Zed
Also known as: Zed Zeta, Agent Client Protocol
Open-source, GPU-accelerated Rust code editor with a native AI agent panel, parallel agents in isolated worktrees, per-tool permissions, and the open Agent Client Protocol that lets Claude Agent, Codex and OpenCode run inside the editor.
Zed is a code editor built from scratch in Rust by Zed Industries, the team behind the Atom editor, and engineered above all for speed. It uses the graphics processor for rendering and multiple cores for indexing and language servers, so startup, scrolling, search, and keystrokes feel instant in a way Electron based editors like VS Code cannot match.
It runs natively on macOS, Linux, and Windows, is open source under the Apache license, and treats real time collaboration as a first class feature, letting teammates share a project with live cursors, chat, voice, and screen sharing without plugins. It is not a VS Code fork, so its extension ecosystem is smaller but its editing engine is its own.
Since 2025 Zed has invested heavily in agentic AI. Its Agent Panel runs a native agent that discovers context, edits files, and runs tools across a multi step task, and parallel agents can run in several threads across different projects at once. An open weight edit prediction model called Zeta anticipates your next edit at keystroke granularity and can run locally. Zed also created the open Agent Client Protocol, so external agents like Claude Code, Codex, and OpenCode plug directly into the editor, and Terminal Threads run any command line agent as a persistent sidebar. Model Context Protocol servers extend an agent's tools and knowledge.
Zed does not lock users into one model. It connects to fifteen providers including Anthropic, OpenAI, Google, Amazon Bedrock, and Mistral, plus local models through Ollama and LM Studio, offering Zed hosted models with no API keys or your own keys, in which case data goes straight to your provider. Oversight is built in: you follow the agent live as it works, review its changes in an editable unified diff and accept or reject each one before committing, and use a checkpoint system to revert to the state before any edit batch. Context is added by mentioning files, symbols, diagnostics, and past threads.
For teams, a Business tier adds an admin layer with org wide model policies, data sharing controls, role based access, and unified spend visibility, though single sign on, SCIM, and a SOC 2 certification are on the roadmap rather than shipped. Being open source and supporting local and open weight models, Zed keeps code under the team's control, with remote development letting the codebase live on a server while the machine runs only the interface.
Pricing is simple and low: a free Personal tier with a monthly edit prediction allowance and unlimited use of your own keys, Pro at ten dollars a month for unlimited prediction and hosted credits, and Business at thirty dollars per seat. Zed is the speed and craft choice, with native agent depth still maturing against dedicated rivals.
Vendor details
Canonical URL
https://zed.dev
Category
Coding agent
Subcategory
Open source native AI code editor with agent panel and ACP (Rust)
Funding status
Zed is built by Zed Industries, Inc., an independent company founded by the creators of the Atom editor, which Microsoft acquired and later sunset before releasing VS Code. The editor is open source under the Apache 2.0 license. Specific venture funding is not detailed in available sources. Zed accepts optional sponsorship through GitHub Sponsors, which does not grant paid features. The product reached stable releases across macOS, Linux, and Windows by 2026 and reports a growing ecosystem of extensions and agent integrations.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Runs a native first-party agent with built-in tools and hosts external agents including Claude Agent, Codex, Copilot, Cursor, Gemini CLI and OpenCode through the open Agent Client Protocol, which Zed itself authored and publishes; Terminal Threads run any command line agent as a persistent sidebar. MCP servers extend agent tools and knowledge, and Tool Permissions govern what those tools may do. The editor has first-class Git support for staging, committing, pushing and diffs, native debugging through the Debug Adapter Protocol, language server support, roughly 800 extensions, and remote development over SSH. Three model paths are documented separately: Zed-hosted models with no API keys, local models through Ollama or LM Studio where nothing leaves the machine, and bring-your-own-key to any provider, spanning fifteen providers including Anthropic, OpenAI, Google, Amazon Bedrock, Mistral, OpenRouter and Vercel. Context is added by mentioning files, directories, symbols, diagnostics, branch diffs, threads and fetched URLs.
In practice
You want a coding agent inside a fast native editor. Zed's Agent Panel edits files while you follow along live and review every change, and you can plug in Claude Code through its open protocol.
You care about editor speed and privacy. Zed is open source, runs local models through Ollama, and lets you bring your own keys so code and prompts go to your provider rather than Zed.
Your team pairs and reviews together. Zed's native multiplayer gives shared buffers, live cursors, voice, and screen sharing, and its Business tier adds org wide model policies, role based access, and unified spend visibility.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
10.0 / 14 capabilities · 71%
| Integrations & Tool Calling | Full |
|---|---|
|
A native first-party agent runs with built-in tools while external agents including Claude Agent, Codex, Copilot, Cursor, Gemini CLI and OpenCode connect through the open Agent Client Protocol, and Terminal Threads run any command line agent as a persistent sidebar; MCP servers extend agent tools and knowledge with dedicated documentation. The editor itself provides first-class Git operations for staging, committing, pushing and diffs, native debugging through the Debug Adapter Protocol, language server support, roughly 800 extensions, and remote development over SSH. Context is added by mentioning files, directories, symbols, diagnostics, branch diffs, previous threads and fetched URLs. Sourcezed.dev/ai, zed.dev/docs/ai/mcp and zed.dev/extensionsread 2026-08-30 |
|
| Workflow Orchestration | Full |
|
Zed Agent's spawn_agent tool starts a subagent with its own context window for a delegated task, such as a parallel investigation while the parent continues and then reviews the findings; parallel agent threads and Terminal Threads run concurrently from the Threads Sidebar across projects, with Git worktrees isolating each thread's changes. Sourcezed.dev/docs/ai/tools and /docs/ai/parallel-agentsread 2026-09-29 |
|
| Knowledge Grounding & RAG | Partial |
|
The agent searches the codebase for relevant context during a task, and developers add context explicitly by mentioning files, directories, symbols, diagnostics, branch diffs, previous threads and fetched URLs, with multi-core indexing and language servers making retrieval fast; Agent Profiles, Skills and Instructions carry standing project guidance, and thread history is restorable as context. No repository index product, semantic retrieval layer, dependency graph or cross-file reasoning surface is documented, and project-wide context depends substantially on what the developer passes in. Sourcezed.dev/docs/ai/agent-panel and zed.dev/docs/ai/overviewread 2026-08-30 |
|
| Human Oversight & Guardrails | Full |
|
Tool approval is controlled by agent.tool_permissions: a global default of allow or confirm, per-tool defaults, always_allow patterns, always_confirm patterns that prompt case by case (for example sudo commands) regardless of other settings, and auto-deny rules for dangerous actions that hold even when the default is allow; Zed Agent tool calls can also run in an OS-level sandbox. After a run the developer reviews changes in an editable unified diff, accepting or rejecting each, and checkpoints revert any edit batch. Sourcezed.dev/docs/ai/tool-permissions, /docs/ai/sandboxing and /docs/ai/agent-panelread 2026-09-29 |
|
| Security, Identity & Governance | Partial |
|
The company's SOC2 page states it is working toward SOC 2 Type 1 certification, and the pricing page lists SSO, SAML and SCIM as planned but not available. The Business admin layer is shipped: organization roles and permissions, org-wide AI model policies, data governance controls and unified spend visibility, with local and bring-your-own-key paths keeping code away from Zed. Sourcezed.dev/docs/soc2, /docs/roles and zed.dev/pricingread 2026-09-29 |
|
| Observability & Auditability | Partial |
|
Developers follow the agent live as it navigates the codebase, using the same multiplayer infrastructure that powers human collaboration, with responses streaming and indicators showing which tools the model is using; token usage is displayed, threads are restorable from history, unified spend visibility is available to teams on Business, and the vendor publishes public agent metrics comparing usage across Zed Agent, Claude Agent and Codex CLI over rolling windows. Thread history, checkpoints and token display depend on the agent integration and vary for external agents. No retained per-action execution trace, audit log or replay is documented. Sourcezed.dev/docs/ai/agent-panel, zed.dev/agent-metrics and zed.dev/businessread 2026-08-30 |
|
| Memory & State Persistence | Partial |
|
A checkpoint system reverts the workspace to the state before any edit batch, and threads are restorable from history so a session can be resumed later, with the docs noting these features depend on the agent integration and vary across external agents; Agent Profiles, Skills and Instructions persist as customer-authored configuration the agent reads on every run, and parallel Git worktrees hold isolated state per agent thread. No memory layer accumulating learned context, preferences or corrections across sessions is documented. Sourcezed.dev/docs/ai/agent-panel and zed.dev/docs/ai/overviewread 2026-08-30 |
|
| Deployment & Data Residency | Full |
|
Zed is open source under Apache 2.0 and runs as a native application on macOS, Linux and Windows entirely on the developer's own machine, with the source publicly cloneable; local models run through Ollama or LM Studio with the vendor stating nothing leaves the machine, bring-your-own-key sends data directly to the customer's chosen provider rather than through Zed, and the Zeta edit-prediction model is open-weight with self-hosted Next Edit models supported. Remote development keeps the codebase on a server the customer controls while the local machine runs only the interface, and the editor works fully without AI features or any signed-in account. Sourcezed.dev/ai, zed.dev/docs/ai/llm-providers and zed.dev/edit-predictionread 2026-08-30 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
The ACP Registry is the primary way to install External Agents: opened from the command palette or Agent Settings, it lists ready agents such as Claude Agent, Codex, Gemini CLI, OpenCode, Copilot, Cursor and Pi Coding Agent, and an installed agent appears in the new-thread menu. Agent Server Extensions publish further agents through the Zed Extension Registry, and Zed Agent is shaped by reusable Agent Profiles, Skills and Instructions. Each agent in the registry installs in one step. Sourcezed.dev/docs/ai/external-agents and /docs/extensions/agent-serversread 2026-09-29 |
|
| Triggers & Channel Coverage | Partial |
|
Zed runs natively on macOS, Linux and Windows with the AI features reached from the Agent Panel inside the editor, alongside a Threads sidebar, Terminal Threads running command line agents persistently, inline assistance in the buffer, remote development over SSH against a codebase held on a server, and real-time collaboration sessions where multiple humans and the agent share the same workspace. Invocation is by the developer typing in the message editor or invoking inline assistance; no event-driven trigger, scheduled run or inbound chat or ticketing channel is documented. Sourcezed.dev/ai and zed.dev/docs/ai/agent-panelread 2026-08-30 |
|
| Model Flexibility & Routing | Full |
|
Three model paths are documented separately: Zed-hosted models including Claude Opus and GPT 5.4 with no API keys to manage, local models through Ollama with Zed stating nothing leaves the machine, and bring your own keys from any provider using the customer's existing accounts and budgets, with LM Studio also supported. Fifteen providers are supported, including Anthropic, OpenAI, Google, Amazon Bedrock, Mistral, OpenRouter and Vercel, and merged community contributions continue to add providers and models including Gemini 3.7 Flash and OpenRouter effort-based reasoning. The Zeta edit-prediction model is open-weight and self-hostable, with self-hosted Sweep Next Edit models also supported. Sourcezed.dev/ai, zed.dev/docs/ai/llm-providers and zed.dev/edit-predictionread 2026-08-30 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
Zed authored and publishes the Agent Client Protocol, an open standard with its own site for connecting any agent to any client, and the vendor states Zed, the Zeta model and ACP are all open source, built in public; the editor is Apache 2.0 with a public repository, a defined extension interface and roughly 800 published extensions, MCP server support extends agent tools and knowledge with its own documentation, and Tool Permissions expose configuration for what those tools may do. Community contributions to the AI subsystem are merged continuously and listed publicly, including provider additions and agent model updates. Sourcezed.dev/acp, zed.dev/ai and zed.dev/extensionsread 2026-08-30 |
|
| Testing, Debugging & Optimization | Partial |
|
Native multi-language debugging is built in through the Debug Adapter Protocol, language servers surface diagnostics that can be mentioned as agent context and acted on, and the agent runs tests and other commands as tools within the workspace, with results visible as it works. The vendor publishes editor performance benchmarks and a public agent metrics page comparing how developers use different agents, both of which measure Zed and the ecosystem rather than providing a harness the customer points at agent behavior. No test generation, evaluation or regression testing product is documented. Sourcezed.dev/ai and zed.dev/docs/ai/agent-panelread 2026-08-30 |
|
| Browser & Computer Use | Not documented |
|
The agent operates through programmatic surfaces throughout: file reads and edits in the workspace, tool calls governed by the Tool Permissions system, MCP servers, Terminal Threads running command line processes, native debugging through the Debug Adapter Protocol, language servers, and Git operations. No browser control, screenshot capture, visual verification or capability to operate software lacking a programmatic interface is documented on zed.dev/ai, the Agent Panel documentation or the MCP documentation. External agents connected over ACP may carry their own browser tooling, which belongs to those agents rather than to Zed. Sourcezed.dev/ai and zed.dev/docs/ai/agent-panelread 2026-08-30 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Recent platform changes
A parent agent in Zed 1.22.0 can now choose the model for each subagent it spawns instead of using one configured subagent model, and subagents now compact their context when it fills up.
Bears on: Workflow orchestration
View sourceZed released version 1.14.2, introducing sandboxing capabilities for the Agent's terminal commands and web fetches. The update also adds support for undoing and redoing file operations directly within the Project Panel and introduces configurable fonts for the Agent Panel.
Bears on: Security / enterprise
View sourceZed shipped versions 1.12.0 (preview) and 1.11.3 (stable), introducing sandboxing for the agent's terminal and fetch tools. The releases enable ACP (Agentic Communication Protocol) elicitations by default to gather structured user input, add support for searching inside terminal threads, and expand AI model support with GPT-5.6 via Amazon Bedrock.
Bears on: Security / enterprise
View sourcePricing
Personal is free forever with 2,000 accepted edit predictions a month and unlimited use of your own keys. Pro is $10 a month for unlimited prediction plus $5 of hosted model credit. Business is $30 per seat. Hosted usage beyond credit is billed at provider list price plus 10 percent.
A free editor with a monthly edit prediction allowance, then a low flat subscription. Paid tiers add unlimited edit prediction and a small hosted model credit, after which hosted usage is billed by tokens at the provider's list price plus ten percent. Bringing your own key or running local models avoids that usage cost entirely.
Included quota
Personal, free: two thousand accepted edit predictions a month, unlimited use of your own keys and external agents, local models, and full real time collaboration. Pro at $10: unlimited edit predictions, $5 of monthly hosted model credit, then hosted usage at provider list price plus ten percent, and priority access to frontier models. Business at $30 per seat: an admin layer with org wide model policies, data governance, role based access, and unified spend, with members able to keep their own Pro plans. Students get Pro features free.
What is public
Public and exact: the three tier prices, the free tier contents, the ten percent usage markup, and the roadmap status of enterprise identity features. Set by providers: the underlying per token model rates.
Billing mechanics
A free Personal tier, Pro at $10 a month, and Business at $30 per seat, with hosted model usage metered by tokens at provider list price plus ten percent above a small credit, and bring your own key or local models as a zero markup path.
Cost watchouts
Hosted model usage beyond the small included credit is billed by tokens at the provider's list price plus ten percent, so heavy use of frontier models like Opus adds up. Bringing your own key or using local models avoids the markup but shifts spend to your provider or hardware. Business seats do not bundle a fixed model credit.
Variable cost rationale
Hosted model usage is token metered at provider list price plus ten percent above a small monthly credit, so cost scales with how much and which models are used, which is real exposure. However, a free tier, unlimited edit prediction on Pro, and a bring your own key or local model path let a developer cap or avoid usage billing entirely, so exposure is moderate and controllable rather than high.
Additional watchouts
Frontier model usage on hosted billing adds up quickly beyond the small credit, and enterprise identity and compliance features, single sign on, SAML, SCIM, and SOC 2, are not yet shipped, which may block regulated buyers despite the strong data control from open source and bring your own key.
Overage / add-ons
Hosted model usage above the included monthly credit is billed by tokens at the provider's list price plus ten percent, charged either at month end or per $10 of usage, whichever comes first. Bringing your own key routes spend directly to the provider with no markup.
Sales call required
No, self serve available
Free / trial
Personal is free forever with 2,000 accepted edit predictions a month and unlimited use of your own keys, external agents and local models. Pro has a 14-day free trial with $5 of GPT-5.6 Luna usage and unlimited edit predictions, no card required. Business has no trial and no seat minimum.
Lowest paid plan
Pro at $10 a month, above a free Personal tier. Pro dropped from $20 to $10 in May 2026.
Commercial notes
At $10 a month, Pro is inexpensive for a paid AI editor plan, and the open-source editor plus the bring-your-own-key path means many developers pay nothing. Transparent token billing at provider list price plus ten percent is a deliberate contrast to bundled credit systems, appealing to developers who want predictable, provider-aligned costs.
Key ambiguities
Because hosted model usage is token metered above a small credit, the real monthly cost depends on which models are used and how heavily, though bringing your own key or local models makes cost predictable. Business seats meter hosted usage separately rather than bundling a fixed allowance.
Cancellation / refund
Self serve monthly plans can be changed or canceled by the user. There is a two week Pro trial with $20 of credits. The editor itself is free and open source, so there is no lock in at that level.
Missing data
Annual pricing, if any, and exact hosted model rates per provider are set by the underlying providers plus the ten percent markup rather than published as fixed Zed prices.
Related vendors
- Cognition — Maker of Devin, an autonomous AI software engineer
- 10Web — Agentic website platform whose specialized AI agents build, host,…
- AgentUI — Managed AI app builder for operations teams that generates and hosts…
- Aider — Open source, model agnostic terminal coding agent that edits your…
- Anthropic Claude Code — Anthropic's agentic coding system across terminal, desktop, IDE, web…
- AppFactor — Agentic platform that runs persistent agents across an enterprise…
Alternatives to Zed
The closest documented capability profiles to Zed among coding agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- JetBrains AI10.5 / 14Fuller documented coverage on Triggers & Channel Coverage
- OpenCode10.0 / 14Fuller documented coverage on Triggers & Channel Coverage
- 10Web8.5 / 14A lighter documented profile than Zed
- Cosine8.5 / 14A lighter documented profile than Zed
- Potpie AI10.5 / 14Fuller documented coverage on Knowledge Grounding & RAG and Memory & State Persistence
- Augment Code12.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and Security, Identity & Governance
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded