HEAD-TO-HEAD · FACTS ONLY
Bay vs Koi: AI endpoint control facts side by side
Short answer
Bay describes 1 of the four product control types on the pages we read and Koi describes 1 and their control positions match. On decisions, Bay decides on a tool call before it runs and Koi describes some control over tool calls. We do not score either; the right choice depends on the AI tools people use, your devices, your current stack and whether you block or guide.
By The Device Desk · Endpoint AI Finder · Published 2026-10-11
Desk research from public vendor pages read on 27 September and 11 October 2026. Profiles follow our published method; we did not run the products.
Vendor pages read September and October 2026 · updated 2026-10-13
Where do their controls sit?
- Filled or hatched marker: Bay
- Dashed teal ring: Koi, described
- DEV Filled marker: described on the product pages we read
- DEV Hatched marker: partly described (another component, a narrower scope, or method not detailed)
- DEV Dashed marker: not in our sources (it may exist; we did not find it)
What are the facts, side by side?
| Fact | Bay | Koi |
|---|---|---|
| Owner | None | Palo Alto Networks (acquisition completed 14 April 2026) |
| Positioning | An endpoint-first control layer for AI agents, coding assistants and MCP servers on employee devices. Bay describes itself as a fully preventive, agentless platform that classifies, governs and prevents risks across every AI agent and tool on enterprise endpoints. | Koi, now sold by Palo Alto Networks as Cortex Agentic Endpoint Security, gives visibility into models, MCPs and extensions on endpoints, removes insecure packages, extensions and AI tools in one click, and stops malicious actions. |
| How it reaches devices | Deploys through your existing EDR or MDM, with no new agents and no new tools; Bay says it scans in minutes and exits cleanly. Source: Bay homepage · Read 27 Sep 2026 | Available as a Cortex XDR module or as a standalone agent. koi.ai now redirects to the Palo Alto Networks page. Source: Palo Alto Networks: Cortex Agentic Endpoint Security · Read 27 Sep 2026 |
| On-device control | DescribedInventories AI agents, MCP servers, tools, credentials and extensions on the device, and its Enforcer returns Allow, Ask or Deny locally on prompts, pre-tool calls and post-tool responses on supported agent surfaces. Delivered as Wave, an ephemeral binary that runs through your existing MDM or EDR. Policies can be scoped by agent, group, endpoint or device. Source: Bay blog: Ghostjacking (re-read) · Read 11 Oct 2026 | DescribedAgent-based visibility into models, MCPs and extensions, one-click removal of insecure packages, extensions and AI tools, and blocking of unsafe AI application interactions. Source: Palo Alto Networks: Cortex Agentic Endpoint Security · Read 27 Sep 2026 |
| Browser extension | Not in our sourcesNo browser extension is described. Bay's Application Control can allow or block browsers and extensions as applications, which is a device rule, not a control on what goes into browser AI. Source: Bay blog: Ghostjacking (re-read) · Read 11 Oct 2026 | Not in our sourcesNot described on the pages we read. |
| Network or SSE proxy | Not in our sourcesNot described on the pages we read. | Not in our sourcesNot described on the pages we read. |
| AI or MCP gateway | Not in our sourcesNot described on the pages we read. | Not in our sourcesNot described on the pages we read. |
| Names the coding agents it covers | MatchesManaged settings for Claude Code, Codex and Claude Desktop lock permission rules, preserve managed hooks, restrict MCP servers and constrain plugin sources. Source: Bay blog: Ghostjacking, an AI agent attack (August 2026) · Read 27 Sep 2026 | MatchesA Palo Alto Networks blog of 18 August 2026 names Claude Code, Cursor, Codex and Antigravity. Source: Palo Alto Networks blog: A Quarter of Innovation with Koi Agentic Endpoint Security · Read 11 Oct 2026 |
| Decides on a tool call before it runsDiffers | MatchesEvaluates prompts, pre-tool calls and post-tool responses and returns Allow, Ask or Deny locally, which Bay states takes under 4ms with no cloud round-trip, on supported agent surfaces. Source: Bay blog: Ghostjacking (re-read) · Read 11 Oct 2026 | PartlyBlocks unsafe AI application interactions and prevents prompt injections; decisions on individual tool calls are not described. Source: Palo Alto Networks: Cortex Agentic Endpoint Security · Read 27 Sep 2026 |
| Controls what goes into browser AI | Not in our sources | Not in our sources |
| Rolls out through EDR or device management you runDiffers | MatchesDeploys through your existing EDR or MDM; no new agents. Source: Bay homepage · Read 27 Sep 2026 | PartlyA Cortex XDR module, or its own standalone agent. Source: Palo Alto Networks: Cortex Agentic Endpoint Security · Read 27 Sep 2026 |
| Can warn or ask instead of only blockingDiffers | MatchesAsk is one of three answers, with Allow and Deny. Source: Bay homepage · Read 27 Sep 2026 | Not in our sources |
| Can block an agent, tool or action | MatchesDeny stops an action. Rules cover MCP servers (an allow or block list, and per-tool rules for read, create, update, delete and execute), applications, CLIs, browsers, IDEs and extensions, credentials and files, network egress allow lists, and capability rules such as shell execution, package installation and cloud CLIs. Source: Bay blog: Ghostjacking (re-read) · Read 11 Oct 2026 | MatchesStops malicious actions and removes insecure packages, extensions and AI tools in one click. Source: Palo Alto Networks: Cortex Agentic Endpoint Security · Read 27 Sep 2026 |
| Has a watch-only or simulation modeDiffers | MatchesSimulation Mode runs a rule without enforcing it. Source: Bay blog: Ghostjacking, an AI agent attack (August 2026) · Read 27 Sep 2026 | Not in our sources |
| Records who or what took each actionDiffers | MatchesRecords who triggered an action, whether a human was involved, and the chain from prompt to system action. Source: Bay homepage · Read 27 Sep 2026 | Not in our sources |
| Also covers SaaS, cloud or homegrown AIDiffers | Not in our sources | PartlyPalo Alto Networks plans integration with Prisma AIRS, its wider AI security platform. Source: Palo Alto Networks press release: acquisition of Koi completed · Read 27 Sep 2026 |
| Has a public price or tiers (vendor site or marketplace) | Not in our sources | Not in our sources |
| Built into | None | Offered as a Cortex XDR module. Source: Palo Alto Networks press release: acquisition of Koi completed · Read 27 Sep 2026 |
| Works with | None in our sources | None in our sources |
| Price | Not published. Bay's site has a demo request and no pricing page (re-checked 11 October 2026). Source: Bay homepage (re-read) · Read 11 Oct 2026 | Not published. Source: Palo Alto Networks: Cortex Agentic Endpoint Security · Read 27 Sep 2026 |
Where do they differ?
- Decides on a tool call before it runs: Bay matches; Koi partly matches.
- Rolls out through EDR or device management you run: Bay matches; Koi partly matches.
- Can warn or ask instead of only blocking: Bay matches; Koi is not in our sources.
- Has a watch-only or simulation mode: Bay matches; Koi is not in our sources.
- Records who or what took each action: Bay matches; Koi is not in our sources.
- Also covers SaaS, cloud or homegrown AI: Bay is not in our sources; Koi partly matches.
Choose Bay if
- Your developers run coding agents with many MCP servers and you want a decision on each action before it runs
- You want to add AI agent control through the EDR or MDM you already run instead of rolling out another agent
Choose Koi if
- You already run Palo Alto Networks Cortex XDR
- You want to find and remove risky extensions and packages across the fleet
Which limits should you check?
Bay
- Endpoint only: no browser extension, network proxy or AI gateway is described, and nothing on SaaS or cloud-hosted AI apps.
- One of the smallest public footprints here: a single homepage, two blog posts, no documentation portal and no published pricing (re-checked 11 October 2026).
- Headline figures such as under 5 minutes to deploy, under 1% false positives and 100% of shadow AI discovered are Bay's own statements and have not been verified.
Koi
- Agent-based: it adds an agent unless you run it as a Cortex XDR module.
- Warn or ask actions and a record that separates a person from an agent are not described on the pages we read, and no pricing is published.