
Coding agents are much more useful when they can read your actual servers. Not push code, not exec commands, not open a shell. Just read logs, tail the current systemd unit output, cat a config file, and grep across the running deployment. That thin permission is enough for the agent to spot the missing environment variable that has been eating your evening for a week.
Setting up read-only server access safely is not the same as ssh-copy-id. You want a locked account with rbash or a forced command wrapper, an SSH proxy that logs everything, and an MCP server or CLI that surfaces the read-only calls to the agent. The seven best apps for AI coding agents with read-only server access below cover the agent, the transport, and the audit trail.
What to look for
- A coding agent that speaks MCP or the equivalent, so a filesystem or SSH tool can be exposed to it.
- Enforced read-only at the SSH layer, not just polite instructions in the prompt.
- Full session logging, so you can review every command the agent issued.
- Rotating short-lived credentials rather than a permanent private key.
- A tight allowlist of paths and commands.
- Local execution for the agent side, so source code never leaves the workstation without a decision.
Quick comparison table
| App | Best for | Platforms | Free plan | Starting price/mo | Rating |
|---|---|---|---|---|---|
| Claude Code | The agent that drives the read-only session | Windows, macOS, Linux | Free tier | $17 (Pro) | The leading coding CLI |
| Aider | Open-source terminal coding agent | Windows, macOS, Linux | Free (OSS) | Free (BYO API key) | Loved by developers |
| Continue | IDE-side coding agent with MCP support | VS Code, JetBrains | Free (OSS) | Free (BYO API key) | Growing quickly |
| MCP filesystem server | Read-only file exposure to any MCP client | Windows, macOS, Linux | Fully free (OSS) | Free | Reference implementation |
| Tailscale SSH | Zero-config, key-rotating SSH | Windows, macOS, Linux | Free (100 devices) | $6/user/mo | Widely used |
| Teleport | Audited SSH with recording | Linux server | Free (community) | $15/user/mo | Enterprise-grade |
| Zed AI | IDE with built-in server-aware agent | macOS, Linux, Windows | Free | $20 (Pro) | Fastest editor with AI |
The apps
1. Claude Code – the agent that drives the read-only session
Claude Code is Anthropic’s terminal-based coding CLI. Give it an MCP filesystem server pointing at a mounted read-only view of the remote host (via SSHFS, sftp, or a Tailscale-tunnelled MCP endpoint) and Claude can tail logs, read configs, and grep across running services while your local codebase stays writable. Claude never gets shell access on the server unless you explicitly hand it a bash tool.
Where it falls short: Requires an Anthropic API key and Pro or higher for meaningful usage. The MCP configuration for a remote host is manual the first time.
Pricing:
- Free tier with usage caps.
- Paid: Pro at $17/month, Max at higher usage tiers, API metered.
Platforms: Windows, macOS, Linux.
Download: Claude.com/code
Bottom line: Claude Code plus an MCP filesystem server plus a read-only SSH account is the current gold standard.
2. Aider – best open-source terminal agent
Aider is the mature open-source coding agent that most developers already run against local repos. Point Aider at a working directory that includes a read-only mount of the production server’s /var/log and /etc, and it will treat those files the same as your source. Aider does not push to the server; it only reads.
Where it falls short: No native MCP client (yet). File access is via the working directory, so the read-only mount has to be reliable.
Pricing:
- Free (Apache 2.0). BYO API key.
Platforms: Windows, macOS, Linux.
Download: Aider.chat
Bottom line: Aider is the free way to try this workflow before you subscribe to anything.
3. Continue – best IDE-side agent with MCP
Continue runs inside VS Code and JetBrains IDEs and speaks MCP natively. Connect an MCP filesystem server to the read-only remote view and Continue picks it up as another data source alongside your source files. Chat and inline edits both see server logs.
Where it falls short: Best experience in VS Code; JetBrains support lags a version. Requires a model provider (OpenAI, Anthropic, or a local Ollama endpoint).
Pricing:
- Free (Apache 2.0). BYO model provider.
Platforms: VS Code, JetBrains IDEs.
Download: Continue.dev
Bottom line: Continue is the IDE-native pick when you do not want to leave VS Code for a terminal.
4. MCP filesystem server – read-only file exposure
MCP filesystem server is the reference Model Context Protocol server that exposes a directory tree to any MCP-aware agent. Mount /var/log and /etc from the remote host over SSHFS, aim the MCP server at the mount, restrict its exposed path list, and every MCP client (Claude Code, Continue, Zed AI) sees the same read-only view.
Where it falls short: Reference implementation, minimal features. No fancy caching or partial reads.
Pricing:
- Free (MIT).
Platforms: Windows, macOS, Linux.
Download: modelcontextprotocol GitHub
Bottom line: The MCP filesystem server is the piece that connects everything else to the actual files.
5. Tailscale SSH – best zero-config secure transport
Tailscale SSH gives every device on your Tailscale tailnet SSH access using short-lived, tailnet-issued credentials instead of long-lived SSH keys. Combined with tailnet ACLs, you can grant your workstation ssh to readonly@server and nothing else. The ACL is one JSON file.
Where it falls short: Free tier caps device count. Requires Tailscale on both endpoints.
Pricing:
- Free: personal use up to 100 devices.
- Paid: Starter at $6/user/month.
Platforms: Windows, macOS, Linux, iOS, Android.
Download: Tailscale.com
Bottom line: Tailscale SSH removes the “how do I safely expose port 22?” step entirely.
6. Teleport – best for audited SSH with session recording
Teleport replaces SSH with a proxy that records every session as video-like playback, enforces role-based access, and issues short-lived certs from a central authority. Overkill for a solo developer; ideal when the same read-only workflow needs to survive a security review.
Where it falls short: Setup involves a Teleport cluster, even if it is a one-node cluster. Community edition is generous but enterprise features cost.
Pricing:
- Free: Community edition, self-hosted.
- Paid: Enterprise starts around $15/user/month.
Platforms: Linux server (proxy), any client OS.
Download: GoTeleport.com
Bottom line: Teleport is the pick when the compliance story matters as much as the coding one.
7. Zed AI – best IDE with built-in server-aware agent
Zed AI ships with an agent that treats MCP servers as first-class context sources. The editor is fast enough that scrolling through a 4 GB log file from the remote host does not freeze the UI. Zed’s collaborative multiplayer sessions let a teammate join your agent’s chat without exposing your credentials.
Where it falls short: Windows support is newer than macOS/Linux and still catching up. Model choice is narrower than Continue’s.
Pricing:
- Free.
- Paid: Zed Pro at $20/month for hosted models and priority.
Platforms: macOS, Linux, Windows.
Download: Zed.dev
Bottom line: Zed AI is the editor that pairs best with big remote log files.
How to pick the right one
- If your daily driver is a terminal, use Claude Code or Aider plus the MCP filesystem server.
- If your daily driver is VS Code or JetBrains, use Continue plus the MCP filesystem server.
- If your daily driver is Zed, use Zed AI and enable MCP directly.
- For the transport, Tailscale SSH is the fastest safe setup. If you need session recording for compliance, use Teleport.
- Never grant write access “just to try.” Read-only is the whole point.
Do not point an agent at a server that logs personal customer data unless you have anonymised the log stream first. The agent will summarise anything it reads.
FAQ
Is it safe to give an AI coding agent SSH access to my server?
Read-only, key-scoped, with session logging, yes. Full shell access, no. The safe pattern is a locked user account, an authorized_keys line with a command= prefix that runs a wrapper limiting reads to specific paths, and a session recorder in front of the proxy.
What is the difference between MCP and just giving the agent SSH?
MCP standardises how the agent asks for files and receives them. Instead of executing cat /var/log/app.log, the agent calls a read_file tool the MCP server implements. The server enforces the allowlist. The agent never touches a shell.
Can Aider or Claude Code write to my server?
Only if you configure them to. Both default to reading and to editing files in the local working directory. Any write to a remote host requires an explicit tool or a mount that is writable. Leaving the remote mount read-only prevents accidents.
Does this work with Windows servers?
Yes, over OpenSSH for Windows or WinRM behind a wrapper. Path conventions differ (C:\Users\... vs /home/...), so the MCP filesystem server needs Windows-style paths in its config.
What is the smallest safe setup?
Tailscale SSH on both ends, a readonly Linux user on the server with rbash and a restricted PATH, and the MCP filesystem server mounting /var/log and /etc from a workstation-side SSHFS mount. Claude Code or Aider on the workstation. Half an hour of setup total.