mcp-chainladder
Enables to perform actuarial chain-ladder reserving calculations, including IBNR estimation, ultimate projections, Mack stochastic error estimates, and diagnostic tests, through natural language conversation with Claude.
README
mcp-chainladder
<!-- mcp-name: io.github.datalattice/mcp-chainladder -->
Actuarial chain-ladder reserving for Claude. A Model Context Protocol server that hands Claude the tools to compute IBNR, project ultimates, run Mack (1993) stochastic error estimates, check Mack (1994) model assumptions, and parse loss triangles from CSV — all from a natural- language conversation.
Calculation only — not actuarial advice. This is a mechanical chain-ladder calculator. Selection of data, factors, tails, and exclusions, and the interpretation of stochastic error measures, require professional actuarial judgement.
What it does
Eight tools, exposed over MCP. Claude picks the right one when you ask a question; you don't have to call them by name.
| Tool | When Claude reaches for it |
|---|---|
compute_chain_ladder |
"What's the IBNR on this triangle?" — the workhorse |
project_triangle |
"What does the full projected triangle look like?" |
mack_stochastic |
"What's the uncertainty on the total reserve?" |
mack_diagnostics |
"Are there outliers or trend issues?" |
parse_csv_triangle |
"Run the chain ladder on this CSV file" |
to_incremental |
"Show me the incremental development pattern" |
to_cumulative |
"Cumulate these incremental values" |
sample_triangle |
"Show me a working example" — quick demo |
All numerical conventions match the Mack 1993 and Mack 1994 papers; results are stable to ~1e-12 against hand calculation on the textbook 10×10 triangle.
Install
pipx install mcp-chainladder
# or, to run on demand without installing:
uvx mcp-chainladder
Python 3.10 or newer; no other runtime dependencies beyond the official MCP SDK.
Hook into Claude Desktop
Edit ~/Library/Application Support/Claude/claude_desktop_config.json
(macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"chainladder": {
"command": "uvx",
"args": ["mcp-chainladder"]
}
}
}
Quit and reopen Claude Desktop. The triangle tools appear under the 🔌 icon in any conversation.
For Cursor / Cline / Continue / Zed — same mcpServers block, each app
has its own config file location (their docs walk you through it).
Examples
Run the chain ladder on a CSV
Me: I have a paid-loss triangle at ~/Desktop/motor_2026Q1.csv.
Run the chain ladder and tell me the IBNR.
Claude: [calls parse_csv_triangle, then compute_chain_ladder]
Loaded a 10×10 cumulative triangle. Using volume-weighted
factors:
• Total paid to date: 49,458,000 USD
• Total ultimate: 65,883,400 USD
• Total IBNR: 16,425,400 USD
Selected factors trend from 1.879 at the front of the
triangle down to 1.002 at the tail.
Check whether you should worry about assumptions
Me: Same triangle — any reason not to trust the projection?
Claude: [calls mack_diagnostics]
Three flags worth a look:
• Inflation trend: slope 0.053, p ≈ 1e-12 — STRONG evidence
link ratios are drifting upward across accident years.
This violates Mack's homogeneity assumption; consider
weighting recent years more.
• Calendar-year: Z = 0.0, p = 1.0 — no diagonal effect.
• Independence: Z = -0.35, p = 0.73 — no obvious dependence
across development columns.
• Outliers: 0 cells exceed |2σ|.
The inflation finding is the one to act on.
Exclude an outlier and re-run
Me: Drop the link ratio at AY 3, dev period 2→3. What changes?
Claude: [calls compute_chain_ladder with excluded=[[2, 1]]]
Excluding (AY 3, 2→3):
• Volume factor at 2→3 moves from 1.331 to 1.328
• Total IBNR moves from 16,425 to 16,389 (−36)
• Mack SE total moves from ±354.6 to ±348.2
Net: small enough to be a "robustness check passes" rather
than a finding.
Tool reference
Each tool returns a JSON object (or a 2-D list, in the case of
project_triangle/to_incremental/to_cumulative). Claude reads the
descriptions and types directly from the server — you don't need to
memorise the shapes — but here's the cheat sheet.
compute_chain_ladder(triangle, selected_factors?, tail?, excluded?)
End-to-end chain ladder.
| Field returned | Meaning |
|---|---|
volume_factors[j] |
All-year volume-weighted age-to-age factor for transition j→j+1 |
simple_factors[j] |
Unweighted average of individual link ratios |
selected_factors[j] |
The factor set actually used to project (defaults to volume) |
individual_factors[i][j] |
Per-row link ratio C[i,j+1] / C[i,j]; null where the pair is unobserved |
cdf[j] |
Cumulative dev factor to ultimate; cdf[-1] == tail |
latest_diagonal[i] |
Most recent observed value per AY |
ultimates[i] |
Projected ultimate per AY |
ibnr[i] |
Ultimate − Latest per AY |
total_* |
Sums of the three above |
n_acc, n_dev |
Triangle dimensions |
mack_stochastic(triangle, selected_factors, excluded?)
Mack (1993) distribution-free standard errors. Returns σ̂²_j per dev period (tail-rule backfilled when only one observation), SE & CV per row, and SE_total / CV_total including cross-row covariance per eq. 5.15.
mack_diagnostics(triangle, selected_factors, excluded?, outlier_threshold?)
Returns standardised residuals, outliers (|r| > threshold, default
2.0), the calendar-year sign test, Spearman independence test across
adjacent dev columns, and the inflation slope of mean log-link-ratio
against accident-year index.
p-value bands to translate to plain English:
p < 0.005 → strong evidence
p < 0.05 → significant
p < 0.10 → borderline
p ≥ 0.10 → no evidence
parse_csv_triangle(path)
Reads a CSV from disk, treats blank / NA / NaN / N/A / − cells as unobserved, strips embedded thousand-separator commas, and skips header / metadata rows. Returns the triangle + dimensions + the absolute path read (useful for the assistant to confirm what it loaded).
project_triangle(triangle, selected_factors)
Fills the lower-right of the triangle with chain-ladder projections.
Returns a fully-rectangular list[list[float]] (no nulls). NaN where
a row has no observation to project forward from.
to_incremental(cumulative) / to_cumulative(incremental)
Two conversions. Unobserved cells stay unobserved; the inverse on observed cells is exact.
sample_triangle()
Returns the textbook 10×10 cumulative paid triangle. Use as a
self-check: pass it to compute_chain_ladder and you should get
Paid 49,458 / Ultimate 65,883 / IBNR 16,425.
Triangle format
[
# AY 1 — fully developed
[1000, 1855, 2423, 2988, 3335, 3483, 3552, 3603, 3624, 3631],
# AY 2 — observed through dev 9
[1113, 2103, 2774, 3422, 3844, 4010, 4090, 4148, 4172, None],
# …
# AY 10 — only the first observation
[2640, None, None, None, None, None, None, None, None, None]
]
- Outer index = accident year, oldest first
- Inner index = development period, 0 = first age
- Use
None(or JSONnull) for unobserved cells - All rows must be the same length — pad with trailing
null
Testing
pipx install --editable mcp-chainladder
pytest -q
Tests pin every public tool against the textbook triangle's well-known parity values to ~1e-9.
Pro tier
<< UNDER REVIEW - COMING SOON >>
The free tier covers all 8 tools listed above. Pro unlocks
additional methods + bulk workflows, gated by a local license file
at ~/.chainladder/license (or wherever $CHAINLADDER_LICENSE_FILE
points).
Pro licenses are currently for internal and testing purposes only — not open to the public. No purchase channel is available at this time. The Pro tools are listed below for reference and will continue to return
pro_license_requiredfor external users.
| Pro tool | What it does |
|---|---|
pro_license_status |
Inspect current license state (free to call) |
interpret_diagnostics |
Mack tests with verdict labels + plain-English summaries + recommended actions |
sensitivity_analysis |
Drop each link ratio one-at-a-time and rank by IBNR impact |
tail_extrapolation |
Fit exponential + inverse-power tail models, recommend best fit |
bornhuetter_ferguson |
BF reserving method with side-by-side CL comparison |
compare_methods |
Run CL + BF in one call, report deltas + largest divergence |
generate_pdf_report (coming v1.2) |
Full 5-page actuarial PDF — cover / triangle / factors / results / 3D loss surface |
batch_csv_processing (coming v1.2) |
Fold the chain ladder over a directory of CSV triangles |
cape_cod, mack_bf (coming v1.3) |
Additional reserving methods, all returning side-by-side comparisons |
License file format
{
"product": "mcp-chainladder-pro",
"owner": "alice@example.com",
"expires": null,
"key": "CL-PRO-1A2B3C4D",
"signature": "…"
}
Drop it at ~/.chainladder/license (the file's directory must exist;
the server doesn't create it). Pro tools immediately respond as
unlocked the next time Claude calls them.
When the license is missing or expired, every Pro tool returns
{"error": "pro_license_required", "status": {...}} instead of
computing — Claude reads the status and points you to the upgrade
URL. The free-tier tools always work regardless of license state.
License
MIT. See LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.