Maxim Bifrost manages LLM routing, caching, and MCP tools. Willow governs employee AI with shadow AI discovery, managed connectors, resource checks, and runtime approvals.
A security team rolls out Claude Code and Cursor to its engineers. Within weeks, laptops carry MCP servers nobody approved, skills copied from public repositories and instruction files that change how agents behave. The team needs to see what is installed, decide which tools each group may use, and make sure risky actions wait for a person. A fast model router helps engineering. It does not answer those questions on its own.
Willow is built for that job. It gives organizations central control over how employees and background agents use AI tools, MCP servers, browser AI and desktop coding assistants. It scans laptops for local instruction files, checks current resource data such as file labels before tools run, provides ready-to-use human approval workflows, flags unused tool permissions and helps admins turn shadow MCP servers into managed integrations.
Maxim's Bifrost is an open-source Go LLM gateway paired with commercial Enterprise controls and Bifrost Edge, an endpoint agent in alpha. It focuses on fast routing across 8+ providers and 1000+ AI models response caching and code-based tool bundling. Bifrost includes enterprise identity, payload guardrails and alpha endpoint discovery, but it leaves per-call human approval screens to host apps and its Enterprise guardrails do not inspect MCP prompt or resource reads. Both products put a gateway in front of MCP servers and both enforce budgets. The difference is who each one is built for. Bifrost serves engineering teams that want routing speed and developer tooling. Willow serves the IT and Security teams who have to approve, inspect and audit what employees' AI tools do.
The rows that matter most to Security teams are approvals and resource checks. In Bifrost's MCP gateway mode, the host application controls per-call confirmation, so custom hosts must build that logic themselves. Willow ships the approval workflow and can block a call based on the current state of the resource it touches, not only on the arguments in the request.
Deeper endpoint file discovery
Both platforms deploy through MDM tools such as Jamf, Intune and JumpCloud and monitor desktop AI apps, coding assistants, MCP servers and browser AI. Bifrost Edge is in alpha with request access, and it separates traffic routing from MCP config discovery. On laptops running the Willow Scan Agent, Willow runs daily full scans and 5-minute shallow scans to inventory local skill files (SKILL.md), agent instruction files (CLAUDE.md, CLAUDE.local.md, AGENTS.md) and plugins.
Willow also flags skill instructions that try to alter persistent agent files such as MEMORY.md or SOUL.md as an approximate risk indicator. WSL MCP servers are logged without enforcement, and admins still need to check files modified before a bad skill was blocked. Bifrost does not scan for these local instruction files, though it offers a central Skills Repository for publishing SKILL.md packages.
Turnkey human approvals for sensitive actions
Both platforms can inspect and block individual MCP tool calls based on argument values. Bifrost Enterprise uses CEL rules linked to regex, secret scanning and LLM judges to block or redact calls, and its Edge dashboard approves or denies whole apps and MCP servers. In gateway mode, per-call confirmation is left to the host: Claude Desktop or Cursor show their own prompts, and custom hosts must build their own.
Willow gives IT a packaged, centrally managed approval workflow out of the box. Admins require sign-off with tool-level Force Approval or with runtime guards that inspect tool inputs using JSONata checks. Users receive an approval link, or inline chat controls when MCP Apps support is enabled, and approved calls can be reused for the same user, tool and arguments within a set time window.
Guardrails that check the live resource
Willow's per-tool Conditions do more than read the arguments in a tool call. They can query the upstream app first with integration credentials, for example to check a Google Drive file's classification label or a Slack channel's identity, and block the call if it matches. Conditions fail closed, blocking if the lookup errors, and currently enforce Block only; conditional approvals run through input runtime guards. Bifrost's CEL rules select checks using tool, argument, identity and header fields.
Willow guards support built-in detectors, regex, JSONata, LLM checks and custom functions, filtered by client, agent, group, risk score and direction. When several guards trigger, a strict order applies: Block beats Require Approval, which beats Redact, which beats Warning. In either platform, blocking on tool output only hides the result from the model; it cannot undo a write that already ran upstream.
Guided shadow-to-catalog onboarding
When Willow discovers an unmanaged MCP server on a laptop, admins can open a prefilled setup form, review the configuration, name it and publish it to the company catalog. That turns shadow AI into a managed integration in a few steps. Admins can also promote employee-built User MCPs into organizational copies and assign them to groups. Bifrost focuses on developer tool delivery through Virtual MCP endpoints, three-level tool filtering, OpenAPI-to-MCP generation (beta) and Starlark orchestration (Code Mode).
Maintained connectors, not directory listings
A raw catalog count hides how much work a team still has to do. Willow builds and maintains its own application connectors with curated tools, service-specific auth setup and tool risk ratings, and it clearly separates third-party registry listings it has not vetted. Bifrost connects to upstream HTTP, SSE and STDIO MCP servers, and its directory lists upstream servers; a listing shows availability, not a connector Maxim maintains.
Willow's built-in connectors support Instant OAuth, org-owned OAuth, shared API keys or per-user API keys depending on the service. In Bifrost, per-user MCP authentication and headers work only over HTTP, so SSE and STDIO servers cannot use its per-user credential model. For REST APIs, Willow lets admins define or AI-generate tools and map arguments with JSONata, while Bifrost's beta OpenAPI importer generates tools from specs with static credentials or per-user headers.
Actionable least-privilege reporting
Both vendors integrate with Okta and Entra ID through SSO and SCIM, support per-user OAuth for MCP tools and govern background agents. To clean up excess permissions, Willow Radar flags enabled tools with zero recorded usage over the past 30 days, and highlights enabled destructive or high-risk tools that lack Force Approval or a blocking or approval condition. It turns least privilege for AI agents into a to-do list.
Audit built for security reviews
Willow's Logs include a visual Tool Call Flow covering authentication, guard checks, execution and outcome, clearly marking Local MCP bypasses, with configurable payload retention. The catalog supports draft-to-published states, and the optional Auto Sync Tools setting keeps previously disabled tools off when an upstream server syncs. Bifrost provides MCP execution metrics and logs with optional payload capture, plus Enterprise admin audit trails.
When Maxim (Bifrost) is the right choice
Bifrost fits engineering teams whose first requirement is the model traffic layer:
These are strong developer capabilities. They pair well with a governance layer that gives Security the approvals, resource checks and catalog control that a router leaves to the host app.
Plans, limits and coverage boundaries
Both vendors place advanced MCP governance in higher tiers. Bifrost's access profiles and payload guardrails require Enterprise. Willow places API wrapping, per-tool permissions, Conditions, runtime approvals and full audit streams in its Enterprise plans; Willow's pricing page lists a $0 Free plan and Enterprise priced per human seat, billed annually.
Neither platform guarantees zero budget overshoot. Bifrost updates spend after an active request finishes, and Willow caches spend totals for about a minute. Both have fail-open cases: if Willow's built-in detector is offline, prompt-injection and person-name checks score zero; if a Bifrost guardrail provider times out, that profile lets processing continue. Willow's Local MCP option runs STDIO servers on the laptop, where group access and disabled-tool blocks still apply but gateway rate limits, Conditions and runtime guards are bypassed.
Should you run both?
Some organizations will keep Bifrost as a fast model router for engineering workloads. Willow then governs the employee AI layer: what is installed on each laptop, which approved tools each group receives, which calls need a person's sign-off and which tools have not been used in 30 days. That split gives each product a clear job.
Where both sit in front of MCP servers, route employee AI clients through one gateway to avoid duplicate policy. For Security-led rollouts, Willow's built-in approvals, resource-aware Conditions and catalog workflows make it the natural owner of the MCP path.
Bifrost gives engineering teams a fast, flexible model router. Willow gives IT and Security the controls to govern employee AI: local file discovery, turnkey approvals, live resource checks and a curated catalog. To see Willow on your own tools, book a demo or start on the Free plan.
Willow is built for IT and Security teams to govern how employees and agents use AI tools, with endpoint discovery, built-in approvals and resource-aware guardrails. Bifrost is built primarily for engineering teams that want fast LLM routing, caching and developer MCP tooling.
In Bifrost's MCP gateway mode, per-call confirmation is left to the host application. Clients such as Claude Desktop or Cursor show their own prompts, and custom hosts must build that logic. The Bifrost Edge dashboard approves or denies whole apps and MCP servers. Willow ships a centrally managed per-call approval workflow.
Both deploy through MDM and monitor desktop AI apps, coding assistants and MCP servers. Bifrost Edge is in alpha with request access. Willow's Scan Agent runs daily full scans and 5-minute shallow scans and inventories local SKILL.md, CLAUDE.md, CLAUDE.local.md and AGENTS.md files.
Bifrost. It offers 8+ providers and 1000+ AI models native adapters, cross-API translation, failover chains and response caching. Willow's LLM Gateway, in beta, supports five core provider types plus OpenAI-compatible services and does not translate across API formats.
Bifrost serves MCP prompts and resources alongside tools, but its virtual-key tool allowlists do not restrict them, and its Enterprise guardrails do not inspect the content of prompt or resource reads.
Give them a Basecamp. Go from AI chaos to AI work, in minutes.