ai-sandbox

Run an autonomous coding agent against your real project, with the escape routes closed: only the current directory is mounted, no ambient environment variables, no root inside, hard resource caps, and outbound network limited to an explicit allowlist that the agent cannot lift.

Single bash script, rootless podman, Ubuntu 24.04. No daemon, no k8s, no config language.

$ cd ~/Projects/billing-api
$ pi --allow aipi.reeseapps.com --env AI_API_TOKEN
[aisb] egress gate aisb-gate-7f3c1a -> 127.0.0.1:18084 (1 domains, 0 ip/cidr entries)
[aisb] egress net=allow enforce=nft proxy=[100.64.0.1:18084] direct=0 dns=0
 pi

pi there is an alias for ai-sandbox. Inside, the agent sees your cwd at the path it has on the host — ~/Projects/billing-api, not a generic /project — read-write, and nothing else on your disk, with the internet only where you pointed it.

Install

git clone <this repo> && cd ai-sandbox
./install.sh --alias pi          # symlinks to ~/.local/bin and adds the alias to ~/.bashrc
ai-sandbox doctor                # checks podman, rootless, subuid, pasta, SELinux, uid mapping
ai-sandbox build                 # optional: build the image now instead of on first run
make test                        # 390+ assertions (unit needs no containers; --e2e builds and runs)

Requirements: podman ≥ 5 running rootless (with a /etc/subuid range for your user), and either pasta (default in podman 5) or slirp4netns. Nothing else — shellcheck, tests and the image build all run through podman.

The alias, and how arguments are read

Sandbox flags are consumed until the first positional argument or unknown flag; those become arguments to the harness:

alias pi='ai-sandbox'

pi                                     # runs pi inside the sandbox
pi --allow aipi.reeseapps.com --env AI_API_TOKEN --model turbo
#     ^^^^^ sandbox ^^^^^^^^^^^^^^^^^^^^^^^^^^^         ^^^^^^^^^^ for pi
pi --shell                             # a bash shell in the same sandbox
pi --net none -- make test             # -- replaces the command entirely

--strict-args turns an unrecognised flag into an error instead of a pass-through, which is what you want in scripts.

Network policy

mode	what it means
--net allow (default)	Only the allowlist is reachable. A netfilter ruleset inside the container plus a domain-filtering proxy on the host loopback. If the ruleset cannot be installed, the sandbox refuses to start (exit 77).
--net none	No network at all.
--net full	Nothing restricted. Prints a warning. Never the default.

ai-sandbox --allow aipi.reeseapps.com --allow registry.npmjs.org --env AI_API_TOKEN
ai-sandbox --allow 10.20.0.5:5432            # a database: pinned IP and port
ai-sandbox --allow files.pythonhosted.org --allow pypi.org
ai-sandbox --no-gate --allow example.com     # no proxy container: pin resolved IPs instead
ai-sandbox --proxy corp-proxy.internal:3128  # route through a proxy you already run

Domains match that name and anything below it, and are enforced by the gate at CONNECT time, so CDNs and rotating IPs are not a problem. IP/CIDR entries become direct netfilter rules for non-HTTP destinations (databases, etc.).

An allowlist is required in allow mode: an empty list is a hard error, never a silent free-for-all. Keep it in the repo so the policy is reviewable:

# .ai-sandbox/allow  (one entry per line, # comments fine)
aipi.reeseapps.com
registry.npmjs.org

If you bridge your pi config (below), ai-sandbox tells you when the model endpoint your own models.json points at is not in the allowlist -- pi would otherwise start and sit there unable to answer, which reads like a broken sandbox.

Environment

Nothing is inherited except TERM LANG LC_ALL LC_CTYPE TZ.

--env AI_API_TOKEN               # from your environment
--env FOO=bar                    # literal (shows up in ps; prefer the two below)
--env-file ~/.config/secrets     # a 0600 NAME=VALUE file

Values are handed to podman through a 0600 file that is removed when the run ends, so they never appear in ps, in the printed dry-run, or in the session log. Asking for a variable that is not set fails the run closed.

Filesystem

Your cwd is mounted at the path it has on the host, and that is the working directory. So pwd, an error the agent quotes back, a commit message and the window title your harness derives from the cwd all name this project instead of reading the same as every sandbox in the session. The mount itself is unchanged by the naming: same directory, same read-write, nothing else visible except the one exception below.

A bind mount hides whatever was already at that path in the image, so a host path that would land on something the image or the sandbox owns (/usr, /etc, /var, /opt, /tmp, /run, /home/sandbox, /) falls back to /project — and says so, because a working directory you did not choose is worth a line:

[aisb] ai-sandbox: cannot mirror /tmp/scratch/billing-api: the sandbox owns /tmp*, so the
project is mounted at /project instead (--workdir DIR to choose)

/tmp is the one case worth naming: the sandbox mounts a tmpfs there, and podman puts that tmpfs over a bind of the same path, so a project under /tmp would come out as an empty directory that happens to be writable and to have the right name. Measured, not guessed — see docs/networking.md.

--workdir DIR picks the mount point yourself (auto, the default, is the mirroring behaviour above). The table below is everything else, all opt-in:

--mount ~/.ssh:ro                # if the agent must be able to push
--mount ../docs:/docs:ro         # read-only reference material
--workdir /src                   # mount and cd there instead of mirroring the host path
--no-cwd                         # a scratch sandbox with nothing mounted (lands in /project)
--read-only                      # read-only root filesystem
--ephemeral-home                 # do not persist /home/sandbox (default: persist in a volume)

Files the agent writes land on your disk owned by your uid, not root (--userns=keep-id, verified by doctor).

Your .agents is shared, read-only

A harness with no skills is a harness you have to re-brief on every run, so ~/.agents — skills, prompt templates, your standing AGENTS.md — is bind-mounted at /home/sandbox/.agents, which is where pi looks for user-level skills. Nothing is copied; it is one mount, and it is read-only:

pi                                  # your skills are there
pi --no-agents                      # a run where the harness sees none of them
pi --agents-dir ~/work/.agents      # a different set, e.g. per-team

Read-only is not a convenience limit. A skill is an instruction the model follows later, on the host, outside the sandbox — so a sandbox that could edit ~/.agents had a persistence path out of itself, planted as a helpful-looking SKILL.md. There is no flag to make it writable, and the ro holds even with --insecure-keep-root.

A project's own .agents/ needs no setting at all: it is inside the project mount, and harnesses that read both (pi reads ~/.agents/skills and .agents/skills) see the project's next to yours. What that costs in exposure is spelled out in docs/threat-model.md: whatever is in those files — including any token someone pasted into a skill — is now readable by the agent you are sandboxing. Run --no-agents if a set of skills is not one you would hand to a stranger.

Pi inside the sandbox

The image has pi baked in, and its home is a volume, so out of the box the sandboxed pi has no provider and no preferences. Rather than making you configure it twice, ai-sandbox bridges your host ~/.pi/agent/{models,settings}.json -- rewriting them, because copying them would ship your credential into a container whose entire purpose is to be unable to use it:

your host config	what the sandbox gets
apiKey: "sk-..." or apiKey: "!pass show ..."	apiKey: "$AI_API_TOKEN" -- pi resolves it itself, no value is written to disk
a header like X-Trace: "!date +%s"	dropped, and reported: a host command is not run in the sandbox
baseUrl: "https://aipi.reeseapps.com/v1"	unchanged, along with every model, limit and thinking level
settings: provider, model, theme, thinking budgets, telemetry	unchanged, merged per key
a settings value pointing into ~	dropped, and reported: that path is not mounted
auth.json, models-store.json, sessions, extensions	never touched

The rewritten pair is seeded into /home/sandbox/.pi/agent at container start, owned by your uid, mode 600. The two files are managed differently, because they are different kinds of thing:

file	rule	why
models.json	replaced when missing or unchanged, left alone once you edit it in the sandbox	it is a catalog; a partial merge of a provider list is worse than either choice
settings.json	merged per key	pi rewrites its own settings — a theme change, "seen the changelog" — and managing the file whole meant that one write froze every host preference out of the sandbox permanently

So in settings.json: a key you change inside the sandbox stays yours, a key you never touched follows the host again, a key the host drops disappears unless you had changed it, and a key pi added (lastChangelogVersion) survives.

Thinking is xhigh, and it is policy

defaultThinkingLevel is the one setting ai-sandbox insists on, because pi's own default is medium and a sandbox run is usually the kind of work you want reasoned about hard:

pi --allow aipi.reeseapps.com        # xhigh, whatever your host settings.json says
pi --thinking max ...                # this run only
pi --thinking off ...                # no reasoning, e.g. a mechanical refactor at scale
pi --thinking host ...               # stop forcing it; your settings.json decides

--thinking is a policy, not a preference: it is written over the merged settings on every run, so a run cannot set its own thinking level down for the runs after it. It is pi-only — codex, claude and anything you put after -- read their own config, and ai-sandbox leaves that alone. --no-pi-config declines the policy too, since "pi starts with its own defaults" is what that flag is for.

The token itself follows the normal rule -- it is the one variable the bridged config names, so it is passed by name (never by value, and never into ps):

pi --allow aipi.reeseapps.com            # AI_API_TOKEN is picked up and announced
pi --no-pi-token-env ...                 # don't pass it; the run then fails closed
pi --pi-token-env GW_TOKEN ...           # a different variable name
pi --no-pi-config --shell                # no bridge at all

Refusing rather than launching a pi with no credential is deliberate -- the alternative is an agent that silently cannot reach its model.

Image

Both images come from one base, docker.io/library/ubuntu:24.04 — including the gate, which is squid from Ubuntu's apt rather than a third-party image. Built once per distinct policy and tagged by content hash: git, curl, jq, build tools, python3, Node 22 (tarball verified against nodejs.org SHASUMS256.txt) and the pi CLI. --install PKG adds packages and produces a different tag, so variants coexist, and --base-image swaps the base.

The base is pinned by digest

A tag is a moving pointer: FROM ubuntu:24.04 means "whatever Canonical points at this week", which makes a build unreproducible and a security patch invisible. So the tag is resolved to a digest, recorded under ~/.cache/ai-sandbox/base/, and the digest goes into the FROM line and into the content hash — refreshing the base produces a new image instead of mutating one a sandbox may be running right now.

ai-sandbox update --check     # what upstream has, what you are pinned to, nothing built
ai-sandbox update             # re-resolve, rebuild sandbox + gate, keep the old image
ai-sandbox build --refresh-base

Freshness is a TTL: every AISB_BASE_TTL_DAYS (default 7) the digest is re-resolved, which is a manifest read (skopeo), not a download. Offline, or with AISB_BASE_AUTO_UPDATE=0, the recorded digest is kept — a stale pin beats a mutable one, and AISB_BASE_TTL_DAYS=-1 pins until you say otherwise. A build with no digest is announced as UNPINNED, labelled that way, and warned about by doctor. Preview modes (--dry-run, --print-*) never touch the network, so they stay scriptable and report the pin you already have; inspect is the one that asks the machine now.

What an image admits to being

podman image inspect on any ai-sandbox image shows aisb.base-digest, aisb.node, aisb.pi-package, aisb.pi-version (plus aisb.role=gate for the gate), and /etc/aisb/manifest.txt inside it lists every resolved apt package version — the answer to "which squid was doing the filtering" once the build log is long gone. The previous image is always kept after an update, so rollback is a tag away.

Inspecting a run

ai-sandbox --dry-run --allow example.com --env KEY    # exact podman invocation, secrets redacted
ai-sandbox --print-policy                             # effective policy as key=value lines
ai-sandbox --print-entrypoint                         # the enforcement script baked into the image
ai-sandbox doctor [--deep]                            # host checks; --deep tests egress for real
ai-sandbox ps | clean [--all] [--volumes[=GLOB]] | inspect | gate up|status|logs|down

State lives under ~/.local/state/ai-sandbox (session dirs, log of every run with no secret values, build logs). Config in ~/.config/ai-sandbox.

/home/sandbox is a named volume (aisb-home) so npm and tool caches survive between runs. It outlives the image, so after changing the image you may be looking at a stale home: ai-sandbox clean --volumes reclaims unused ones. It only deletes volumes in the aisb- namespace that no container references, and --volumes=GLOB narrows it to one name.

Why it is hard to undo from inside

The container starts as root with CAP_NET_ADMIN, the entrypoint installs the egress ruleset, and then it execs your command through

setpriv --reuid=1000 --regid=1000 --clear-groups --inh-caps=-all --bounding-set=-all --no-new-privs

No root, no capabilities, no-new-privileges. Flushing the ruleset needs CAP_NET_ADMIN; unshare -n needs CAP_SYS_ADMIN; setuid helpers cannot hand either back. Both are asserted against a real container in tests/e2e/11_net.sh. Details and the measurements behind the design: docs/networking.md.

Read docs/threat-model.md — it says plainly what this does not protect you from (allowlisted destinations, container escapes, anything you explicitly mounted).

An unattended pipeline: ai-sandbox-autophase

A phase pipeline (.agents/phases/todo/…) is the harshest thing to ask of a sandbox: hours of unattended work by an agent that installs dependencies, runs the tests and wants to commit and push. It gets its own driver, and the split inside that driver is the design:

in the sandbox	on the host
the executor agent of every task, and the phase's final pass	the phase commit
.agents/validate.sh, the pass/fail gate	the push — and the re-push of anything an earlier run left behind

The git remote stays unreachable from inside on purpose: an agent that could push could ship what you are about to review. git status, diff, log and blame work in there (the repository is the mount, git is in the image), and the agent is told in its own prompt that push/fetch/pull cannot work, so it does not spend attempts discovering that. Delivery is the driver's job: it pushes when a phase commits, again before the next phase starts, and again at the end of the run, so nothing sits unpushed between runs. A push failure stops the run and prints the command that finishes it — the commit is local and safe, and the next run takes it out.

The gate runs in the sandbox for the same reason the agent does: .agents/validate.sh sits in the writable mount, so running it on the host is running agent-editable bash on the host. --validate-in host exists for a project whose toolchain only exists on the machine, and says so on every run.

[2026-10-28 02:11:03] preflight: one sandboxed shell, to prove the launcher, the image and the policy
  aisb-ok aisb=1 git=2.43.0 user=sandbox cwd=/home/you/Projects/foo
[2026-10-28 02:11:04] checking for commits a previous run left unpushed
  (up to date with origin/main)
[2026-10-28 02:11:04] ━━ 01_one/01_task.md — attempt 1/3 ━━
  (every line prefixed with the time, as the pipeline prints them)
  (committed: 9f1c2ab phase: 01_one)
  (pushed: 9f1c2ab phase: 01_one)
[2026-10-28 02:41:52] ✓ pipeline done — 2 unit(s) delivered this run: 01_one/01_task.md 01_one/00_phase.md

It runs your command, not a copy of it

The driver does not know your allowlist and does not try to. It runs your launcher — by default the sandbox-pi shell function you keep in ~/.bashrc.d:

sandbox-pi () {
    export AI_API_TOKEN=$(pass aipi/dev-token)
    ai-sandbox --allow aipi.reeseapps.com --allow github.com \
               --allow pi.dev --allow release-assets.githubusercontent.com \
               --env AI_API_TOKEN "$@"
}

That last "$@" is the whole contract. A function launcher is called with the run: the flags, then --, then the command to run inside, and one that ends at --env AI_API_TOKEN drops all of it. It then starts the image's default harness with no prompt — and a pi with no prompt on the other side of a launcher that attached your terminal sits waiting for keystrokes while its output goes into a pipe nobody is watching. The driver names a launcher like that in --status and refuses to start the pipeline; aliases and whole command lines do not need it, because the shell appends the words to those by itself.

A script cannot see an interactive shell's functions, so the driver looks the name up in ~/.bashrc.d/*, ~/.bashrc, ~/.bash_aliases and ~/.profile (or AISB_AP_RC_FILES) and takes the definition — probing never calls it, so a preview cannot fetch a token. Reading that definition is done in a throwaway session of its own, with no stdin and a bounded read: an interactive bash started from your terminal takes your terminal, gets stopped for it, and a stopped process ignores the timeout sent to it, so the probe used to be able to wedge the whole script before it printed a single line.

Each child then runs exactly what your shell would have, pass included, once per unit so an overnight run is not holding yesterday's token. An alias or a plain program works too; --pi-cmd also accepts a whole command line.

cd your/project
ai-sandbox-autophase --status     # launcher, pending units, git plan. Runs nothing.
ai-sandbox-autophase --dry-run    # the exact child and gate commands
ai-sandbox-autophase              # every pending task, in pipeline order

Every child is started as <your command> --no-tty --no-interactive -- pi <args> <prompt>, with its stdin on /dev/null: your flags first, then the two that keep an unattended run off your terminal (no TTY to corrupt the JSON stream, no stdin for your keystrokes to fall into), then --, after which ai-sandbox reads nothing but the command. Those two are flags, and a launcher is free to lose them, so the /dev/null is what actually guarantees the second one.

flag
--pi-cmd NAME	the launcher: function, alias, program or whole command line
--rc-files A:B	which rc files to look in
--sandbox-args ARGS	flags added to every run (default --no-tty --no-interactive)
--validate-in sandbox|host	where the gate runs
--no-push	commit locally, deliver nothing
--no-preflight	skip the throwaway container that proves the chain works before spending tokens
--trust-git-changes	deliver even though the sandbox edited .git/config or .git/hooks/
--unit, --phase, --max-units	scope a run

It does all the pushing — including the pipeline library's own — so that one function answers for every commit that leaves the machine. And because a repository is not something git merely records but something it runs, host-side git is started with an empty hooks path, no pager, no system config and no ext:: transport, and the run refuses to deliver if a sandboxed unit rewrote .git/config or planted a hook. A pre-commit written inside the sandbox is the sandbox executing code on the host, and the phase commit is not worth finding that out the hard way.

MAX_FIX_ATTEMPTS, PHASE_MODEL, PHASE_THINKING, PI_TRUST, FRESH_FIX, QUIET, PHASE_NOTIFY and PHASE_COMMIT behave as the phased-execution skill documents them — the driver reuses that pipeline and replaces only the parts that assume the agent is a local process.

Project layout

bin/ai-sandbox            the application: one script, ~2.2k lines
bin/ai-sandbox-autophase  the pipeline driver: agent in the sandbox, git on the host
install.sh                symlink both into ~/.local/bin + optional alias
contrib/ai-sandbox.bash   bash completion for both
config/                   example config, allowlist and env file
docs/                     networking.md, threat-model.md, recipes.md
tests/run.sh              runner; tests/unit/* (no containers) tests/e2e/* (real ones)
tests/fixtures/phase-lib  contract fixture: the pipeline library's shape, without skill
                          or containers, for the driver's tests
tests/spike_network.sh    re-measure podman's behaviour on a new host
Makefile                  make test | e2e | lint | check | doctor | build

Development

make test     # unit: args, config, argv construction, env handling, rules, pipeline driver
make e2e      # + real containers: build cache, egress blocking, capability drop, lifecycle
make lint     # shellcheck in a container
make style    # bash -n

The Containerfile, the container entrypoint and the squid config are all generated from bin/ai-sandbox, so the script remains the single source of truth (--print-containerfile, --print-entrypoint).