notes, which on this product move weekly.

The Agent Client Protocol is Zed's answer to the question of how an editor hosts an agent it did not write. The External Agents page lists Gemini CLI, Claude Agent, which is the docs' name for Anthropic's agent installed from the registry, and Codex, plus OpenCode, Copilot, Cursor, Pi Coding Agent and Poolside, with the note that the list "is curated, not exhaustive". Custom agents are configured with a command, arguments and environment. The docs call the ACP Registry "the primary way to install common External Agents in Zed".

That registry went live on January 28, 2026, in a post by Ben Brandt of Zed describing a place where developers register an agent once and it works in every ACP-compatible client. The post credits JetBrains with pushing the registry specification forward, and both Zed and JetBrains IDEs ship built-in support; JetBrains joined ACP on October 6, 2025. The agents the launch post names are Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode and Gemini CLI.

Three boundaries from the docs decide whether an external agent feels native. Zed Agent profiles and Zed Skills "generally don't apply" to external agents. Zed's configured MCP servers "may be forwarded over ACP". And extension-provided agents are deprecated in favour of registry equivalents. Recent release notes show the seam being worked: 1.25.2 on October 21 added session notices and context compaction for ACP agents, 1.24.4 on October 20 fixed OpenCode models failing to load, and 1.26.2 on October 28 made ACP config defaults apply only to new conversations. The page does not say whether ACP or External Agents are beta, so we do not call them that.

Edit prediction is Zed's name for the inline next-edit suggestion, and its default provider is Zeta, which the edit prediction docs describe as "an open source model developed by Zed". Zeta predictions are included with Pro, the Pro trial, and the Student and Business plans; Free-plan users need another provider. The alternatives the docs list are GitHub Copilot, Mercury Coder from Inception Labs, Codestral from Mistral, and local models through Ollama or any server that implements the OpenAI completions endpoint, with vLLM, the llama.cpp server and LocalAI named.

By Zed's blog titles, the model was rebuilt this year and then revised: "We Rebuilt Zeta from the Training Data Up" on March 25, 2026 and "Zeta2.1: 3x Fewer Tokens, 50ms Faster" on May 8, 2026. We read only the titles of those posts, so the numbers in the second are Zed's headline claim and nothing more. Provider choice arrived with a February 4, 2026 post. Display has two modes: eager, the default, and subtle, where predictions "only appear inline when holding a modifier key". The page does not mention beta.

The privacy consequence is specific. Zed's privacy page says that with edit prediction, "Each keystroke can send local editing context to the selected provider." Zeta processes that transiently unless training-data collection is on, and training data is collected only if the user opts in, the project is open source and the file is not excluded. A team that enables a third-party provider has changed where every keystroke goes, which is why the Business plan lets admins turn edit predictions off organisation-wide.

Zed's August 5, 2026 sandboxing post is unusually direct about scope, and the scope is narrower than the word suggests. Sandboxing is "enabled by default for all users, starting on the 1.14 release", and it applies to the native agent's terminal and fetch tools. Default rules "block writes outside project directories, writes to .git, and network requests", enforced by the operating system, Seatbelt on macOS, Bubblewrap on Linux and WSL only on Windows, because, as the post puts it, you should "not rely on an agent following instructions".

The example that follows: a team that chose Zed for the sandbox and then installed Claude Agent from the registry has the sandbox for none of its agent work. Whether that is acceptable depends on the external agent's own controls, which our comparison of Claude Code, Cursor and Codex covers; the point here is only that Zed's docs say the boundary is there.

Zed moved to token-based billing for hosted models on September 24, 2025 and launched a Business plan on May 6, 2026. The pricing page on November 5, 2026 lists three plans and a trial.

Two things the table implies. A developer who already pays Anthropic, OpenAI or GitHub for an agent subscription can run it in Zed at the Personal tier for nothing, which is where the external-agents strategy pays off for the user rather than for Zed. And a Pro subscription's $5 of tokens is small against hosted frontier models, so the plan is better read as $10 for edit predictions plus a metered account than as an AI allowance. Our coding-agent cost calculator puts these numbers beside the other tools.

Zed's AI privacy page opens with the statement that "Zed does not retain your prompts or code context by default", keeping data only when a user explicitly shares or opts in. For hosted models it goes further: provider commitments prohibit training on prompts and code, and "Provider agreements require zero data retention for Zed-hosted inference requests, with one exception". The exception is Anthropic, which retains prompts and outputs for the models it designates Covered Models, the page naming Claude Fable 5.1 and Claude Fable 5, for at least 30 days for trust and safety purposes. The page states that Zed cannot opt out and that bringing your own key does not avoid it for those models, and that the retained data is used for safety review, not training.

For the other two paths the rule is simpler and less reassuring: the external agent or CLI "handles model requests, auth, tools, and data under its own setup". Zed's guarantees stop at the ACP boundary. The router below is how we would read the page for a team deciding which path to permit.

Delta is Zed's "multiplayer environment for coding with agents and reviewing their work", introduced in an August 12, 2026 post and opened as a public beta on September 16, 2026. The idea is to invite teammates into agent threads and to replace pull requests with review subthreads, built on a storage layer Zed described in a June 11, 2026 post under the title "Software Is Made Between Commits". It runs on macOS, Linux, Windows and the web, is free during the beta, and Zed says it plans paid plans for individuals and teams while keeping a free version.

Two honest markers from Zed itself. The company says Delta is still "forming", and its own repository stays on GitHub for now, with the beta post citing 33 people landing 570 changes to main. That is the only part of the AI surface Zed explicitly labels beta; a buyer should treat it as the experimental tier and the Agent Panel as the product. Business administrators can switch collaboration features off for everyone, per the Zed for Business post, which is the control a regulated team will want while Delta is forming. If you are choosing an editor for a team and want these paths mapped against your own agent subscriptions and data obligations, our AI transformation engagements include that evaluation, and our AI IDE comparison covers the alternatives.