AgentLink
A coding agent harness built into VS Code
AgentLink gives coding agents VS Code’s intelligence, reviewable changes, visible execution, and adjustable supervision. Separate, narrower Desktop and CLI previews run on macOS Apple Silicon.
Review the change, not just the answer
The full VS Code extension keeps proposed edits in the editor you already use.
Why a harness matters
A capable model is only part of a capable coding agent. It also needs a good working environment: editor intelligence instead of text guesses, fast feedback instead of a surprise broken build, and a way for you to supervise or redirect work without taking the controls away.
Inside the VS Code extension
Better context, better feedback, and better control make agents more useful on real codebases.
The editor is the runtime
Agents get language server navigation, diagnostics, symbols, and workspace-aware rename. Edits arrive as native diffs you can accept, reject, or adjust.
Autonomy is a dial, not a switch
Review every action on risky work, or let familiar tasks move faster with focused rules and Approve for Me. Checkpoints make experiments reversible.
Nothing happens in the dark
Tool calls, approvals, questions, and background agents stay visible in the chat and Activity Shelf, with a browser remote for supervision on the go.
A second opinion from another provider
Run background research and review with visible Fleet progress and steering. Route reviews to another configured provider or external ACP agent for a second opinion.
Context for the code, not the harness
Local lexical and structural retrieval, progressive tool disclosure, bounded terminal output, and session condensing keep the active task in focus.
Bring your own models and workflow
Use ChatGPT/Codex, OpenAI, or configured OpenAI-compatible providers on your own accounts. In VS Code, connect MCP tools, resources and prompts, plus skills, hooks and project instructions.
One engine, several surfaces
VS Code has the fullest editor and review integration. Desktop offers standalone Ask Agent chat and a view into existing local VS Code sessions; the CLI runs terminal coding sessions. Sharing a core does not mean sharing every tool or permission.
VS Code extension
The fullest experience: language-server intelligence, native diff review, visible terminals, checkpoints, background agents, Fleet and configurable approvals.
Browser remote
Supervise a running VS Code session, answer questions and inspect read-only diffs. There is no remote shell or file-writing path.
Desktop app
Standalone Ask Agent chat and Quick Ask without VS Code. Optionally view existing AgentLink VS Code sessions on the same Mac; standalone chat does not inherit editor tools.
CLI
A self-contained terminal coding agent with reviewed single-file edits and commands, approved MCP tools, scoped background agents and optional TypeScript/JavaScript intelligence. Narrower than VS Code.
Shared foundations, different capabilities
- Use your own provider accounts. Model requests send the context needed for the task to your selected provider; MCP services receive the data you send them.
- Lexical and structural codebase retrieval and default telemetry stay local. Optional embedding services have their own data boundary; local retrieval does not mean offline model inference.
- Tool sets and approvals depend on the surface. Desktop’s VS Code view supervises the existing host session rather than granting standalone chat coding tools.
- Desktop and CLI are separate unsigned, not-notarised Apple Silicon previews, not bundled with the VSIX. Neither supports Intel Macs, Linux or Windows.
Desktop: ask without opening an editor
A separate macOS Apple Silicon preview—not the full VS Code tool environment.
CLI: a local terminal coding session
A self-contained Apple Silicon preview with reviewed single-file edits and commands, project sessions, and optional TypeScript/JavaScript intelligence.
Get started with VS Code
curl -fsSL https://raw.githubusercontent.com/reefbarman/agentlink/main/scripts/install.sh | bashThe extension installer selects the extension-host target and verifies a checksum when published. For remote or emulated hosts, set AGENTLINK_VSCE_TARGET explicitly (for example linux-x64). Use bash -s -- --dry-run to preview or --version X.Y.Z to pin. Extension targets include macOS Apple Silicon and Windows/Linux x64 or arm64; Intel Macs are no longer supported.
Use VS Code 1.140 or newer. Install with the script above, or download the matching platform .vsix from the latest extension release.
Reload VS Code, open AgentLink from the Activity Bar, and choose Agent.
Sign in with ChatGPT/Codex, add an OpenAI API key, or configure another compatible provider.
Start with a bounded request, review the diffs, and use /checkpoint before risky work.
Desktop preview
curl -fsSL https://raw.githubusercontent.com/reefbarman/agentlink/main/scripts/install.sh | bash -s -- --surface desktopmacOS Apple Silicon only; unsigned and not notarised. No Intel Mac support. Drag AgentLink from the DMG to Applications. macOS may require explicit permission for first launch. Check the setup guide for security warnings and limitations.
CLI preview
curl -fsSL https://raw.githubusercontent.com/reefbarman/agentlink/main/scripts/install.sh | bash -s -- --surface climacOS Apple Silicon only; unsigned and not notarised. No Intel Mac support. The installer verifies the SHA-256 checksum and links ~/.local/bin/agentlink. macOS may warn or block downloaded executables. Check the setup guide for security warnings and limitations.
Embedding AgentLink in a Node host? The SDK’s first release is still pending; use a source checkout for now, not npm or a release installer. See the embedding guide.
What AgentLink is not
It does not train or sell a proprietary model, and there is no cloud middleman between you and your provider.
It is not a VS Code fork, so your extensions, marketplace, and language tooling stay yours.
Desktop and CLI can run without VS Code, but are narrower previews—not replacements for the full extension.
The browser remote is for supervision only: diffs are read only and there is no remote shell or write path.
Best-of-N and scheduled automation are still maturing, not generally available workflows. The Node SDK’s first release is pending; embedding currently requires vendoring from a source checkout, not an npm install.
Rolling out coding agents across your team?
I help engineering teams adopt AI coding agents with the right tooling, guardrails, and workflows, whether that is AgentLink or another harness.
- Tooling and harness selection
- Shared AGENTS.md, skills, and MCP setup
- Review practices and guardrails