# Connect Muse chat to Kody Muse chat (Meta's Muse assistant at [muse.ai](https://muse.ai); the older agent.meta.ai address redirects there) can use Kody over MCP, but it has no native MCP server settings screen, and it is not the same host as Muse Code. Muse Code connects from a machine with `~/.config/muse/settings.json` and `muse mcp login kody` (see the Muse Code entry in the [host notes](https://raw.githubusercontent.com/kody-hq/kody/main/docs/use/connect-your-agent.md)). Do not follow those instructions for Muse chat, and do not paste `npx @kodycodes/cli install` into Muse chat: that runs in Muse's Linux VM and cannot configure Muse Code on your machine. Muse chat connects through a custom connector: Muse builds a small MCP client for Kody's hosted MCP server and stores the OAuth credential in its Secure credentials store. The client speaks Streamable HTTP, and every tool call is gated behind an explicit allowlist. This guide covers the search-only setup. `execute` is a deliberate opt-in after search works, not part of the default setup. ## What approving access means Approving an agent gives it full access to that Kody account, not a limited permission set. You can revoke it on Kody's Agents page at `/@/-/agents`, where `` is your Kody username. One MCP connection binds one org. Kody's MCP server offers `search`, `execute`, and `api`. For this setup, allowlist `search` first and prove it before considering anything else. Do not allowlist `api`: it calls Kody API operations directly, including writes and token operations such as `tokenCreate`, and it is out of scope for a first connection. ## Prerequisites - A Kody account, signed in, with a verified email address. Authorize cannot finish until the account email is verified. - A Muse account. - No API key is involved. Kody's MCP server is OAuth-only; Kody API tokens do not work on `/mcp`. ## 1. Pre-register a static OAuth client on Kody Muse's custom-connector flow asks you to supply a Client ID and Client Secret in its Secure credentials store; it does not complete dynamic client registration for you. Kody's MCP OAuth clients page exists for exactly this case: it pre-registers a confidential client for a host that cannot finish dynamic OAuth registration. Open Workspace → Agents, then **MCP OAuth clients** in the **Advanced** panel (`/account/mcp-oauth-clients`). Under **Create a client**, set **Label** to something you will recognize (for example `Muse chat`). In **Redirect URIs**, enter these two values, one per line: ```text https://muse.ai/api/hatch/oauth/callback https://agent.meta.ai/api/hatch/oauth/callback ``` Copy them exactly. Kody accepts only a redirect URI that matches one saved on the client, and a different callback is the most common failure in this setup (see Troubleshooting). The `agent.meta.ai` callback is the one this setup was first proven with; Muse has since moved to muse.ai, which serves the same `/api/hatch/oauth/callback` path. Registering both keeps the client working whichever one Muse sends. Select **Create client**. Kody shows the Client ID and Client secret once under **New client credentials** and does not store the secret after you leave the page. Keep them for step 2 only, and never paste them into chat. Do not confuse this page with two other Kody pages: - **Your clients**, further down the same page, lists the OAuth clients you registered. A client existing there does not by itself prove an agent grant. - **Agents** (`/@/-/agents`) lists inbound agents granted access. A Muse entry there may be for Muse Code, not Muse chat, so read its connection instructions before concluding that Muse chat is granted. ## 2. Ask Muse to create the custom connector Ask Muse to create a custom connector for Kody with these details. They are the values for [kody.codes](https://kody.codes), whose MCP URL is `https://kody.codes/mcp`: ```json { "provider": "kody", "auth_scheme": "oauth2_code", "issuer_domain": "kody.codes", "authorization_url": "https://kody.codes/oauth/authorize", "token_url": "https://kody.codes/oauth/token", "token_endpoint_auth_method": "client_secret_basic", "scopes": "openid profile email", "resource": "https://kody.codes/mcp", "api_hosts": ["kody.codes"] } ``` On another deployment (a preview or a self-hosted Kody), replace every `kody.codes` value above (`issuer_domain`, `authorization_url`, `token_url`, `resource`, and `api_hosts`) with that deployment's host, and register the client in step 1 on that same deployment. Every deployment publishes its OAuth endpoints at `/.well-known/oauth-authorization-server` if you want to confirm them. Muse will open its Secure credentials store. Enter the Client ID and Client secret from step 1 there, and nowhere else. Muse then redirects to Kody's authorize page. Sign in if needed, approve access, and the connection is stored. ## 3. Prove it with a harmless read Do not start with `execute`, and do not allowlist a tool before its schema has been returned. Ask Muse to list tools first, then run the harmless search that Kody's own onboarding guide uses: ```text list-tools search({ entity: "guide:onboarding" }) ``` Success looks like: `list-tools` returns `search`, `execute`, and `api` with their input schemas, and the search returns the onboarding guide, which describes Kody as your agents' cloud (not a gateway) and starts by moving what is already in the agent into Kody, with six first-win ideas as the fallback. Only `search` needs to be allowlisted to reach this point. ## What failed on the first attempt Three failures shaped this guide. Each looked like a different problem until the error itself was read: 1. **Wrong redirect URI.** The first client was registered with an older example callback instead of the `hatch` callback above. Authorize failed with `Invalid redirect URI` before any consent screen appeared. The fix was a new client with the exact `hatch` callback, not an edit to the old client. 2. **A 403 that was not a Kody refusal.** `list-tools` returned HTTP 403, which first read as a scope, audience, or agent-grant problem. The error body was Cloudflare Error 1010, `browser_signature_banned`: the edge had blocked the client's signature before the request ever reached Kody, so no token, scope, or grant diagnosis applied. See Troubleshooting. 3. **A Muse entry that was the wrong host.** The Agents page showed a connected Muse entry, which first read as proof that Muse chat was granted. Its connection instructions were for Muse Code. A connection marked connected, and an Agents entry, do not prove that Muse chat tools work; only the proof step in step 3 does. ## Troubleshooting ### Invalid redirect URI at authorize This is a pre-consent check: Kody compares the `redirect_uri` Muse sends against the Redirect URIs saved for that exact Client ID. If none matches exactly, authorize fails before any consent screen with `Invalid redirect URI. The redirect URI provided does not match any registered URI for this client.` Fix: create a client with both `hatch` callbacks from step 1, and enter that client's ID and secret in Muse's Secure credentials store. Kody's MCP OAuth clients cannot be edited, so if the saved redirect is wrong, select **Revoke**, then **Revoke now**, on the old client under **Your clients** and create a new one rather than trying to change it in place. ### HTTP 403 on list-tools Read the error body before diagnosing scope, audience, or grants. A 403 from Cloudflare, not from Kody, has a body like `error code: 1010` in plain text, or JSON like this: ```json { "type": "https://developers.cloudflare.com/support/troubleshooting/http-status-codes/cloudflare-1xxx-errors/error-1010/", "title": "Error 1010: Access denied", "error_name": "browser_signature_banned" } ``` That error means Cloudflare blocked the request based on the client's signature before it ever reached Kody. A library default User-Agent (for example Python's `Python-urllib/3.x`) can trigger it. Cloudflare recommends a descriptive User-Agent that identifies the client instead. If Muse does not provide a way to configure that header, contact the deployment owner to review the Cloudflare rule; do not spoof a browser User-Agent to evade the block. Once the error body is from Kody rather than Cloudflare, the usual MCP OAuth distinction applies: 401 means a missing, invalid, or expired token (Kody's 401 carries a `WWW-Authenticate: Bearer resource_metadata=...` header), and 403 means a valid token that lacks the required scope or resource audience. ### The connection is marked connected, but tools still fail Check, in order: the error body (Cloudflare vs Kody, above), whether the client is the one whose redirect matches the `hatch` callbacks, and whether the Agents page entry you are reading is for Muse chat or for Muse Code. A Muse Code entry does not prove a Muse chat grant. ## Security notes - Never put a Client Secret, token, password, or cookie in a doc, a helper, a command argument, an environment variable, or a log, and never paste one into chat. The Secure credentials store is the only place for it. - A redirect URI from a browser address bar is not a secret and is safe to share when diagnosing a redirect mismatch. - Keep `execute` and `api` out of the allowlist until search is proven and you have separately decided you want writes. Approving an agent already gives full account access; the allowlist is the second, narrower gate you control.