350+ Tools CoveredSource Data Updated Weeklydates

Strands Box

Box runs AI agents in an operating-system-enforced sandbox and applies Dogwood policy to their shell commands, file operations, network requests, and MCP calls. It is in preview, and local execution now covers Linux ARM64 alongside Apple-silicon Macs, so the first platform question for a pilot is which of the two hosts your agents will actually run on.

Visit Strands Box →
Type
AI Security
Category

By Egor Burlakov · Edited · Source dates below

Snapshot

Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.

AttributeDetail

Deployment

Local execution on Linux ARM64 and macOS with Apple silicon

License

Apache License 2.0

Policy model

Checked operations are denied by default; a matching permit rule allows them and a matching forbid rule overrides that permission

Enforcement

Seatbelt on macOS sets which files and programs the agent can reach, and each tool and local MCP server runs in its own sandbox

Linux package requirements

glibc 2.34 or newer and the GCC runtime library (libgcc_s.so.1)

Take

Choose Box when an agent needs OS-level containment together with policies that can weigh a request's context or earlier recorded actions. Preview status still applies, so pin your version. If the pilot targets Linux, check first that the release you pin ships a box-<version>-linux-arm64.tar.gz asset with its SHA256SUMS.txt entry, since a release can contain macOS assets only.

Evaluate Strands Box

Comparisons

Strengths

  • Dogwood rules can depend on the requested operation, its arguments, earlier actions, and elapsed time; a file read through one interpreter can affect whether a later network request is allowed.

  • The egress gateway authenticates permitted requests with configured API credentials or AWS SigV4 signing, and the agent does not receive the underlying secrets.

  • Box records policy decisions in OTLP JSON, with records going by default to <box_dir>/private/telemetry/records.jsonl.

  • strands_box::reload replaces a running box's policy from a host thread and waits until the change takes effect. Invalid policy returns ReloadRefused and leaves the current policy in place, policy files and the OS sandbox stay unchanged, and an empty list denies all policy requests.

Trade-offs

  • The directory where the downloader stages its version check must be writable and allow execution.

  • Direct filesystem grants in box.toml are enforced by the OS but do not produce individual Dogwood decisions. To apply policy to file operations, use the interpreters and keep those files out of the agent's direct grants.

  • Using Box as a Rust library means one Box workload per host process; later run calls fail, even if the first call failed.

  • A rule for an MCP tool can reload only after Box loads that server's tool definitions. Without them, even reloading the unchanged policy is refused, and if policy blocks discovery, waiting will not help.

  • The Strands CLI walkthrough in the getting-started guide uses macOS 15 or later on Apple silicon, Node.js 22.21 or later from Homebrew, and access to Claude Opus 5 on Amazon Bedrock in us-west-2.

  • The release downloader tests run on Linux and macOS, but they use local fixture archives and simulated HTTP clients and do not qualify a published binary or its containment behaviour.

Cost and licensing

Box is open source under Apache License 2.0 and runs on your own machine. You install it with the download script, which checks the release checksum and unpacks the binaries into ./box-core without changing your PATH or system directories, or you build the binaries from the Cargo workspace. The spend therefore sits in the host you run the box on and the model access its agent needs, such as the Bedrock access the Strands CLI walkthrough requires.

Public signals

About these signals

Verified factual signals from public sources. They indicate observable activity or interest, not total adoption, product quality, or cost.

203 GitHub commits 90d548 GitHub stars

See all signals from 1 source
Source
Signals
Last updated
GitHub
Commits 90d:203↑46Stars:548↑44
November 16, 2026

Alternatives to Strands Box

The reviewed substitutes for Strands Box among the AI security, and what would make each one the better answer.

Other approaches

A different approach to the same problem. Each substitutes only for the workload named beside it.

AgentVault
Strands Box confines an agent at the OS level, using default-deny policies on what it can execute, read, write and reach on the network, and it injects credentials so the agent never holds secrets. AgentVault sits as a proxy in front of agents like Claude, with a live dashboard, dangerous-command blocking, permission approvals and network monitoring; the choice comes down to hard isolation that stops actions up front versus visibility and approval of actions as they happen.Applies to: constraining what a local coding agent can do on a developer machine
EarlyCore
EarlyCore scans AI agents for prompt injection, data leakage and jailbreaks before release and then monitors them in production on Bedrock, Vertex AI or custom stacks. Strands Box doesn't detect attacks; it limits the damage a compromised agent can do by sandboxing its system and network access, so the choice depends on whether the buyer wants detection and monitoring or containment.Applies to: protecting agents against prompt-injection-driven data exfiltration
Explore all Strands Box alternatives →

Sources

This review is based on the Strands Box repository and the Strands Agents product site, observed on 2026-11-09.