## What RepoRelay does
RepoRelay is a local, first-party MCP bridge that gives ChatGPT **bounded
access to exactly one approved repository on your computer** — nothing more.
ChatGPT reviews your code through RepoRelay, and can leave a structured task
for a separate local coding agent (like Codex or Claude) through fixed handoff
files. RepoRelay is the security boundary between ChatGPT and your machine.
```text
ChatGPT reviews/plans
↓
RepoRelay
↓
Repository reads/searches
+
fixed .ai-handoff writers
↓
Codex / another local coding agent implements
```
## How the safety model works
MCP (Model Context Protocol) is the standard that lets ChatGPT call tools.
ChatGPT is the MCP client. RepoRelay is the local MCP server and security
boundary: it decides what ChatGPT may access and exposes exactly one approved
repository at a time. `tunnel-client` is only the secure networking pipe that
carries ChatGPT traffic to your computer.
| Component | Job |
| --- | --- |
| ChatGPT | MCP client — chooses RepoRelay tools. |
| Secure MCP Tunnel | Carries traffic from ChatGPT to your computer. |
| `tunnel-client` | Local network forwarder; points the tunnel at RepoRelay. |
| RepoRelay | MCP server + security boundary; enforces authentication and allowed access. |
| Repository | The one directory ChatGPT is allowed to inspect. |
**What ChatGPT can do through RepoRelay (the normal 7-tool setup):**
```text
✓ inspect files open_workspace, list_files, read_file
✓ find instructions list_files with instructionsOnly
✓ search the repository search_files
✓ write to three predetermined handoff targets
write_next_task, write_review, update_handoff_state
```
In read-only mode, only the four read/search tools are exposed.
**What ChatGPT cannot do:**
```text
✗ run shell commands
✗ run PowerShell
✗ run Git
✗ launch processes
✗ arbitrarily edit source files
✗ delete files
✗ choose arbitrary write targets
✗ access outside the approved repository
```
This is one of RepoRelay's strongest differentiators: ChatGPT can read and plan
against your code, but it gets **no execution capability** and can only write to
a few fixed handoff files you control.
## Before you install
### Which mode should I use?
RepoRelay runs in two modes. The right one depends on your ChatGPT plan and
workspace:
| Mode | Start command | Tools in Scan Tools |
| --- | --- | --- |
| **Read-only** | `reporelay quickstart "" --no-handoff-writes` | **4** |
| **Handoff** (default) | `reporelay quickstart ""` | **7** |
- **Personal ChatGPT workspaces:** if developer mode and the **Tunnel**
connection are available for your account, start with **read-only mode**.
ChatGPT can open, read, and search the repository, but cannot write handoff
files. Expect **4 tools**.
- **Business / Enterprise / Edu workspaces:** available capabilities depend on
your plan and workspace/admin settings. Where developer features and
handoffs are enabled, the normal setup exposes **7 tools** (4 read/search +
3 fixed handoff writers).
Both modes share the same security boundary; read-only mode simply disables
the three handoff writers. RepoRelay cannot enable ChatGPT developer mode or
tunnel access. OpenAI plan and workspace behavior changes over time, so check
the current
[OpenAI Secure MCP Tunnel guide](https://developers.openai.com/api/docs/guides/secure-mcp-tunnels)
for what your account offers.
You need:
- **Node.js** `>=22.19` and `<27` (npm is included). Check with
`node --version`.
- **An existing local project or repository** you want ChatGPT to review. It
must be a real folder on your computer — not a drive root and not your
whole user folder.
- **OpenAI Secure MCP Tunnel access.** RepoRelay reaches ChatGPT through
OpenAI's Secure MCP Tunnel. See the current
[OpenAI Secure MCP Tunnel guide](https://developers.openai.com/api/docs/guides/secure-mcp-tunnels)
for availability, permissions, and plan details.
- **Permission to create/use a custom MCP app** in the target ChatGPT
workspace (ChatGPT developer mode).
You do **not** need to download anything else. RepoRelay installs the official
OpenAI `tunnel-client` automatically during `reporelay tunnel setup`.
Do not worry about the handoff protocol yet. The quickstart sets up working
handoff files for you and explains them as you go.
## Quick setup
### Where setup happens
Setup touches three different surfaces, and the tunnel and the ChatGPT
integration are configured in **different places**:
```text
YOUR MACHINE reporelay quickstart ""
starts RepoRelay against ONE repository
↓
OPENAI PLATFORM create the Secure MCP Tunnel + runtime API key
(reporelay tunnel setup walks you through it)
↓
CHATGPT enable developer mode, then add RepoRelay
through the Plugins / custom MCP app flow,
choosing the tunnel you created
```
The **Secure MCP Tunnel** is created and configured in OpenAI Platform. The
**ChatGPT integration** that uses the tunnel is added **inside ChatGPT** — not
on the Platform page where you created the tunnel. If you already have a
tunnel and are wondering "where do I actually add RepoRelay to ChatGPT?", that
is step 6 below, inside ChatGPT itself.
> **Path examples.** Always quote the repository path:
>
> ```bash
> # macOS / Linux
> reporelay quickstart "$HOME/Projects/my-app"
> ```
>
> ```powershell
> # Windows PowerShell — keep the backslashes
> reporelay quickstart "C:\Users\you\Projects\my-app"
> ```
>
> `C:\Users\you\Projects\my-app` is correct. `C:Users\you\Projects\my-app` is
> not — the backslashes matter.
### 1. Install RepoRelay
```bash
npm install -g reporelay-mcp@latest
```
npm is the supported and tested installer. Check the install:
```bash
reporelay --version
```
If `reporelay` cannot be found (`command not found` on macOS/Linux, `is not
recognized` on Windows), see [Troubleshooting](#troubleshooting).
### 2. Start RepoRelay on one repository
```bash
reporelay quickstart "$HOME/Projects/my-app"
```
Replace the path with the repository you want to expose (Windows example:
`reporelay quickstart "C:\Projects\my-app"`). Keep this terminal open.
**No path? The current directory is only a default.** If you omit the
repository path, `quickstart` uses the directory you are standing in and says
so in its summary. Only this default depends on the current directory —
`reporelay tunnel setup`, `tunnel doctor`, and `tunnel run` read stored
per-user configuration and work from any directory. The tunnel does not need
to be started from the repository directory.
You should see:
```text
Ready.
Local MCP: http://127.0.0.1:7676/mcp
```
The normal quickstart enables the **7-tool handoff surface** (4 inspection
tools + 3 fixed handoff writers). RepoRelay now creates a small `.ai-handoff`
workspace and an `AGENTS.md` note so ChatGPT can leave structured tasks and
reviews for a separate local coding agent:
```text
.ai-handoff/NEXT_TASK.md
.ai-handoff/REVIEW.md
.ai-handoff/RESULT.md
.ai-handoff/STATE.json
AGENTS.md
```
Using read-only mode instead (`--no-handoff-writes`, the recommended starting
mode for personal ChatGPT workspaces that expose developer mode and Tunnel)?
RepoRelay exposes the 4 read/search tools and creates
none of these files. See [Read-only mode](#read-only-mode---no-handoff-writes).
**Why does RepoRelay create these?** ChatGPT still cannot run commands, use
Git, or arbitrarily edit your repository. These files are simply a place where
ChatGPT can leave a task, and a separate local coding agent (running on your
own machine, with your own permission) can leave its result. More below in
[ChatGPT ↔ coding-agent handoff](#chatgpt--coding-agent-handoff).
To stop RepoRelay later, press **Ctrl+C** in this window. There is no
`reporelay quickstart --stop`.
### 3. Audit it
Immediately after quickstart, verify RepoRelay's actual security boundary:
```bash
reporelay audit "$HOME/Projects/my-app"
```
You should see:
```text
RESULT: PASS
```
Audit starts its own temporary loopback listener and exercises the real
authenticated MCP surface, containment checks, and handoff restrictions. It
does not modify your repository. This validates RepoRelay *before* ChatGPT is
connected. If you started quickstart with `--no-handoff-writes`, add the same
flag to the audit command.
### 4. Run RepoRelay tunnel setup
In a **second** terminal window, run:
```bash
reporelay tunnel setup
```
This starts the RepoRelay setup wizard. It does everything for you:
```text
✓ installs a RepoRelay-supported OpenAI tunnel-client
✓ verifies it (pinned version + official SHA-256)
✓ opens OpenAI tunnel setup
✓ asks for your tunnel ID
✓ opens OpenAI runtime-key setup
✓ securely stores the pasted key
✓ creates the tunnel profile
✓ tests the complete connection
```
You provide exactly two things, both in OpenAI Platform:
1. **Your Secure MCP Tunnel ID** — the wizard opens
in your
browser. Create or select a tunnel, associate it with your ChatGPT
workspace, and paste its `tunnel_id` back in the terminal.
2. **A runtime API key** — the wizard opens
. Create a
secret key for the project you use with the tunnel and paste it in the
terminal. Input is hidden: nothing appears while you paste. This key
authenticates `tunnel-client` to OpenAI; it is **not** the RepoRelay
bridge secret.
Creating or editing a tunnel needs the **Tunnels Read + Manage** permission;
running `tunnel-client` or selecting the tunnel needs **Tunnels Read + Use**.
These are organization-level permissions granted by your org owner or RBAC
admin. Follow the current
[OpenAI Secure MCP Tunnel guide](https://developers.openai.com/api/docs/guides/secure-mcp-tunnels)
for the exact UI and permission details.
When the wizard finishes, you should see:
```text
Testing connection...
✓ OpenAI runtime credential
✓ RepoRelay reachable
✓ Bridge authentication
Setup complete.
Next:
reporelay tunnel run
```
These checks are genuine: setup validates the runtime API key against the
OpenAI control plane (the same read-only tunnel lookup `tunnel-client` performs
at startup) and verifies the bridge secret against the RepoRelay that is
actually running. A wrong, expired, or mis-pasted key is caught here with a
clear message — not after `tunnel run` starts.
If your RepoRelay runs on a custom port (for example `--port 7677`),
quickstart already recorded the live endpoint and setup follows it
automatically — no extra flags needed.
Useful options:
- `reporelay tunnel setup --no-open` — do not launch the browser (headless,
SSH, or CI); the URLs are still printed.
- `reporelay tunnel setup --replace-tunnel` — prompt for a new tunnel ID.
- `reporelay tunnel setup --replace-runtime-key` — prompt for a new runtime
API key.
- `reporelay tunnel setup --tunnel-client-path "/path/to/tunnel-client"`
(Windows: `tunnel-client.exe`) — **advanced override** for unusual
environments; RepoRelay does not verify or manage a custom binary.
Re-running `reporelay tunnel setup` reuses your existing verified client,
tunnel ID, and stored key, and re-tests the connection without asking for
anything again.
### 5. Run the tunnel
```bash
reporelay tunnel run
```
Keep this window open alongside the RepoRelay quickstart window. Stop it with
Ctrl+C when you are done.
If the connection ever stops working, `reporelay tunnel doctor` remains
available as a standalone troubleshooting command (expect `Ready.` when
everything is healthy; add `--verbose` for redacted diagnostics).
### 6. Add RepoRelay to ChatGPT
This step happens **inside ChatGPT** — not on the OpenAI Platform page where
you created the tunnel.
Naming note: the current ChatGPT UI may have you enter through **Plugins**,
while OpenAI documentation may still refer to the underlying integration as an
**App** or **custom MCP app**. This guide just says "the RepoRelay
integration." See the current
[ChatGPT developer-mode and MCP apps guide](https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt)
when labels move.
```text
ChatGPT
→ enable Developer Mode in settings
(exact location varies by plan/workspace as OpenAI updates the UI)
→ Plugins / developer features
→ create custom MCP app
→ connection: Tunnel
→ choose RepoRelay's tunnel
→ authentication: No authentication
→ Scan Tools
→ verify the tool count: 4 in read-only mode, 7 with handoffs
→ create/use the integration
→ start a new chat
```
Follow this sequence:
1. Enable **Developer Mode** in ChatGPT settings, or ask your workspace
administrator for access. The exact location of the setting may vary by
plan or workspace as OpenAI updates the interface.
2. Create the custom MCP app.
3. Choose the **Tunnel** connection.
4. Select the RepoRelay/OpenAI Secure MCP Tunnel.
5. When ChatGPT asks for authentication, select **No authentication**.
6. Save or create the integration.
7. Run **Scan Tools**.
8. Verify the expected RepoRelay tools appear.
9. Start a new chat and select the RepoRelay integration.
> **Authentication: No authentication.** RepoRelay already authenticates the
> local bridge through the protected `X-RepoRelay-Bridge-Secret` used by the
> tunnel. Do not configure OAuth or another ChatGPT-side authentication method.
Never paste `127.0.0.1`, `localhost`, the RepoRelay bridge secret, or an OpenAI
runtime API key into ChatGPT. The tunnel connection does all the networking.
### 7. Scan and verify the RepoRelay tools
In the ChatGPT integration flow, run **Scan Tools** and confirm that the tool
count matches your mode.
**Read-only mode (`--no-handoff-writes`) — exactly 4 tools:**
```text
open_workspace
list_files
read_file
search_files
```
**Normal handoff mode — exactly 7 tools** (the four above, plus):
```text
write_next_task
write_review
update_handoff_state
```
A different count usually means the other mode is running. If Scan Tools shows
shell, Git, process execution, generic file editing, delete, patching, or any
other unexpected capability, **stop and investigate** before using the
integration — run `reporelay audit "" --json` and confirm the tool
list.
### 8. Test it
Start a new chat, select the RepoRelay integration, and try:
```text
Open the approved repository and list its top-level files.
```
Then test the boundary:
```text
Try to read .env.
```
The second request should be blocked.
## You're connected — now what?
You now have ChatGPT reviewing your repository through a verified security
boundary. Useful prompts:
```text
Read README.md and explain how this project starts.
```
```text
Search the repository for "authentication".
```
```text
Find every AGENTS.md and CLAUDE.md instruction file before reviewing code.
```
RepoRelay handles that request through the existing `list_files` tool with
`instructionsOnly: true`. This recursively discovers recognized instruction
files without adding a broader filename-glob tool or changing the four-tool
read-only surface.
```text
Review src/server.ts for error-handling issues and write your findings.
```
## ChatGPT ↔ coding-agent handoff
The normal RepoRelay setup lets ChatGPT plan and review while a separate local
coding agent (Codex, Claude, or another) does the implementation. RepoRelay
coordinates them through a small `.ai-handoff` workspace:
```text
.ai-handoff/
├── NEXT_TASK.md ChatGPT writes the task here
├── RESULT.md the local coding agent writes its result here
├── REVIEW.md ChatGPT writes its review here
└── STATE.json coordinates the cycle
```
Conceptually:
```text
ChatGPT
↓ writes NEXT_TASK.md
Codex / local coding agent
↓ implements
↓ writes RESULT.md
ChatGPT
↓ reviews result
↓ writes REVIEW.md
```
`STATE.json` coordinates the cycle. RepoRelay itself **does not run Codex or
Claude** — they are separate local applications you start yourself. The
handoff files are just a structured place to hand work back and forth.
ChatGPT can only write `NEXT_TASK.md`, `REVIEW.md`, and `STATE.json`. The
implementer-owned `RESULT.md` is never writable by ChatGPT.
See [docs/handoff-cycle.md](docs/handoff-cycle.md) for the detailed protocol
and [examples/](examples/) for reviewer and implementer prompts.
## Daily use
After the one-time setup, do not recreate the tunnel or the ChatGPT
integration. Each day:
1. Start RepoRelay:
```bash
reporelay quickstart "$HOME/Projects/my-app"
```
2. Start the tunnel (in a second window):
```bash
reporelay tunnel run
```
3. Open ChatGPT, start a new chat, and select the existing RepoRelay
integration.
Both commands can be run from any directory: quickstart takes the repository
path explicitly, and the tunnel reads its stored per-user configuration.
Keep both windows open while you use the integration. If the connection stops
working, run `reporelay tunnel doctor` again.
## Switch repositories
**RepoRelay exposes one repository at a time.**
1. Press **Ctrl+C** in the RepoRelay terminal.
2. Start RepoRelay for the new repository:
```bash
reporelay quickstart "$HOME/Projects/another-repo"
```
3. Keep `tunnel-client` running. It reconnects to the restarted RepoRelay
automatically (same port and protected bridge-secret file). If you used a
different port, the managed tunnel follows it automatically.
4. Start a new ChatGPT conversation and select the RepoRelay integration.
5. Ask ChatGPT to open the new repository.
Do not rescan tools just because the approved repository changed. Rescan only
if the tool definitions changed or ChatGPT asks you to.
## Troubleshooting
| You see | What to do |
| --- | --- |
| `'node' is not recognized` / `node: command not found` | Node.js is not installed, or the terminal was opened before the install finished. Install Node.js from , close and reopen the terminal, and check `node --version`. |
| `RepoRelay requires Node.js >=22.19 and <27` | Your Node version is unsupported. Install a supported Node.js LTS release, reopen the terminal, and check `node --version`. |
| `'reporelay' is not recognized` (Windows) | The npm install did not finish or the terminal was opened before it finished. Re-run `npm install -g reporelay-mcp@latest`, close and reopen the terminal, and try `reporelay --version`. |
| `reporelay: command not found` (macOS/Linux) | The `reporelay` binary is not on your `PATH`. npm is the supported installer — re-run `npm install -g reporelay-mcp@latest` and reopen the terminal. If you installed globally with another package manager (Bun, pnpm, yarn), its global-bin directory may not be on your `PATH`; fix that package manager's PATH setting or reinstall with npm. Invoking `node .../node_modules/reporelay-mcp/dist/cli.js` through a nested package-manager path is a workaround, not a supported install. |
| `Cannot find module ... dist\cli.js` | You are running from a source checkout in the wrong folder. `cd` into the `RepoRelay` folder and run `npm run build` first. |
| `C:Users\you\...` (path looks mangled) | You dropped the backslashes. Quote the full Windows path: `reporelay quickstart "C:\Users\you\Projects\my-app"`. |
| Repository does not exist / not a directory | RepoRelay requires an existing directory. Double-check the quoted path and that the folder exists. |
| Repository root is too broad | The approved root must be a real project folder, not a drive root or your whole user folder. |
| `Port 7676 is already in use` | Another RepoRelay or program is listening on that port. Press Ctrl+C in its window to stop it, or rerun quickstart with a custom port (`--port 7677`) — the managed tunnel follows the new port automatically. There is no `quickstart --stop`; stop RepoRelay with **Ctrl+C**. On Windows, find the listener with `Get-NetTCPConnection -LocalPort 7676 -State Listen | Select-Object LocalAddress, LocalPort, OwningProcess` and inspect it with `Get-Process -Id `. RepoRelay never kills processes for you. |
| Quickstart stops about an existing `AGENTS.md` | The repository already has an `AGENTS.md` without the RepoRelay marker. RepoRelay will not overwrite it. Review the file first; if you want RepoRelay to preserve it and append the marked handoff instructions, rerun `reporelay quickstart "..." --append-agent-instructions`. |
| `tunnel-client` missing | Rerun `reporelay tunnel setup`; it re-downloads and verifies the managed client. |
| Invalid tunnel ID | The ID must look like `tunnel_` followed by 32 hex characters. Copy it again from Platform tunnel settings. |
| Runtime credential rejected | The runtime API key or tunnel context was not accepted. Check that the tunnel is associated with the target ChatGPT workspace (not only a Platform organization), that the key belongs to the same OpenAI organization/project as the tunnel, and that your account has **Tunnels Read + Use**. Rerun `reporelay tunnel setup --replace-runtime-key`. |
| Control plane unreachable | RepoRelay could not contact OpenAI to validate the credential. Check your internet connection, then rerun `reporelay tunnel doctor`. |
| Tunnel doctor cannot reach MCP | Keep the RepoRelay quickstart window running on the configured port, then rerun `reporelay tunnel doctor`. |
| Bridge authentication failure | Do not paste a secret. Confirm quickstart is using the canonical bridge-secret file, then rerun `reporelay tunnel setup`. |
| ChatGPT cannot see the tunnel | Check that the tunnel is associated with the target ChatGPT workspace (not only a Platform organization) and that you have tunnel-use permission. |
| Scan Tools returns zero tools | Keep both the RepoRelay and tunnel windows open, confirm `reporelay tunnel doctor` reports `Ready.`, then rescan in a **new** chat with the RepoRelay integration selected. |
| Scan Tools shows unexpected tools | Stop and investigate before using the integration: run `reporelay audit "" --json` and confirm the expected list — 4 tools in read-only mode, 7 with handoff writes. |
| RepoRelay window was closed | RepoRelay stopped. Restart it with `reporelay quickstart ""`, then try again in a new chat. |
| Tunnel window was closed | `tunnel-client` stopped. Restart it with `reporelay tunnel run`, then try again in a new chat. |
| Custom port mismatch | Confirm the RepoRelay quickstart port matches what `reporelay tunnel doctor` reports as the local MCP endpoint. Quickstart records the live endpoint automatically, or set it explicitly with `reporelay tunnel setup --port `. |
For anything else, run `reporelay doctor`. It prints configuration and security
status without printing secret values.
## Read-only mode (--no-handoff-writes)
Read-only mode is the recommended starting workflow for **personal ChatGPT
workspaces where developer mode and Tunnel are available** (see
[Which mode should I use?](#which-mode-should-i-use)), and for anyone who wants
ChatGPT to **inspect only** — no handoff files, no writes at all:
```bash
reporelay quickstart "$HOME/Projects/my-app" --no-handoff-writes
```
In this mode RepoRelay exposes exactly four tools:
```text
open_workspace
list_files
read_file
search_files
```
Read-only mode does **not** create `.ai-handoff`, does **not** create or modify
`AGENTS.md`, and leaves the approved repository unchanged. All containment,
authentication, and security checks still apply. Use the matching audit flag:
```bash
reporelay audit "$HOME/Projects/my-app" --no-handoff-writes
```
When you use this mode, expect **4 tools** in Scan Tools instead of 7.
## Security
RepoRelay's enforced boundary includes:
- loopback-only binding and authentication required;
- exactly one existing canonical approved root;
- sensitive paths blocked, including `.env`, VCS metadata, credential stores,
and private-key formats;
- traversal, absolute outside-root paths, symlink/junction/reparse escapes, and
hard-link bypasses blocked;
- bounded reads, searches, results, and handoff content;
- no shell, PowerShell, Git, process, generic write, patch, delete, artifact,
worktree, skill, subagent, or local-agent tool;
- handoff writes limited to the three fixed pre-existing targets.
RepoRelay is a least-privilege application boundary, not an operating-system
sandbox against malicious software already running as the same local user. An
external Secure MCP Tunnel is a separate security boundary and must be secured
independently. Choose the approved repository carefully.
See [SECURITY.md](SECURITY.md) for the full security model and
`reporelay audit` documentation.
## Advanced configuration
Most users never need these. When you do:
- [Configuration](docs/configuration.md) — `REPORELAY_*` environment variables,
bridge authentication, ports, logging, and handoff mode.
- [ChatGPT Web setup details](docs/chatgpt-web.md) — the full tunnel and app
reference.
- [Windows operations](OPERATIONS.md) — lifecycle scripts, scheduled tasks, and
managed Windows setup.
- [`.env.example`](.env.example) — placeholder configuration only. Never put a
real bridge secret or tunnel credential in it.
- **Local MCP clients** — if your MCP client runs on the same computer as
RepoRelay, you do not need the tunnel: point it at the local MCP URL printed
by quickstart and send the `X-RepoRelay-Bridge-Secret` header loaded from the
protected file.
For configuration and security status without printing secret values:
```bash
reporelay doctor
```
## Requirements and platform support
- Node.js `>=22.19 <27` (npm is included);
- Git for the clone-based install — optional if you download the ZIP instead.
The quickstart in this README works on Windows, macOS, and Linux; npm is the
supported installer. Windows 10/11 is the fully validated lifecycle and
operational platform, including the PowerShell scripts and tunnel-managed
runbook.
## Limitations
- One approved repository is exposed at a time.
- RepoRelay is not an operating-system sandbox.
- ChatGPT Web needs external OpenAI Secure MCP Tunnel infrastructure to reach a
private local bridge.
- Windows is the fully validated lifecycle platform; other platforms do not
have identical PowerShell/tunnel operations.
- Codex, Claude, and other coding agents are separate local applications, not
components of RepoRelay.
## Documentation
- [ChatGPT Web setup](docs/chatgpt-web.md)
- [Security model](SECURITY.md)
- [Configuration](docs/configuration.md)
- [Windows operations](OPERATIONS.md)
- [Setup notes](docs/setup.md)
- [Handoff cycle](docs/handoff-cycle.md)
- [Handoff examples](examples/)
- [Contributing](CONTRIBUTING.md)
## License and project lineage
RepoRelay is released under the [MIT License](LICENSE). The project keeps its
required upstream attribution and does not bundle the SDKs or runtimes of
Codex, Claude, or other implementers.
## Development
```bash
npm ci
npm run typecheck
npm test
npm run verify:release
npm audit --audit-level=low
npm pack --dry-run --json
git diff --check
```
