diff --git a/architecture/ai-pipeline.mdx b/architecture/ai-pipeline.mdx
index f284263..243b2c9 100644
--- a/architecture/ai-pipeline.mdx
+++ b/architecture/ai-pipeline.mdx
@@ -59,13 +59,11 @@ The routing is opinionated, not magic: clear rules in `~/CLAUDE.md` and `AGENTS.
 
 ## Local AI gateway (Bifrost)
 
-Bifrost is the OpenAI-compatible HTTP gateway that sits between every AI tool on the workstation and whichever provider eventually answers the call. It exposes `http://localhost:30080/v1/chat/completions` and fans out to OpenAI, Gemini, OpenRouter, and the local MLX server based on the task class.
+[Bifrost](/tools/bifrost) is the OpenAI-compatible HTTP gateway that sits between every AI tool on the workstation and whichever provider eventually answers the call. It exposes `http://localhost:30080/v1/chat/completions` and fans out to OpenAI, Gemini, OpenRouter, and the local MLX server based on the task class.
 
-- **Never hardcode model identifiers in committed config.** Models change weekly; identifiers rot. Tools resolve task classes (Research, Coding, Review, Pre-commit) to a current model at call time via `listmodels`.
-- **Local MLX models carry an `mlx-local/` prefix** when called through Bifrost (it expects `provider/model` format). Calling the vllm-mlx server directly on port 11434 uses the bare HuggingFace model ID — no prefix.
-- **Cloud models go in unprefixed.** Bifrost handles routing; do not add a provider prefix yourself.
+Tools resolve task classes (Research, Coding, Review, Pre-commit) to a current model at call time — never hardcode model identifiers in committed config. When `localOnlyMode` is enabled, every request routes exclusively to the local MLX inference server on port 11434.
 
-When `localOnlyMode` is enabled (or the `--local` flag is passed), every task routes to the MLX inference server on port 11434 and no cloud API calls happen. Verify the LaunchAgent is running before invoking: `launchctl list | grep vllm-mlx`.
+See [Bifrost](/tools/bifrost) for routing conventions, local-only mode details, and provider capabilities.
 
 ### PAL MCP
 
diff --git a/docs.json b/docs.json
index 5b4addd..171e666 100644
--- a/docs.json
+++ b/docs.json
@@ -158,7 +158,8 @@
             "pages": [
               "tools/overview",
               "tools/mlx-benchmarks",
-              "tools/automation"
+              "tools/automation",
+              "tools/bifrost"
             ]
           },
           {
diff --git a/tools/bifrost.mdx b/tools/bifrost.mdx
new file mode 100644
index 0000000..673f44c
--- /dev/null
+++ b/tools/bifrost.mdx
@@ -0,0 +1,68 @@
+---
+title: "Bifrost AI gateway"
+description: "OpenAI-compatible HTTP gateway that routes AI requests from local tools to the right provider — OpenAI, Gemini, OpenRouter, or local MLX inference."
+tier: 2
+---
+
+> One endpoint for every AI tool. Bifrost handles the routing.
+
+Bifrost is the OpenAI-compatible HTTP gateway that sits between every AI tool on the workstation and whichever provider eventually answers the call. It exposes `http://localhost:30080/v1/chat/completions` and fans out to OpenAI, Gemini, OpenRouter, and the local MLX server based on the task class.
+
+- **GitHub:** https://github.com/maximhq/bifrost
+- **Homepage:** https://www.getmaxim.ai/bifrost
+
+## Model routing conventions
+
+Never hardcode model identifiers in committed config. Models change frequently; identifiers rot. Tools resolve task classes (Research, Coding, Review, Pre-commit) to a current model at call time via `listmodels`.
+
+| Context | Format |
+| --- | --- |
+| Local MLX models through Bifrost | `mlx-local/<model>` — Bifrost expects `provider/model` |
+| Direct vllm-mlx on port 11434 | bare HuggingFace model ID — no prefix |
+| Cloud models through Bifrost | unprefixed — Bifrost routes by task class |
+
+## Local-only mode
+
+When `localOnlyMode` is enabled or the `--local` flag is passed, every request routes to the MLX inference server on port 11434. No cloud API calls occur.
+
+Verify the LaunchAgent is running before enabling local-only mode:
+
+```bash
+launchctl list | grep vllm-mlx
+```
+
+## Priority in the AI gateway stack
+
+Bifrost is the second layer in the gateway priority order:
+
+1. **Anthropic official** — Claude Code plugins, skills, patterns
+2. **Bifrost AI gateway** — multi-provider routing at `localhost:30080`
+3. **PAL MCP** — only for `clink` (parallel calls) and `consensus` (multi-model agreement)
+4. **Personal or custom** — only when no alternative exists
+
+## Capabilities
+
+Bifrost supports 23+ providers including OpenAI, Anthropic, AWS Bedrock, Google Vertex, and local inference servers. Key features:
+
+- **Intelligent failover** — transparent routing to a configured fallback when a provider is unavailable
+- **Semantic caching** — caches responses by semantic similarity, reducing cost and latency
+- **MCP support** — Model Context Protocol integration for multi-tool coordination
+- **Prometheus metrics** — built-in observability for latency, throughput, and cost tracking
+
+Performance at scale: <100 µs gateway overhead at 5,000 RPS.
+
+## Deployment
+
+Bifrost runs locally as a lightweight gateway process. Options:
+
+```bash
+npx bifrost@latest       # 30-second startup via NPX
+```
+
+Docker containers and a Go SDK are also available for embedded or orchestrated deployments.
+
+## See also
+
+- [AI development pipeline](/architecture/ai-pipeline) — how Bifrost fits into the full model-routing pipeline
+- [PAL MCP](/architecture/ai-pipeline#pal-mcp) — the multi-model coordination layer on top of Bifrost
+- [nix-ai](/nix/nix-ai) — Nix package and config layer that manages the Bifrost process