- Starting locally with the desktop app helps is a great way to figure out Coral out before investing in the shared server deployment. The shared server needs you to bring your own IdP, among other things.
- Don't design custom tools up front. Let agents discover what they need, but keep the human in the loop to review additions/updates.
- Decide how you want your users to authenticate with the upstream provider (shared token vs. as themselves). Most prefer the latter when possible, and accept users will need to complete a one off OAuth flow for each provider. We're working on ID-JAG support to streamline this step.
I work for Coral but in my personal opinion, it's great how easy is to integrate Coral with my own workflow without having to make big changes to it. It makes a breed to add external services and manage agents permissions
Got really confused by the name. Apparently Google has a "Coral platform" which used to ship TPUs as USB devices (although that's discontinued, and Google barely markets the Coral brand these days). If you search "Coral AI" there are tons of products/companies called by that name, and this one is not nearly at the top. Maybe something to think about.
What would be great too is a library for mini apps built with agents. I find these little apps to be useful all the time, but despite usually being written in very versatile HTML that is cross-platform, these things are confined to the specific platform I used and usually the chats too. I wish there was a provider agnostic way to build and view mini apps anywhere (desktop and mobile).
It's certainly something we've been thinking about too.
At the moment, we see teams solve this in two opposite ways:
- building their own generic artefact host (essentially a host for the html/css) and letting users (employees) create whatever they want. Coral handles all the data access and policy on the backend
- the other way is a much more limited form factor, essentially a set of HTML templates, for creating dashboards or mini widgets. Again, all data access is offloaded to Coral.
I appreciate you keeping the order, but that's false. We let preview users know we’d announce the product today and it looks like a handful of them created new accounts. I apologise if that was against the guidelines.
> Don't solicit upvotes, comments, or submissions. Users should vote and comment when they run across something they personally find interesting—not for promotion.
Coral ships with a UI, which includes a page for all pending and past approvals.
How to prompt a user/agent to visit this page is a fun on. Short version: we started with MCP elicitation, but harnesses today implement it wildly differently. Claude Desktop has had an issue open since August where it just auto answers "decline" on every elicitation (https://github.com/anthropics/claude-code/issues/89858). Under some settings, Codex does the opposite.
Now we return the URL in the tool call response as the next action, and where supported, send a prettier approval dialogue as an MCP App.
disrpt | 9 hours ago
[OP] jcs | 8 hours ago
greenBottle | 9 hours ago
jurgenaut23 | 9 hours ago
[OP] jcs | 8 hours ago
ocamoglu | 9 hours ago
Looks like this is already used in real world by your customers. Any learnings you can share for deployment?
[OP] jcs | 9 hours ago
- Starting locally with the desktop app helps is a great way to figure out Coral out before investing in the shared server deployment. The shared server needs you to bring your own IdP, among other things.
- Don't design custom tools up front. Let agents discover what they need, but keep the human in the loop to review additions/updates.
- Decide how you want your users to authenticate with the upstream provider (shared token vs. as themselves). Most prefer the latter when possible, and accept users will need to complete a one off OAuth flow for each provider. We're working on ID-JAG support to streamline this step.
ocamoglu | 9 hours ago
nikolicmilena | 9 hours ago
[OP] jcs | 8 hours ago
antonmry | 8 hours ago
latissi | 8 hours ago
fg137 | 8 hours ago
[OP] jcs | 8 hours ago
anonyonoor | 8 hours ago
rkeene2 | 8 hours ago
[OP] jcs | 7 hours ago
At the moment, we see teams solve this in two opposite ways:
- building their own generic artefact host (essentially a host for the html/css) and letting users (employees) create whatever they want. Coral handles all the data access and policy on the backend
- the other way is a much more limited form factor, essentially a set of HTML templates, for creating dashboards or mini widgets. Again, all data access is offloaded to Coral.
Both work really nicely.
jrflo | 8 hours ago
wavemode | 8 hours ago
[OP] jcs | 8 hours ago
jasonjmcghee | 6 hours ago
> Don't solicit upvotes, comments, or submissions. Users should vote and comment when they run across something they personally find interesting—not for promotion.
[OP] jcs | 2 hours ago
eb379 | 7 hours ago
[OP] jcs | 7 hours ago
How to prompt a user/agent to visit this page is a fun on. Short version: we started with MCP elicitation, but harnesses today implement it wildly differently. Claude Desktop has had an issue open since August where it just auto answers "decline" on every elicitation (https://github.com/anthropics/claude-code/issues/89858). Under some settings, Codex does the opposite.
Now we return the URL in the tool call response as the next action, and where supported, send a prettier approval dialogue as an MCP App.