Hey!
It’s been a few months since I last used Foundry, and I’m noticing lots more tools available. Like new generations of AIFDE, Pilot, improved MCP, etc.
I’m wondering if there is a single point of reference about how they all link together so we can best make use of them. I’m particularly interested in agentic workflows working symbiotically with agents such as Claude Code // Codex.
Moreover, what is Pilot and how should we think about it vs AI FDE and productisation?
Thanks!
Hey Jpa, the surface really has expanded, here is the cleanest map I can offer although everything requires nuance and with LLM, these different views can live together and compound, this is just a simplified mental model to help you organized your thoughts and share some of our experience.
Inside Foundry
Two distinct agents, plus the analyst layer:
AI FDE is a general purpose, conversational Foundry agent. It runs in a closed loop (executes, observes, iterates) and can do data integration (Python transforms, Pipeline Builder), ontology editing, functions (Logic, TypeScript, Python with AIP Evals), governance audits, platform Q&A, and yes also OSDK React app and custom widget building. Branching is on by default, so every change goes through a Global Branch proposal or a Code Repository PR before merge. Model support is first-class for Anthropic, OpenAI, Google, and xAI, with native tool APIs, and you choose the model per session, along with which tools and which data the agent can touch.
Pilot (beta) is another philosophy. You give it one prompt and it generates an end to end OSDK plus React hosted application, with specialised sub-agents for ontology generation, a full design specification (palette, typography, layout, interaction patterns), the React frontend with OSDK hooks, seed data in an isolated sandbox, and the Developer Console deployment with CI and a subdomain. Same model families available (Claude, GPT, Gemini, depending on enrollment). Ontology still promotes via Global Branching and proposals, so governance applies. Output target is fixed: a production-grade hosted app.
Outside Foundry
Local agentic IDE (Claude Code/Design, Codex, Cursor) with MCP Foundry connected, exposing the ontology in both directions: read objects, query via OSDK, write through Actions, run Functions, search. Add one or more org MCPs alongside (knowledge base like Notion or Confluence, issue tracker like Linear or JIRA, calendar, CRM, code repo, observability), and the IDE agent reasons across all of them (that’s the real power of it, you can have an environment accessing a quite large context landscape). Note: From that standpoint the MCP can be understood both as a lever for your developer toolkit but also as an asset that you can use to fuel external to foundry workflow (we’re using it at Sibyl to interact with our internal operations ontology - HR, Finance, Ops and fuel internal operations)
Nuance: inside and outside compose
The inside/outside split is a useful first cut but it is a simplification. In practice the two are coupled in the same project or be mixed (e.g. a claude design hand-off file can perfectly be plugged to AI-FDE).
No single “official map” page that ties it all together yet, but the three lenses I would suggest are: direction of integration (inside vs outside, with the caveat that they compose), output target (Foundry asset vs hosted OSDK app), and degree of orchestration (single asset, end to end app, or cross-tool workflow). Happy to compare notes on a concrete use case.
Thank you! Very helpful reply.
One follow up question:
Do you think we someday will be getting AI FDE (i.e. Ontology change capability) level functionality in Palantir MCP or similar? To work on branches I mean
Hi @jpa,
I dont have a good answer for this and this is more a question for product teams at Palantir as to how they are envisioning complementarity between AI-FDE and MCP.
That said, my feeling is that AI-FDE is probably poised to be benefiting from more specific, tailored and rich context of Foundry than the MCP, but this is just an opinion, not an evidence-based fact !
Cheers
Nicolas @ Sibyl