diff --git a/AGENTS.md b/AGENTS.md
index 2ee7643..8bbe20c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -224,6 +224,19 @@ make release  # GoReleaser dry-run
 make install  # Install to GOPATH/bin
 ```
 
+### CLI Commands
+
+| Command | Purpose |
+|---------|---------|
+| `replicator init` | Per-repo setup: creates `.hive/` with empty `cells.json` |
+| `replicator setup` | Per-machine setup: creates `~/.config/swarm-tools/` + SQLite DB |
+| `replicator serve` | Start MCP JSON-RPC server on stdio |
+| `replicator cells` | List hive cells (work items) |
+| `replicator doctor` | Check environment health |
+| `replicator stats` | Display activity summary |
+| `replicator query` | Run preset SQL analytics queries |
+| `replicator version` | Print version, commit, build date |
+
 ## Project Structure
 
 ```
diff --git a/README.md b/README.md
index ad9a1fa..aabf0fe 100644
--- a/README.md
+++ b/README.md
@@ -32,6 +32,9 @@ go install github.com/unbound-force/replicator/cmd/replicator@latest
 ## Usage
 
 ```bash
+# Initialize a project for swarm operations
+replicator init
+
 # Start MCP server (for AI agent connections)
 replicator serve
 
diff --git a/cmd/replicator/init.go b/cmd/replicator/init.go
new file mode 100644
index 0000000..925d9c2
--- /dev/null
+++ b/cmd/replicator/init.go
@@ -0,0 +1,52 @@
+package main
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+
+	"github.com/spf13/cobra"
+)
+
+func initCmd() *cobra.Command {
+	var pathFlag string
+	cmd := &cobra.Command{
+		Use:   "init",
+		Short: "Initialize a project directory for swarm operations",
+		Long: `Creates a .hive/ directory with an empty cells.json in the target
+directory. Idempotent — safe to run multiple times.
+
+This is the per-repo initialization command. It does not require the
+global database (replicator setup) or any external services.`,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			return runInit(pathFlag)
+		},
+	}
+	cmd.Flags().StringVar(&pathFlag, "path", ".", "Target directory for .hive/ initialization")
+	return cmd
+}
+
+// runInit creates the .hive/ directory and seeds cells.json.
+func runInit(targetDir string) error {
+	hiveDir := filepath.Join(targetDir, ".hive")
+
+	// Check if already initialized.
+	if info, err := os.Stat(hiveDir); err == nil && info.IsDir() {
+		fmt.Println("already initialized")
+		return nil
+	}
+
+	// Create .hive/ directory.
+	if err := os.MkdirAll(hiveDir, 0o755); err != nil {
+		return fmt.Errorf("create .hive directory: %w", err)
+	}
+
+	// Write empty cells.json.
+	cellsPath := filepath.Join(hiveDir, "cells.json")
+	if err := os.WriteFile(cellsPath, []byte("[]\n"), 0o644); err != nil {
+		return fmt.Errorf("write cells.json: %w", err)
+	}
+
+	fmt.Println("initialized .hive/")
+	return nil
+}
diff --git a/cmd/replicator/init_test.go b/cmd/replicator/init_test.go
new file mode 100644
index 0000000..ae2f5dd
--- /dev/null
+++ b/cmd/replicator/init_test.go
@@ -0,0 +1,78 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestRunInit_FreshDirectory(t *testing.T) {
+	dir := t.TempDir()
+	if err := runInit(dir); err != nil {
+		t.Fatalf("runInit: %v", err)
+	}
+
+	hiveDir := filepath.Join(dir, ".hive")
+	info, err := os.Stat(hiveDir)
+	if err != nil {
+		t.Fatalf(".hive/ not created: %v", err)
+	}
+	if !info.IsDir() {
+		t.Fatal(".hive/ is not a directory")
+	}
+
+	cellsPath := filepath.Join(hiveDir, "cells.json")
+	data, err := os.ReadFile(cellsPath)
+	if err != nil {
+		t.Fatalf("cells.json not created: %v", err)
+	}
+	if string(data) != "[]\n" {
+		t.Errorf("cells.json content = %q, want %q", string(data), "[]\n")
+	}
+}
+
+func TestRunInit_AlreadyInitialized(t *testing.T) {
+	dir := t.TempDir()
+
+	// First init.
+	if err := runInit(dir); err != nil {
+		t.Fatalf("first runInit: %v", err)
+	}
+
+	// Write something to cells.json to verify it's not overwritten.
+	cellsPath := filepath.Join(dir, ".hive", "cells.json")
+	os.WriteFile(cellsPath, []byte(`[{"id":"test"}]`), 0o644)
+
+	// Second init — should be idempotent.
+	if err := runInit(dir); err != nil {
+		t.Fatalf("second runInit: %v", err)
+	}
+
+	// Verify cells.json was NOT overwritten.
+	data, _ := os.ReadFile(cellsPath)
+	if string(data) != `[{"id":"test"}]` {
+		t.Errorf("cells.json was overwritten: got %q", string(data))
+	}
+}
+
+func TestRunInit_CustomPath(t *testing.T) {
+	parent := t.TempDir()
+	target := filepath.Join(parent, "myproject")
+	os.MkdirAll(target, 0o755)
+
+	if err := runInit(target); err != nil {
+		t.Fatalf("runInit with custom path: %v", err)
+	}
+
+	cellsPath := filepath.Join(target, ".hive", "cells.json")
+	if _, err := os.Stat(cellsPath); err != nil {
+		t.Fatalf("cells.json not created at custom path: %v", err)
+	}
+}
+
+func TestRunInit_InvalidPath(t *testing.T) {
+	err := runInit("/nonexistent/path/that/cannot/exist")
+	if err == nil {
+		t.Fatal("expected error for invalid path")
+	}
+}
diff --git a/cmd/replicator/main.go b/cmd/replicator/main.go
index ad9448e..ca0a148 100644
--- a/cmd/replicator/main.go
+++ b/cmd/replicator/main.go
@@ -42,6 +42,7 @@ via Dewey).`,
 	root.AddCommand(statsCmd())
 	root.AddCommand(queryCmd())
 	root.AddCommand(setupCmd())
+	root.AddCommand(initCmd())
 
 	if err := root.Execute(); err != nil {
 		fmt.Fprintln(os.Stderr, err)
diff --git a/openspec/changes/add-init-command/.openspec.yaml b/openspec/changes/add-init-command/.openspec.yaml
new file mode 100644
index 0000000..796d2d7
--- /dev/null
+++ b/openspec/changes/add-init-command/.openspec.yaml
@@ -0,0 +1,2 @@
+schema: unbound-force
+created: 2026-04-06
diff --git a/openspec/changes/add-init-command/design.md b/openspec/changes/add-init-command/design.md
new file mode 100644
index 0000000..5d13cb5
--- /dev/null
+++ b/openspec/changes/add-init-command/design.md
@@ -0,0 +1,29 @@
+## Context
+
+Replicator has `setup` (per-machine: `~/.config/swarm-tools/` + SQLite) but no `init` (per-repo). The `uf` CLI needs `replicator init` to exist so it can delegate to it during `uf init`, following the same pattern as `dewey init`.
+
+The existing `setup.go` pattern: cobra command → `runSetup()` function → filesystem ops → print status. Follow this exactly.
+
+## Goals / Non-Goals
+
+### Goals
+- Add `replicator init` command that creates `.hive/cells.json` in the target directory
+- Support `--path` flag for specifying a non-CWD target
+- Idempotent (exit 0 whether creating or already exists)
+- No database, no git, no network
+
+### Non-Goals
+- Creating `.hive/memories.jsonl` or other hive files (those are created by specific tools when needed)
+- Running `git add` or `git commit` (that's `hive_sync`)
+- Initializing the SQLite database (that's `replicator setup`)
+- Adding `.hive/` to `.gitignore` (user decides tracking policy)
+
+## Decisions
+
+**D1: Follow the `setup.go` pattern.** New file `cmd/replicator/init.go` with `initCmd() *cobra.Command` and `runInit(targetDir string) error`. Register in `main.go` via `root.AddCommand(initCmd())`.
+
+**D2: Seed file is `cells.json` with `[]`.** This matches the format `hive_sync` writes. An empty JSON array signals "initialized but no cells yet" rather than the ambiguity of a missing file.
+
+**D3: `--path` defaults to `.` (current directory).** Same UX as `dewey init` which operates on CWD by default.
+
+**D4: Testable via `t.TempDir()`.** The `runInit` function takes a directory path, making it directly testable without mocking.
diff --git a/openspec/changes/add-init-command/proposal.md b/openspec/changes/add-init-command/proposal.md
new file mode 100644
index 0000000..8926a98
--- /dev/null
+++ b/openspec/changes/add-init-command/proposal.md
@@ -0,0 +1,33 @@
+## Why
+
+The `uf init` command delegates per-repo initialization to each sub-tool (`dewey init`, `gaze init`). Replicator lacks an `init` command, so `uf init` cannot prepare a repository for swarm operations. Without it, the `.hive/` directory only appears organically on first `hive_sync` call, leaving new projects in an ambiguous state where the directory structure doesn't signal that swarm coordination is available.
+
+This blocks `unbound-force/unbound-force#82` (replace Swarm plugin with Replicator in the `uf` CLI).
+
+## What Changes
+
+### New Capabilities
+- `replicator init` creates a `.hive/` directory with an empty `cells.json` in the current working directory
+- `replicator init --path /some/dir` creates `.hive/` at a specified location
+- Idempotent: safe to run multiple times (prints "already initialized" if `.hive/` exists)
+
+## Impact
+
+- `cmd/replicator/init.go` (new file, ~40 lines)
+- `cmd/replicator/main.go` (add `initCmd()` registration)
+- `README.md` (add `init` to usage section)
+- `AGENTS.md` (add `init` to commands table)
+
+## Constitution Alignment
+
+### I. Autonomous Collaboration
+**PASS**: `init` creates a well-known directory (`.hive/`) that other tools can discover without coordination.
+
+### II. Composability First
+**PASS**: No database or external service required. Works standalone before `replicator setup` has run.
+
+### III. Observable Quality
+**PASS**: Prints machine-parseable status (`initialized .hive/` or `already initialized`). Exit 0 on success, exit 1 on error.
+
+### IV. Testability
+**PASS**: Uses `t.TempDir()` for filesystem tests. No external dependencies.
diff --git a/openspec/changes/add-init-command/specs/init-command.md b/openspec/changes/add-init-command/specs/init-command.md
new file mode 100644
index 0000000..6855efd
--- /dev/null
+++ b/openspec/changes/add-init-command/specs/init-command.md
@@ -0,0 +1,34 @@
+## ADDED Requirements
+
+### Requirement: Per-repo initialization
+
+`replicator init` MUST create a `.hive/` directory with an empty `cells.json` file in the target directory.
+
+#### Scenario: Fresh directory
+- **GIVEN** a directory without `.hive/`
+- **WHEN** `replicator init` is run
+- **THEN** `.hive/cells.json` is created containing `[]`
+- **AND** stdout prints `initialized .hive/`
+- **AND** exit code is 0
+
+#### Scenario: Already initialized
+- **GIVEN** a directory with existing `.hive/`
+- **WHEN** `replicator init` is run
+- **THEN** no files are modified
+- **AND** stdout prints `already initialized`
+- **AND** exit code is 0
+
+#### Scenario: Custom path
+- **GIVEN** a valid directory at `/some/path`
+- **WHEN** `replicator init --path /some/path` is run
+- **THEN** `.hive/cells.json` is created at `/some/path/.hive/cells.json`
+
+#### Scenario: Invalid path
+- **GIVEN** a path that does not exist and cannot be created
+- **WHEN** `replicator init --path /nonexistent/path` is run
+- **THEN** an error is printed
+- **AND** exit code is 1
+
+### Requirement: No external dependencies
+
+`replicator init` MUST NOT require the SQLite database, git, network access, or any other external service to function.
diff --git a/openspec/changes/add-init-command/tasks.md b/openspec/changes/add-init-command/tasks.md
new file mode 100644
index 0000000..c0a972b
--- /dev/null
+++ b/openspec/changes/add-init-command/tasks.md
@@ -0,0 +1,19 @@
+## 1. Implement init command
+
+- [x] 1.1 Create `cmd/replicator/init.go` with `initCmd() *cobra.Command` (--path flag, default ".") and `runInit(targetDir string) error`
+- [x] 1.2 In `runInit`: check if `.hive/` exists (print "already initialized", return nil), otherwise `os.MkdirAll` + `os.WriteFile("cells.json", "[]")`
+- [x] 1.3 Register `initCmd()` in `cmd/replicator/main.go` via `root.AddCommand(initCmd())`
+
+## 2. Tests
+
+- [x] 2.1 Create `cmd/replicator/init_test.go` with tests: fresh dir creates `.hive/cells.json`, already-initialized is idempotent, custom --path works, cells.json content is `[]`
+
+## 3. Documentation
+
+- [x] 3.1 Add `replicator init` to the Usage section in `README.md`
+- [x] 3.2 Add `make init` is NOT needed (clarify `init` is a CLI command, not a Makefile target) -- update AGENTS.md commands section to include `replicator init`
+
+## 4. Verify
+
+- [x] 4.1 Run `make check` -- all tests pass
+- [x] 4.2 Run `./bin/replicator init` in a temp dir -- verify `.hive/cells.json` created