Kontext Cli

Kontext is runtime security for an AI workforce. The daemon runs inside the agent's own environment, at pre-action control points, and decides allow-or-deny before the tool call executes. That means an `rm -rf` against a protected path never reaches the shell, a force push to a protected branch waits for a human yes, and a credential file read gets flagged and captured with redaction applied on the machine. Install is one command. The self-serve quickstart is macOS and Homebrew: `brew install kontext-security/tap/kontext`, grab a personal API key from the dashboard get-started page, run `kontext setup`, then `kontext doctor` to verify. The same flow asks you to trust Kontext hooks in Codex when prompted and to send a read-only test trace. There is no code change and no gateway sitting in the request path. The pitch is millisecond decisions without a proxy. Kontext splits two things most tools blur: risk detection flags activity for investigation, while policy enforcement blocks it — so a flagged call can be investigated rather than silently stopped. Deterministic policy handles hard boundaries, and two local models do the judgment work. Kestrel screens agent tool calls in roughly 22 microseconds and, per Kontext's own benchmark write-up, beats seven frontier LLMs by about 0.40 F1 on ShellRisk-Bench. Merlin uses the user request, tool history, and tool schemas to decide whether a call is unsafe before it runs. Policies layer across org, group, user, agent, repo, and branch; payload capture is configurable as omitted, summary, or full. Deployments start in observe mode, recording the decision each tool call would get without blocking anything, which backtests policy against real traffic. When the would-deny log matches intent, you flip to enforce. Discovery covers coding assistants, workplace agents, and custom-built agents across endpoints and cloud environments, all streaming sessions, decisions, and devices into one console with an exportable audit trail.