cd /news/ai-safety/show-hn-grith-syscall-level-supervis… · home topics ai-safety article
[ARTICLE · art-111878] src=github.com ↗ pub= topic=ai-safety verified=true sentiment=· neutral

Show HN: Grith – syscall-level supervision for AI coding agents on Linux

Grith, a new open-source syscall-level security supervisor for AI coding agents on Linux, intercepts every syscall an agent makes and scores it before execution, denying risky operations such as a Claude Code attempt to POST .env to an external host. The tool, available for Linux x86_64 and aarch64, offers eleven built-in profiles for tools like Claude Code, Codex, and Cursor, and includes supervision-escape enforcement by default as of v0.2.5.

read3 min views2 publishedAug 26, 2026
Show HN: Grith – syscall-level supervision for AI coding agents on Linux
Image: Michielbdejong (auto-discovered)

grith is an OS-level security supervisor for AI coding agents. It intercepts every syscall your agent makes and decides what actually runs.

Claude Code ships the feature, then tries to POST .env

to an outside host. grith denies it at the kernel boundary.

grith.ai · Documentation · Security model

Install on Linux (x86_64 or arm64):

curl -fsSL https://grith.ai/install | sh

Wrap the agent you already use:

grith exec -- claude-code "fix the failing test"

Or run grith's own agent, with the same filters in front of every tool call:

grith run "list every TODO in this repo"

Every file read, shell command, network call, and process spawn is scored before the kernel executes it:

Score Verdict What happens
under 3.0 allow
the call proceeds
3.0 to 8.0 queue
the process freezes until you approve or deny it
over 8.0 deny
the call never runs

Nothing runs on a maybe. Eleven built-in profiles (claude-code, codex, aider, cursor, cline, copilot, goose and others) auto-allow each tool's routine work, so the queue only sees the calls worth your attention.

Supported platforms, other install methods, and building from source #

Platform Architecture Status
Linux x86_64 supported (kernel 4.8+)
Linux aarch64 supported (kernel 5.3+)
macOS Apple Silicon / Intel v2.0 - needs an Endpoint Security backend
Windows x86_64 v2.0 - needs an ETW backend

The installer auto-detects your platform, verifies the SHA-256 checksum, and installs to ~/.local/bin

. Pass --global

to install to /usr/local/bin

, or --version <version>

to pin a release:

curl -fsSL https://grith.ai/install | sh -s -- --global

You can also download a binary directly from the latest release, or build from source with Rust 1.88+ and Node 22+:

git clone https://github.com/grith-ai/grith.git && cd grith && make dist

Full build instructions are in the documentation.

Every release ships a static musl binary with a SHA-256 checksum, a cosign keyless signature, a CycloneDX SBOM (itself signed), and SLSA build provenance. If cosign

is on your PATH, the installer verifies the signature against the release workflow's identity automatically - no flags needed. To verify by hand, see release verification.

By default, nothing. The free tier runs entirely offline: no account, no telemetry, and the audit log stays in local SQLite. Paid tiers validate their licence against grith.ai roughly once a day, and sync audit records only until you turn that off with general.audit_sync = false

(licence validation continues; air-gapped deployments disable it too).

Supervision-escape enforcement is on by default as of v0.2.5: spawning something that hands work to an unsupervised peer (systemd-run

, docker

, tmux

) reaches the review queue rather than running unseen. Non-interactive sessions have no one to ask, so they fail safe and deny - if a CI script legitimately delegates, permit the binary in its profile or set supervisor.enforce_authority_delegating_spawn = false

.

Getting started- installation, configuration, and the CLI referenceSecurity model- the filter pipeline, scoring, and what it does not coverSupervisor profiles- per-tool allowlists and how to write your ownCHANGELOG- including the bypass classes we have not closed yet

Pull requests are welcome - start with CONTRIBUTING.md, which covers the CLA and the local setup. grith is developed in a private monorepo and exported here per release, so this repository's history is one commit per export rather than per change; PRs are reviewed on GitHub and applied upstream with attribution.

Found a security issue? Please follow SECURITY.md rather than opening a public issue.

Repository code: MPL-2.0

Pro and Enterprise capabilities ship in the same binary and are unlocked by signed licenses. Hosted billing, license issuance, and cloud sync infrastructure are not part of this repository.

── more in #ai-safety 4 stories · sorted by recency
── more on @grith 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-grith-syscal…] indexed:0 read:3min 2026-08-26 ·