# Setting Up Devbox

> Guide a PostHog engineer through spinning up, connecting to, running commands on, and mirroring local code to a remote devbox (a Coder workspace running the full PostHog stack). Use when asked to set up a devbox, start or connect to a devbox, configure remote dev, get gh CLI / Claude Code authed on a devbox, run a command on a devbox, sync a local checkout so you can edit locally while the stack runs remotely (devbox:sync), or diagnose why a devbox command fails. Covers the tailnet prerequisite, hogli devbox commands, Coder user secrets for auth, one-way local->remote sync via mutagen, and verifying with devbox:exec. How each dev personalizes their box is left to them.

- Skill: `posthog/setting-up-devbox` (Agent Skill)
- Install (CLI): `npx skillmds@latest add posthog/setting-up-devbox`
- Raw SKILL.md: https://api.skillmd.com/api/skills/posthog/setting-up-devbox/raw
- Safety review: pending (external: skill-scanner PASS, skillspector CAUTION)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: posthog (https://skillmd.com/u/posthog)
- Updated: 2026-09-09
- Page: https://skillmd.com/skills/posthog/setting-up-devbox

---


# Setting up a PostHog devbox

A devbox is a Coder workspace running the full PostHog stack on an EC2 instance, managed through `hogli devbox:*` (the only supported interface — drive those commands, don't reimplement them). It ships ready to use: the repo cloned at `~/posthog`, the stack pre-warmed, and Claude Code installed. This skill gets a dev connected and working; how they personalize beyond that is their choice, not something to push.

## Prerequisite: tailnet access (the thing people miss)

The devbox control plane lives inside a private VPC reachable only over Tailscale. Two things have to be true, and they fail in ways that look identical at the reachability check — _every_ `hogli devbox:*` command dies there, and it is neither an auth nor an install problem, so no amount of re-running `devbox:setup` fixes it:

1. **You are signed into the `posthog.com` tailnet.** PostHog has several tailnets — `dev`, `prod-us`, `prod-eu` and `internal` exist for CI runners and subnet routers, and none of them route you to devboxes. Humans want `posthog.com` and nothing else.
2. **Your email is in `group:engineering`** in [`tailnet-policy.hujson`](https://github.com/PostHog/posthog-cloud-infra/blob/main/tailnet-policy.hujson) (`posthog-cloud-infra`), the ACL that grants the route to the Coder control plane (`10.70.0.1:443`).

Check the tailnet first — it's cheap, and it's the more common of the two:

```bash
tailscale switch --list                 # the active tailnet is marked; you want posthog.com
tailscale switch posthog.com            # if you've signed into it before
tailscale logout && tailscale login     # otherwise — pick posthog.com at the tailnet picker
# on macOS when `tailscale` isn't on PATH:
/Applications/Tailscale.app/Contents/MacOS/Tailscale switch --list
```

Being on the wrong tailnet is sticky and invisible: the picker only appears at first sign-in, so someone who clicked past it a year ago has been on `dev` ever since and was never asked again. In the GUI it's **Add account** → sign in as usual → select `posthog.com`. Suspect this whenever a dev says "it used to work on my other laptop" or is on a new machine.

If the tailnet is right and `hogli devbox:doctor` still reports the control plane unreachable, the fix is a PR adding the user to `group:engineering` in `tailnet-policy.hujson` (then ask Team DevEx if still blocked). Diagnose both of these before touching anything else.

### Control plane unreachable with a DNS cause?

When doctor shows `[ok] Tailscale connected` but fails reachability with a **DNS** cause (`DNS lookup for coder.dev.posthog.dev failed`), the name is never reaching the internal resolver. In order:

1. **Wrong tailnet** (above). The other tailnets have no route to `dev.posthog.dev`, and this is what it looks like. Rule it out first.
2. **MagicDNS off.** "Use Tailscale DNS" in the client's DNS settings is what points the machine at the internal resolver; without it nothing internal resolves.
3. **Stale resolver upstream.** MagicDNS on and names still failing, but `tailscale ping <internal-ip>` pongs and the service loads over its raw IP? That's the machine's router/ISP resolver. Adding `8.8.8.8` or `1.1.1.1` to the host's DNS settings has fixed this for several people.

To confirm it's resolution rather than the grant: `dig coder.dev.posthog.dev @10.90.0.2` answering while the system resolver fails proves the name exists and only the resolution path is missing.

**Do not reach for an exit node.** Devboxes are reached as tailnet peers, so an exit node is not required for devbox access — it just routes all of the dev's traffic through infra, which is noticeably slower, and it masks whichever of the three causes above is the real one. Don't suggest `/etc/hosts` or `/etc/resolver` workarounds either: they hardcode internal ELB IPs that rotate.

Full write-up, kept current by Team Cloud Foundations: [wiki.posthog.com/access/vpn](https://wiki.posthog.com/access/vpn#which-tailnet).

## Workflow

### 1. Check state — `hogli devbox:doctor`

```bash
hogli devbox:doctor          # read-only: tailnet access, reachability, auth, ssh config, saved setup
```

A safe probe — it never prompts or mutates host config (unlike `devbox:setup`). It names the active tailnet on its own line (`Tailnet: … (need posthog.com)`) and, when that's the problem, says so outright (`Cause: Signed into the 'dev' tailnet, not 'posthog.com'.`) — trust that over any other symptom. If it flags the control plane unreachable, resolve the tailnet (and then the ACL grant) before anything else. For more detail: `hogli devbox:list` (your boxes), `hogli devbox:status` (state, template freshness), `hogli devbox:secret:list` (secret names only).

### 2. One-time local setup — `hogli devbox:setup`

Interactive: checks Tailscale + Coder reachability, installs and authenticates the `coder` CLI (plus the pinned mutagen binary that backs `devbox:sync`), and writes the SSH host entries that `devbox:ssh`/`devbox:exec` rely on. It then _offers_ git identity, git signing, a dotfiles repo, and your Claude token — all optional; `--skip-*` anything you don't want. Re-run one step with its flag, e.g. `hogli devbox:setup --configure-git-signing`.

### 3. Start and connect — `hogli devbox:start`

```bash
hogli devbox:start           # create or resume your box
hogli devbox:ssh             # shell in
hogli devbox:open --vscode   # or --cursor / --web
hogli devbox:stop            # when done — preserves disk, stops billing
```

### 4. Fast QA and agent resume

Before recreating sync, restarting PostHog, or making a new devbox, check whether the existing box is already usable:

```bash
hogli devbox:status
hogli devbox:exec -- bash -lc 'cd ~/posthog && git status --short --branch && git rev-parse --short HEAD'
hogli devbox:sync --status
hogli devbox:exec -- bash -lc "curl -sf -o /dev/null -w '%{http_code}' http://127.0.0.1:8010/"
```

Use `-n <name>` on each command for a labeled box. If the forwarded app serves the intended branch/SHA, the target route loads, and route-critical APIs work, keep going and note unrelated degraded units rather than chasing perfect all-process health. When startup time matters, record rough timings for devbox start/resume, sync readiness, first route response, and first target-route load.

For agent-managed startup, `hogli devbox:start --start-app` is the supported path for new/stopped boxes. The flag is sticky and starts the regular PostHog stack in the background. If the box is already running without the app, start it inside the box:

```bash
hogli devbox:exec -- bash -lc 'cd ~/posthog && ./bin/hogli up -d -y'
```

Use the target route, route-critical APIs, and process-specific phrocs checks as the readiness gate.

### 5. Auth, if you want it (optional)

To have `gh` or Claude Code authenticated on the box, store the token once as a Coder user secret. It's injected as an env var into every box you start, so you set it once rather than per box:

```bash
hogli devbox:secret:set GH_TOKEN --env GH_TOKEN
hogli devbox:secret:set CLAUDE_CODE_OAUTH_TOKEN --env CLAUDE_CODE_OAUTH_TOKEN
# also supported: ANTHROPIC_API_KEY, OPENAI_API_KEY, OP_SERVICE_ACCOUNT_TOKEN, AWS_CREDENTIALS (--file)
```

Authing `gh` / Claude on a devbox is fine — that's what these are for. Set the value from `--file` or the hidden prompt; never paste a token into a command line or into this conversation. Restart a running box to pick up a newly set secret.

### 6. Make it yours — your call

The box is usable as shipped; personalize it however suits you, or not at all. Two supported paths, neither required, don't push one over the other:

- **Tweak the box directly** — `devbox:ssh` in and install tools, add aliases, clone repos. Changes under `/home` survive stop/start and template updates, but a `devbox:destroy` (or a brand-new box) starts fresh.
- **A dotfiles repo** — if you'd rather keep portable, version-controlled config that re-applies to every box: `hogli devbox:setup --configure-dotfiles` points the box at your `dotfiles_uri`, and Coder clones it (running an executable `~/dotfiles/install.sh` if present) on each start.

### 7. Run commands on the box — `hogli devbox:exec`

`devbox:exec` runs one command over SSH and propagates its exit code — handy for scripts, agents, and quick checks without opening a shell:

```bash
hogli devbox:exec -- bash -lc 'gh auth status'
hogli devbox:exec -- bash -lc 'cd ~/posthog && git status'
hogli devbox:exec -n api -- bash -lc 'uname -a'    # -n targets a labeled box
```

Wrap commands in `bash -lc '...'`: a non-login shell doesn't reliably source `~/.bashrc`/`~/.zshrc`, so a bare `gh auth status` can report "command not found" for anything on a login-shell `PATH` (e.g. `~/.local/bin`) — a false negative. The login shell also keeps the exit code trustworthy, so `&&` chaining and `if` checks work. Use `--` to separate hogli's flags from the command's own.

`devbox:exec` is not side-effect-free: like every `devbox:*` command it runs the reachability check first, which on Linux may `sudo tailscale set --accept-routes` and prompt for a password. Run `hogli devbox:setup` once interactively so routes and SSH config are in place before an agent drives `devbox:exec` unattended.

## Editing locally, running on the box — `hogli devbox:sync`

When you want your fast local checkout to stay the place you edit but the heavy stack (`hogli up`) to run on the box, `hogli devbox:sync` mirrors your repo onto the box over [mutagen](https://mutagen.io), one-way: local is the source of truth, nothing comes back. Reach for this in an agentic loop — edit with your normal local tools, let the mirror carry each change, and drive the remote stack with `devbox:exec` — instead of committing and pushing every iteration or editing over Remote-SSH.

```bash
hogli devbox:start                                   # the box must be running first
hogli devbox:sync                                    # create the mirror (idempotent: re-run just reports status)
# edit files locally — changes propagate within seconds
hogli devbox:exec -- bash -lc 'cd ~/posthog && pnpm --filter=@posthog/frontend typescript:check'
hogli devbox:sync --status                           # watching / paused / conflicts
hogli devbox:sync --terminate                        # tear the mirror down when done
```

Before relying on sync, verify the remote branch/SHA and `devbox:sync --status`. If the box already matches and sync is watching without source-file conflicts, do not recreate it for cleanliness. Source-file conflicts block reliable QA until resolved; box-local config conflicts such as `.env` can be acceptable when the tracked source is clean.

The non-obvious parts:

- **It runs on your machine and pushes to the box — not the reverse.** Don't invoke it through `devbox:exec`. It mirrors whichever checkout you run it from (it walks up from the cwd for `hogli.yaml` + `.git`), so run it from the repo root you are editing — including a `/wt` worktree.
- **`one-way-safe` preserves remote-only files.** The AMI's prewarmed `node_modules`, venv, and `target/` are never deleted — they aren't in your local checkout and the mode leaves remote-only content alone. Lockfiles _do_ sync, so the box reconciles deps on its next start.
- **The first sync of a feature branch conflicts per diverged file.** The AMI is always on `master`; every file your branch changed relative to the box's `master` surfaces as a conflict in `--status`. That is expected one-way-safe behavior, and it is per-path — non-conflicting files (including brand-new ones) still sync. Resolve a path, or check the matching branch out on the box, only if you specifically need that file mirrored.
- **Don't also edit those files on the box.** Editing over Remote-SSH while the mirror is live fights the local source of truth; `devbox:open --vscode|--cursor` warns when a sync is active for exactly this reason.

The packaged ignore defaults are seeded once to `~/.hogli/mutagen.yml` and never overwritten — it is yours to tweak. If a newer hogli ships updated ignore defaults, `rm ~/.hogli/mutagen.yml` and re-run `devbox:setup` to pick them up.

## Persistence & multiple boxes

- `devbox:stop` → `devbox:start` and template/AMI updates preserve `/home` (the instance is stopped, not terminated). A `devbox:destroy` wipes it — intentional, so don't keep anything irreplaceable only inside a box.
- You can run more than one box. Box-local changes don't carry between them; user secrets do (user-scoped), and a dotfiles repo does if you use one. That's the practical reason to reach for those if you find yourself re-doing setup — but it's a choice, not a requirement.

## Gotchas

- **Never echo secret values** into the transcript, logs, a PR, or a command line. `devbox:secret:set` reads from a hidden prompt or `--file`; `secret:list` shows names only. Keep it that way.
- **Secrets need a restart.** A new or changed secret only reaches boxes started afterward — `hogli devbox:restart` to pick it up on a running box.
- **`devbox:exec`/`devbox:ssh` need `devbox:setup` to have run** (it writes the `coder.*` SSH host config). Without it they fail at connection; `devbox:doctor` shows whether SSH access is configured.
- **`code-server` (browser IDE) has no SSH agent forwarding**, so commit signing via a forwarded key won't work there — use VS Code Desktop / Cursor / JetBrains (SSH-based) when you need to sign.