Integrations

LLM API keys, OAuth connections, MCP servers, and persistent browser sessions.

LLM API keys

Backbend doesn’t bundle inference. You bring your own provider keys. Add at least one before you can chat with an agent.

Supported providers

  • Anthropic (Claude): the default and the most tested. Recommended for almost all use cases.
  • OpenAI (GPT-4, GPT-5): works for chat and tool use.
  • Google (Gemini): supported, lighter testing.
  • OpenRouter: works as a fallback or cost-shaping provider, gives access to dozens of models behind one key.

Adding a key

Open Dashboard → LLM, click + Add key, pick the provider, paste the key, and save. The key is encrypted at rest with a Fernet key derived from your account secret; the browser never sees the plaintext again after you save.

Choosing a default model

Each agent has a default model in its Settings. When you call the agent without overriding the model, this is what runs. The in-chat model picker lets you override per-message; tasks pick the agent default unless overridden in the task config.

Rotation

To rotate, delete the old key and add a new one. Agents pick up the new key on the next message; no restart needed.

Lost your account secret? You’ll have to re-enter every key (we can’t decrypt them). Same for any connection field stored encrypted.

Cost tracking

Every model call is priced at the provider’s public rate and rolled into the agent’s daily spend on the home page. You can also export per-task token usage as CSV from the Tasks page.

Connections

Connections are OAuth links to the apps your agent needs to touch. One approval gives the agent persistent access; no copy-pasting tokens, no re-authing every week.

Supported services

  • GitHub: read/write repos, issues, PRs, actions.
  • Google: Gmail, Calendar, Drive, Sheets, Docs.
  • Slack: read channels and DMs, post messages, react to messages.
  • Notion: pages, databases, search.
  • LinkedIn: profile, messaging, posts.
  • X / Twitter: read your timeline, post, DM.
  • Custom OAuth: bring your own client_id/client_secret for any OAuth 2.0 provider not in the list.

Connecting

Open Dashboard → Connections, find the service, hit Connect. A popup opens the provider’s OAuth flow; once you approve, the row switches to a green “Connected” state. Disconnect any time with the menu on the right.

Agent vs account scope

Connections live at the accountlevel. Every agent in your workspace can use them by default. To restrict which agents see which connections, open the agent’s Settings tab and toggle access per connection.

Tokens

OAuth refresh is automatic. If a connection goes stale (the provider revoked the grant, the user rotated their password, the app got uninstalled), the row turns red and the agent stops using it until you reconnect.

MCP servers

Backbend supports the Model Context Protocol as a first-class integration surface. Any MCP-compatible server (yours or a third party’s) plugs in and exposes its tools to your agent just like a built-in tool.

Why MCP

MCP is an open protocol for “here’s a tool, here’s what it does, call it like this.” The ecosystem includes servers for databases, ticketing systems, file storage, custom SaaS apps, and dev tools. Hook one in and the agent can use it without you writing a single line of integration code.

Adding an MCP server

  1. Open Dashboard → Connections and switch to the MCP servers tab.
  2. Click + Add server. Pick a transport: http (a hosted MCP server with an HTTPS URL) or stdio (a local binary we shell out to inside the agent container).
  3. Fill in the URL or command, plus any auth headers / environment variables the server needs. Sensitive values are stored encrypted.
  4. Hit Test. We probe the server, list its tools, and show them inline. If the connection succeeds, hit Save.
  5. Open the agent’s Settings, find the MCP server row, and toggle access. Tools become available on the agent’s next turn.

Popular MCP servers people use

  • Filesystem MCP: give the agent read/write access to a scoped directory tree.
  • Postgres MCP: query your database with named, scoped queries instead of raw SQL.
  • Linear MCP: full Linear API surface.
  • Sentry MCP: investigate errors, link issues to PRs.

Building your own

The MCP spec lives at modelcontextprotocol.io. If you ship a server that’d be useful to other Backbend users, email us and we’ll feature it in the Connections list.

Persistent browser sessions

Some sites don’t expose an OAuth surface that covers what you need, or the OAuth scope they offer is too narrow. For those, the agent can use a persistent browser session: log in once through the agent’s browser, and the session is stored encrypted on our servers and reused on every later run.

How it works

  1. Open Dashboard → Connections and switch to the Browser sessions tab.
  2. Click + New session, give it a name (e.g. “HackerNews”), and hit Open browser.
  3. A live browser window opens in your tab. Navigate to the site, log in normally, do whatever 2FA dance is required.
  4. When you’re logged in, hit Save session. The cookies, local storage, and any saved auth state are captured and encrypted at rest.
  5. Reference the session by name from any agent that has browser-use turned on. The agent reuses it transparently.

When sessions expire

Sessions live until the site invalidates them. Most last weeks; some last days. When a session expires, the row turns yellow and the agent falls back to OAuth (if available) or fails the tool call. Re-login the same way you set it up; the saved session ID stays the same so existing tool calls just start working again.

We never share browser sessions across accounts. Each user gets an isolated browser profile per session.