diff --git a/README.md b/README.md
index c6552ae..08a34d1 100644
--- a/README.md
+++ b/README.md
@@ -1,119 +1,157 @@
 # cli-mcp-server
 
-MCP server that provides AI agents with a sandboxed bash shell inside per-investigation Kubernetes pods.
+MCP server that gives AI agents a **sandboxed bash shell** inside per-session Kubernetes pods.
 
-## Architecture
+Agents call a single `bash` tool — pipes, redirects, chaining, and whatever CLIs are installed in the sandbox image. The MCP server is a stateless control plane: it creates and routes to sandbox pods, then proxies commands. Session state (working directory, env vars, files) lives in the pod’s persistent bash process.
 
-Two binaries in one repo:
+Designed to run against a Kubernetes (or OpenShift) cluster — the server manages sandbox pods via the API. It is not a local-only shell MCP.
 
-- **cli-mcp-server** (`cmd/server/`) — Control plane. Manages sandbox pod lifecycle via client-go, exposes the `bash` MCP tool, proxies commands to sandbox agents.
-- **sandbox-agent** (`cmd/agent/`) — Data plane. Runs inside each sandbox pod, manages a persistent bash process with `oc`/`kubectl`/`jq`/`yq`/`curl`.
+## Features
 
-## Project Structure
+- **Persistent bash per session** — cwd, environment, and `/workspace` files survive across tool calls
+- **Full shell, not a command allowlist** — security comes from pod isolation, RBAC, and network policy
+- **Per-session agent auth** — MCP server → sandbox `/exec` uses an HMAC-derived bearer token (defense in depth beyond NetworkPolicy)
+- **Customizable CLIs** — available tools are whatever is in the sandbox agent image (the default image includes `oc`/`kubectl`)
+- **Stateless, multi-replica ready** — any server replica can handle any request; Kubernetes is the source of truth
+- **Optional warm pool** — pre-warmed pods cut cold-start latency when enabled
 
-```
-cli-mcp-server/
-├── cmd/
-│   ├── server/main.go          # MCP server entry point
-│   └── agent/main.go           # Sandbox agent entry point
-├── pkg/
-│   ├── session/                 # Pod lifecycle, warm pool, cache
-│   ├── agent/                   # HTTP client + shared types
-│   ├── sandbox/                 # Bash session + HTTP handlers
-│   ├── server/                  # MCP server setup
-│   ├── tools/                   # MCP tool handlers (bash)
-│   └── version/                 # Build version info
-├── docs/
-│   ├── proposals/               # Design documents
-│   └── implementation/          # Implementation plan and Jira stories
-├── Makefile
-└── go.mod
-```
+## Usage
 
-## Build
+Works with any MCP client that can call tools over HTTP (or stdio) and send an `X-Session-ID` header. Over HTTP this is straightforward; over stdio it depends on whether the client/SDK can attach request headers.
 
-```bash
-make build          # Build both server and agent
-make build-server   # Build only the MCP server
-make build-agent    # Build only the sandbox agent
-make test           # Run tests
-make lint           # Run linter
-make build-prod     # Production build (static, CGO disabled)
+### MCP tool: `bash`
+
+| Parameter | Type | Required | Description |
+|-----------|------|----------|-------------|
+| `command` | string | yes | Shell command (full bash — pipes, redirects, chaining) |
+| `timeout` | int | no | Max execution time in seconds (default 60, max 300) |
+
+Example calls:
+
+```json
+{"command": "kubectl get pods -n kube-system --context=prod"}
+{"command": "kubectl get pods -o json | jq '.items[].metadata.name'", "timeout": 120}
 ```
 
-## Session & Pod Lifecycle
+Returns `stdout`, `stderr`, `exit_code`, and `duration_ms`. Non-zero exit codes are tool results (not transport errors) so the agent can use failure output.
 
-Each investigation gets its own sandbox pod. The MCP server is stateless — session
-state lives in Kubernetes (pod labels/annotations) and is cached in memory for performance.
+### Session routing
 
-### Request Flow
+Every request must include:
 
+```http
+X-Session-ID: <session-id>
 ```
-TARSy (LLM)
-  │  MCP tool call + X-Session-ID header
-  ▼
-MCP Server (stateless, multi-replica)
-  │  HTTP POST /exec + Bearer HMAC token
-  ▼
-Sandbox Pod (per-session, persistent bash)
-  │  stdout / stderr / exit_code
-  ▼
-MCP Server → TARSy
-```
 
-### GetOrCreatePod — Three-Tier Lookup
+The server only uses this value to find or create a sandbox pod. Session IDs must be RFC 1123 DNS labels (`^[a-z0-9]([a-z0-9-]{0,61}[a-z0-9])?$`); invalid IDs fail at pod create.
+
+Typical client practice (agent harness / orchestrator — not the LLM):
+
+1. Generate a new session ID when starting work that should share one sandbox
+2. Send that same ID on every follow-up `bash` call to reuse the persistent shell and `/workspace`
+3. Call `DELETE /sessions/{id}` when finished (HTTP transport only) — removes the sandbox pod, auth secret, and cache entry
 
-When a command arrives for a session, `SessionManager.GetOrCreatePod` resolves the
-target pod using a three-tier strategy:
+If the client never deletes the session, the sandbox is garbage-collected after `--idle-timeout` (default 30m).
 
-| Tier | Method | Latency | Description |
-|------|--------|---------|-------------|
-| 1 | **PodCache** | ~0ms | In-memory map of session ID → pod IP with 30s TTL |
-| 2 | **Label discovery** | ~50-100ms | `List` pods by `tarsy.redhat.com/session-id` label |
-| 3 | **Create** | ~3-8s | Create auth Secret + Pod, wait for readiness probe |
+## Configuration
 
-Cache misses fall through to tier 2; if no pod exists, tier 3 creates one
-idempotently (handles `AlreadyExists`).
+### Server flags
 
-### Warm Pool
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--transport` | `stdio` | `stdio` or `http` (`http` requires `--stateless`) |
+| `--address` | `localhost:8080` | Listen address (HTTP; must be loopback) |
+| `--stateless` | `false` | Required for HTTP / multi-replica |
+| `--namespace` | `tarsy` | Namespace for sandbox pods |
+| `--sandbox-image` | _(required)_ | Container image for sandbox pods |
+| `--hmac-key-file` | _(required)_ | Path to shared HMAC secret |
+| `--kubeconfig` | _(in-cluster)_ | Kubeconfig for managing sandbox pods |
+| `--idle-timeout` | `30m` | Delete idle sandbox pods after this duration |
+| `--warm-pool-size` | `0` | Pre-warmed pods (`0` = create on demand) |
 
-When `--warm-pool-size N` is set, N pre-warmed pods are maintained in a ready state.
-New sessions are assigned a warm pod instantly (~0ms) instead of waiting for a cold
-start. Disabled by default (`N=0`).
+### Sandbox image (available CLIs)
 
-### Pod Spec
+CLIs available to the agent are determined by the **sandbox agent image**, not by server code.
 
-Each sandbox pod runs with a hardened security context:
+The default image (`Containerfile.agent`) is based on `oc-client-base-minimal` and includes `oc`, `kubectl`, plus utilities such as `jq`, `yq`, and `curl`.
 
-- **Non-root**: UID/GID 1001, `allowPrivilegeEscalation: false`, all capabilities dropped
-- **Resources**: 100m/500m CPU, 128Mi/512Mi memory
-- **Readiness**: HTTP GET `/health` on port 8090
-- **Volumes**: `kubeconfig` (read-only Secret mount), `workspace` (emptyDir)
-- **Auth**: HMAC-SHA256(key, sessionID) → bearer token delivered via Secret → env var
+To add other CLIs (for example `helm` or `virtctl`):
 
-### Cleanup
+1. Extend `Containerfile.agent` (or build a custom image from it)
+2. Build and push the image
+3. Point the MCP server at it with `--sandbox-image`
 
-| Method | Trigger | Description |
-|--------|---------|-------------|
-| **Explicit delete** | `DELETE /sessions/{id}` | TARSy calls when investigation ends. Deletes pod, Secret, and cache entry |
-| **Stale cleanup** | Background ticker | Deletes pods idle > 30 minutes (via `last-activity` annotation) |
+No MCP server code changes are required. Tell the LLM what is available via your client’s server instructions (or equivalent); the `bash` tool description stays generic.
+
+## Architecture
+
+```mermaid
+flowchart TB
+  Client[MCP Client]
+  Server["cli-mcp-server<br/>stateless · N replicas"]
+  Sandbox["Sandbox pods<br/>assigned sessions · optional warm pool"]
+  Target["Target infrastructure<br/>e.g. Kubernetes API"]
+
+  Client -->|"bash + X-Session-ID"| Server
+  Server -->|"POST /exec + HMAC token"| Sandbox
+  Sandbox -->|"CLIs from sandbox image / config"| Target
+```
 
-If a sandbox pod crashes mid-investigation, shell state (env vars, working directory,
-files in emptyDir) is lost. The MCP server detects the transport error, invalidates
-the cache, and creates a new pod on the next call. TARSy retains the investigation
-context.
+Bash commands run in the sandbox pods. What they can reach (for example a Kubernetes API via `kubectl`) depends on the sandbox image and the kubeconfig/RBAC mounted into those pods — not on the MCP server binary.
 
-## Design Documents
+### Scalability
 
-- [Sketch](docs/proposals/cli-mcp-server-sketch.md) — Problem statement and approach
-- [Design Overview](docs/proposals/cli-mcp-server-design-overview.md) — Architecture overview
-- [Detailed Design](docs/proposals/cli-mcp-server-design.md) — Go types, flows, YAML specs
+The MCP server holds no durable session state. Pod identity is stored in Kubernetes labels; an in-memory cache speeds up routing. Any replica can serve any request, so you can scale the Deployment horizontally behind a load balancer with no sticky sessions.
+
+Optional `--warm-pool-size` keeps ready pods on hand so new sessions skip cold start (image pull + container boot).
+
+### Sandboxing and security
+
+Each session gets its own pod. That pod is the security boundary:
+
+- **Isolation** — non-root (runAsNonRoot), no privilege escalation, all capabilities dropped, resource limits
+- **Credentials** — read-only kubeconfig mounted from a dedicated investigation ServiceAccount (typically view/read-only RBAC)
+- **Network** — NetworkPolicy can restrict ingress to the MCP server and egress to intended APIs
+- **Agent auth** — per-session HMAC bearer token; unauthenticated `/exec` calls are rejected
+- **Ephemeral workspace** — `/workspace` is an `emptyDir`; destroyed with the pod
+
+The server does not filter shell commands. Capability is controlled by what is in the image and what RBAC allows.
+
+### Components
+
+| Component | Role |
+|-----------|------|
+| **cli-mcp-server** | Control plane: MCP `bash` tool, pod lifecycle, command proxy |
+| **sandbox-agent** | Data plane inside each pod: persistent bash over HTTP (`/exec`, `/health`, `/assign`) |
+
+## Development
+
+```bash
+make build          # Build both server and agent
+make build-server   # Build only the MCP server
+make build-agent    # Build only the sandbox agent
+make test           # Run tests
+make lint           # Run linter
+make build-prod     # Production build (static, CGO disabled)
+```
+
+```
+cli-mcp-server/
+├── cmd/server/     # MCP server entry point
+├── cmd/agent/      # Sandbox agent entry point
+├── pkg/session/    # Pod lifecycle, warm pool, cache
+├── pkg/sandbox/    # Bash session + agent HTTP handlers
+├── pkg/tools/      # MCP tool handlers
+├── pkg/server/     # MCP server + HTTP mux
+└── docs/           # Design and implementation docs
+```
 
-## Implementation
+### Further reading
 
-- [Implementation Plan](docs/implementation/implementation-plan.md) — Phased plan
-- [Jira Stories](docs/implementation/stories.md) — Epic SANDBOX-1803 breakdown
+- [Sketch](docs/proposals/cli-mcp-server-sketch.md) — problem statement and approach
+- [Design Overview](docs/proposals/cli-mcp-server-design-overview.md) — architecture overview
+- [Detailed Design](docs/proposals/cli-mcp-server-design.md) — types, flows, deployment specs
 
 ## License
 
-The code is available under the [Apache License 2.0](LICENSE).
\ No newline at end of file
+[Apache License 2.0](LICENSE)