Flatland vs.
Pigment.
Pigment is an enterprise FP&A platform with a proprietary modeling language and, as of 2026, an MCP server agents can query. Flatland is a typed reasoning substrate AI agents call to construct models from scratch. Both expose MCP. They sit at different tiers of the stack.
Pigment · founded 2019 · enterprise FP&A platform · MCP server and AI Modeler agent shipped in the 2025-2026 product cycle. · Flatland · 2026 · MCP + REST day one · metered billing · typed IR.
> agent calls pigment.query_model(
model_id="fy26_plan",
cell="revenue.q3_2026"
)
← $4.2M
# The model already exists in Pigment.
# The analyst built it. The agent reads it.Pigment's MCP server is a window into models that have already been authored, typically by FP&A analysts in Pigment's UI. The agent queries cells, runs scenarios on existing structure, asks questions of already-built logic.
> agent calls flatland_build_model(
drivers=[ /* typed IR */ ],
assertions=[ /* guardrails */ ]
)
> agent calls flatland_compile()
← { values: {...}, assertions: {...}, excel_url: ... }
# No model existed yet.
# The agent authored it. The compiler validates it.Flatland's MCP server accepts typed IR from the agent and compiles it into a deterministic model with types, assertions, and a real Excel export. The model is built in the conversation; the conversation is the authoring surface.
Both products ship an MCP server. Both products integrate with Claude, ChatGPT, and the rest of the agent ecosystem. The distinction is what the agent does: in Pigment, it queriesan FP&A analyst's authored model. In Flatland, it authors a typed model the compiler then validates.
You're a 50-to-5000-person company with a finance team that already has an FP&A platform budget and wants the modern Anaplan alternative. You need broad native connectors to your ERP / CRM / HRIS that you don't want to engineer. Your FP&A analysts are the authors and your finance team is the audience; the agent is a productivity overlay, not the primary builder. Pigment is built for that team shape, and it shows.
Your agent is the author. You want the model to be a typed compilation artifact, not a workbook. You'd rather pay per answer than per seat. Your auditor and investor want Excel, but you want the source of truth to be IR a developer can read.
Both ship MCP.
Only one of us builds the model.
Compile a typed model.
See where your agent draws the line.
Three questions. A real compile. A real Excel export. If you've been wondering what an agent-authored model looks like with types and assertions underneath: this is the answer in sixty seconds.