Product · Safeguard Code

Safeguard Code. A local AI coding agent that already knows your supply chain.

A terminal-first AI agent that runs locally on your machine. It drives the editor, the shell, the build, and the test loop the same way a human engineer does — but it reasons about security as a first-class concern, not an afterthought. Reachability, policy, SBOM, license posture — all of it is in scope before the agent writes a line.

Local-first
Runs on your machine, not a tab
Mac · Win · Linux
Native binary, no runtime
Lino
On-device inference for the hot path
Griffin
Cloud-burst on the hard reasoning
What it does

Four jobs. One agent.

Drives the editor, the shell, and the build

Operates the developer's actual workflow, not a chat sidebar. Edits files in place, runs the build, watches tests, and fixes the regressions it just caused. The loop is the product — type, run, observe, repair — and the agent stays inside it.

Supply-chain aware by default

Every code change is evaluated against the project's policy, SBOM, and reachability graph before it gets committed. The agent knows which transitive packages are KEV, which sinks are reachable, and which license shifts your legal team will care about — without being prompted.

Local-first inference for the hot path

Lino runs on-device for inline findings and short reasoning. Network is only used when Griffin cloud-burst is genuinely needed — multi-hop reachability, cross-package exploit hypothesis. The default mode is offline-capable; the cloud is the optional escalation.

Auditable trace per session

Every action is logged with a structured trace: intent → tool call → result → fallback. The session is exportable as a single artifact for code review, regulatory audit, and reproducibility. The trace is not a debug log — it is the work product.

How it compares

Versus a generic AI coding agent.

General coding agents are tuned for breadth. Safeguard Code is tuned for the supply-chain-aware engineer.

CapabilitySafeguard CodeGeneric AI coding agent
Knows your SBOM
Reads your policy gates
Runs offline by default
Reasoning model is security-tuned
Auto-fix uses a cited trace
Audit log per session
Comparison against the median of widely-deployed general-purpose AI coding agents.
Install + invoke

Three commands to a working agent.

1. Install the agent binary globally. One command, every platform.

$ npm i -g @safeguard-sh/code

2. Initialise the workspace. Discovers the project, reads the policy, indexes the SBOM.

$ safeguard-code init

3. Launch the agent. It now drives your editor, shell, build, and tests in a single session.

$ safeguard-code
Workflows it handles well

The shape of work it does best.

These are the loops where supply-chain context turns a coding agent from a generalist into a specialist.

Fix every reachable CVE before the release branch cuts

Point the agent at a service. It enumerates reachable findings, ranks by SLA and CVSS, applies the lowest-risk patches that satisfy your version constraints, and runs the test suite after each. The release branch cuts clean.

Apply a library upgrade across N services

Hand it a target package and version, plus a list of repos. The agent fans the upgrade out, runs each service's test suite, captures the diff, and reports the survivors and the casualties — with a per-repo patch ready for review.

Triage a fresh disclosure

Drop in a CVE that just landed. The agent scopes blast radius across reachability and SBOM, drafts the upstream patch against the affected version, and prepares the disclosure thread with the trace attached. Hours, not days.

Onboard a new repo

Hand it a fresh repository. It reads the code, gets the build green, and surfaces the policy gaps in priority order — secrets in history, missing license headers, unsigned dependencies, reachable CVEs. A first-day report your security lead trusts.

Security posture

The agent that fails closed.

Tenant-level commitments

  • Weights signed and verified at install. The binary refuses to run an unsigned model file.
  • No source code or prompts leave the machine without explicit, logged network actions.
  • Air-gapped mode supported. Lino runs locally; Griffin escalations queue and fail closed.
  • Opt-in anonymised telemetry only. The default install ships zero telemetry.
  • Per-session trace stored locally first; export is an explicit user action.
  • Model weights pinned by SHA; provenance attestation available under NDA.

Run the agent against your repo.

Local-first, supply-chain-aware, audit-trace by default. The coding agent that fits the way security engineers already work.