Reliafy

Use Reliafy from Claude (MCP)

Reliafy runs a remote Model Context Protocol server, so Claude — on the web, the desktop and mobile apps, or Claude Code — can work in your Reliafy account for you: fit distributions, read your models, build and analyse RBDs, and run the maintenance calculators. It sees what you see: your own data plus the shared samples. Self-hosted instances allow every tool for everyone.

Free to try, unlimited on Pro

On Reliafy Cloud, using Reliafy from AI agents (MCP) is part of Pro (US$19/month): every tool, no limit and unlimited storage. Free gives you 20 tool calls a month to try it — every tool except fitting, fleets and availability simulation, within the Free storage limits; the calls reset on the 1st of each month (UTC). Past them, each tool says so, and upgrade_link gives you a link to subscribe.

Claude on the web, desktop and mobile

In Claude, open Customize → Connectors (claude.ai/customize/connectors), choose Add custom connector, enter https://reliafy.com/mcp, then choose Connect and sign in to Reliafy. Reliafy asks you to approve the connection; you can disconnect it any time under Settings › Connected apps. Connectors added on claude.ai are available in the desktop and mobile apps too.

Claude Code

claude mcp add --transport http reliafy https://reliafy.com/mcp

Then run /mcp in Claude Code, pick reliafy and sign in — a browser window opens for you to approve access. After that, ask something like “fit a Weibull to these bearing lives” or “build an RBD of my pump skid and tell me its MTTF”. claude mcp list shows the connection.

Prefer a token (scripts, CI, headless machines)? A personal API token (rlf_…) still works — sign in and create one under Settings › API access. Give it the Read scope for the reading and calculating tools, and Write too for the ones that save, change or delete.

claude mcp add --transport http reliafy https://reliafy.com/mcp --header "Authorization: Bearer rlf_YOUR_TOKEN"

Other MCP clients

Any client that supports Streamable HTTP can connect to https://reliafy.com/mcp — with OAuth (discovered automatically from the server; dynamic client registration and client ID metadata documents are both supported), or with a token header:

{
  "mcpServers": {
    "reliafy": {
      "type": "http",
      "url": "https://reliafy.com/mcp",
      "headers": { "Authorization": "Bearer rlf_YOUR_TOKEN" }
    }
  }
}

Sending files

Files don’t need to go through the conversation. uploads (action create) returns a single-use upload URL; Claude sends the file there from its shell, then imports it by its upload_id:

curl -X PUT --data-binary @plant.rsgz \
  -H "Content-Type: application/octet-stream" \
  "https://reliafy.com/api/uploads/UPLOAD_ID?t=TOKEN"
  • The link works once, expires after 15 minutes and takes at most 30 MB. Its answer, errors and privacy: Files and photos.
  • The upload URL points straight at Reliafy’s app service, which is fastest. Some sandboxes and corporate proxies refuse that host (for example CONNECT tunnel failed, response 403). For those, uploads create also returns url_fallback (with a ready curl_fallback): the same link on reliafy.com. Send the same PUT there once. It’s slower for large files. A PUT that a proxy refused never reached Reliafy, so the link is still unused.
  • If Claude’s environment can’t reach either URL (no network), small text files (up to 200 KB) can be pasted instead — CSV into datasets upload or rbd upload_outage_log, Open-PSA or Galileo text into uploads import_rbd. Other files of up to a few hundred KB (an Excel workbook, a BlockSim project) can go through the MCP connection itself with uploads send_part: base64 parts into the same upload, checked against the file’s SHA-256. An FMEA / RCM worksheet can also be given as data with rcm create. Anything larger is uploaded in the app.

Tools

A few tools, each doing one area’s work: an agent picks the tool, then the action (or, for get and list, the kind). Each tool’s description lists its actions and what each takes, and guide gives the agent how-tos by topic.

guideShort how-tos by topic — life data and censoring, units, accelerated tests, repairable systems, building RBDs, analysing them, strategy, fleets, uploads, RCM, sharing, the workspace, plans — so the agent reads the one it needs before a task. Never counted against the allowance.
getOne saved item in full, by kind and id.
  • model — A saved life or recurrent model — distribution, parameters with confidence intervals, goodness of fit, MTTF and B-lives with lower bounds. Regression models add how good the model is (Harrell’s C, the integrated Brier score, the time-dependent AUC) and whether it holds (the proportional-hazards test with a plain verdict; for Cox PH robust standard errors, the tie method and strata). A model fitted by failure mode reports each mode’s share with 95% intervals and Gray’s test. Recurrent models give the growth verdict, ROCOF and MTBF with bounds and the trend tests.
  • dataset — A dataset’s columns and rows, a page at a time.
  • rbd — A diagram’s structure.
  • rbd_run — One availability simulation in full — its results with their confidence intervals, the inputs and engine versions it ran on, the precision it achieved and what was quoted before it ran.
  • job — A calculation queued on the calculation service (a long simulation, a what-to-improve or interval search, an ALT’s bootstrap bounds) — its place in the queue, then the result.
  • rcm_study — An RCM study’s worksheet — functions, functional failures and failure modes with effects, consequences and decisions — with every node’s id and each decision’s evidence check.
  • upload — An uploaded file before importing it: a workbook’s sheets with their header rows, columns, row counts and a guessed mapping; a diagram file’s diagrams and block counts.
  • account — Your plan; on Free, the tool calls used and left this month and the date they reset (Pro is unlimited); storage used against your plan’s limits, and whether simulation is included. Never counted against the allowance.
listYour saved items of one kind, with the shared samples.
  • models / datasets / rbds / rcm_studies — Your models (life or recurrent), datasets, diagrams and RCM studies.
  • rbd_runs — Simulation history, from the app or any connector: who ran each run and from where, its status, method, runtime and result. Kept 7 days on the free plan and 90 on Pro.
  • fleets / fleet_alerts — Your fleet forecasts and a fleet’s email alerts (Pro).
  • share_links — Your live share links.
  • workspace_changes — What changed in your workspace over a period — the last 30 days by default, a number of days, a date, or since your last visit — worked out from what Reliafy stores, never estimated: new and changed items; refits whose B10 or MTTF moved by more than 10%; fleet forecasts that moved and alerts that fired (Pro); RCM decisions whose linked model was refitted; simulations that moved. It says what it can’t compare.
life_data_analysisLife data analysis that saves nothing.
  • fit — Fit a distribution (or “best”) to failure times given inline or from a saved dataset — every model id, from Beta and Royston-Parmar to Buckley-James; shared frailty with group_column; Cox PH with cox_options; covariates that change over time with id_column; by failure mode with causes and “competing_risks”. A fit whose data don’t pin the model down says so (fit_ok: false) so its numbers aren’t quoted. With “best”, include_mixtures lets two-mode mixtures compete. Fitting is Pro.
  • fit_per_demand — Per-demand (one-shot) reliability — a valve that must open, a standby start — from batches of demands and failures, with exact (Clopper-Pearson) bounds; zero failures is the success-run demonstration.
  • reliability_at — Reliability, failure probability and hazard of a saved model at given times — for a unit that has already survived to some age too, with confidence bounds where the model has them. A regression model is evaluated at your covariates (or along a covariate schedule), with its B-lives and their lower bounds.
  • check_life_requirement — Does a saved life model meet a B-life requirement — “B10 ≥ 50,000 cycles at 90% confidence”? Decided on the B-life’s one-sided lower bound, never the estimate.
  • compare_groups — Is A better than B? The log-rank test, each group’s average life over a common window (RMST) with the difference and its interval, Gray’s test per failure mode, and a one-line verdict — without fitting.
  • alt_use_level — Extrapolate a saved accelerated-life model to the use stress: characteristic, mean and B-lives, the mission reliability and the acceleration factor, with Wald, likelihood-ratio or bootstrap bounds (bootstrap: Pro, as a job).
  • plan_demonstration_test — Plan a reliability demonstration test — units, test time per unit and allowed failures — with a trade-off table; with producer_risk a plan that keeps both risks; the Weibull shape from a saved model if you like.
repairable_analysisRepairable systems from a saved recurrent model.
  • next_failure — A system’s next failure at its current age: the chance of a failure within a time (say, before its next service), and the mean and percentiles of the time to it.
  • growth_projection — Reliability growth projection (MIL-HDBK-189C, AMSAA-Crow): the MTBF a growth test reaches once its delayed fixes are in. Inline data is Pro (it fits).
strategyMaintenance strategy calculators and fleet forecasts.
  • optimal_replacement / failure_finding_interval — Cost-optimal replacement interval, and proof-test interval for a hidden function — for a saved Weibull fitted to data, at each end of the shape’s 95% interval too, with a note when that makes the recommendation less firm.
  • optimal_overhaul — Optimal overhaul interval from a minimal-repair recurrent model.
  • fleet_forecast — Expected failures across a fleet of in-service items, with exact P10–P90 intervals per period and each item’s chance of failing; regression and ALT fleets at each item’s covariates or stress; recurrent fleets count every repeat failure. Pro.
rbd_analysisAnalyse a saved reliability block diagram; nothing is saved.
  • analyze — System reliability, MTTF, B-lives and importance; for a repairable diagram exact availability over time, the window’s expected failures, downtime and cost, a safety function’s PFDavg and SIL, a priced diagram’s cost of ownership and capacities’ production availability — from new or from the blocks’ state now. The Monte-Carlo simulation adds the spread and the time to the next failure (Pro; a long one comes back as a job).
  • quote_simulation — How long a simulation should take, before anything runs.
  • sensitivity — What to improve first in a repairable diagram: every lever moved a step, ranked by the gain in availability or cost, in plain words — by benefit per unit of cost when you price the changes — plus over_time, shares, joint, rate and uncertainty measures.
  • optimise_intervals — Choose a repairable diagram’s age-replacement or proof-test intervals together (lowest cost, a target availability, PFDavg or SIL, or a cost cap), staggering tests if you like; returns the edits that apply the plan.
  • cheapest_design — The redundancy that owns a repairable diagram at the lowest total cost, whole trains included, over several horizons. Pro.
  • compare — Is design B better than design A? The difference in availability (and cost of ownership) over a window. Pro.
  • fault_tree — The diagram as a fault tree: top event, ranked minimal cut sets, basic events by importance and each common-cause group’s share. Exact and free.
  • allocate_targets — What reliability, availability or MTTF / MTTR each block needs for the system to meet a target. Exact and free.
  • system_history — The observed availability from an outage log, each outage with the block that caused it, and the blocks ranked by downtime share.
  • export_python / export_json — A standalone SurPyval + RePyability script, or RePyability’s JSON (rbd_from_json loads it; uploads import_rbd brings it back).
modelsSave and change models (Claude asks before running it).
  • fit — The same fit as life_data_analysis fit, saved as a model; with demand_batches, a per-demand model. Pro.
  • save — Save a model from a distribution and parameters your agent fitted itself — e.g. locally with SurPyval.
  • fit_alt — Fit and save an accelerated life test model — failure times at several stress levels and a life-stress relationship, inspection data and delayed entry included. Pro.
  • fit_recurrent — Fit and save a recurrent-event (repairable-system) model: Crow-AMSAA and the other minimal-repair models, proportional intensity, or imperfect repair (Kijima I/II, ARA, ARI, G1). Pro.
  • update / delete — Rename or annotate your own models, or delete one (never shared samples; not one a fleet forecast runs on).
datasetsSave and change datasets.
  • upload — Save CSV text, or a CSV or Excel file sent with uploads create.
  • update / delete — Rename or annotate your own datasets, or delete one (not one a model was fitted to).
rbdBuild and change reliability block diagrams.
  • create — Build and save a diagram — series, parallel, k-of-n, standby, sub-systems — that opens in the RBD builder: mixtures of failure modes, imperfect repair, common-cause groups (beta factor or multiple Greek letter), capacities and a demand, phased missions, networks, covariate schedules, costs, scheduled and condition-based replacement and proof tests.
  • clone / edit — Copy a sample or one of your diagrams, and change a diagram with a short batch of edits, all or nothing, instead of rebuilding it.
  • delete — Delete one of your own diagrams.
  • upload_outage_log — Save a real outage log (asset, down, back up) — CSV text or an uploaded file — against one of your diagrams.
uploadsFiles into Reliafy without pasting them.
  • create — A single-use link (15 minutes) to send a file with a plain HTTP PUT — a ready curl command — with a reliafy.com fallback URL for networks that block the direct one. Up to 30 MB; deleted after its import, or after an hour.
  • send_part — For an agent that can’t reach the upload URL at all: the file in base64 parts through the MCP connection, into the same upload. For files up to a few hundred KB; not counted against the Free allowance.
  • import_rbd — Import diagrams from ReliaSoft BlockSim, Open-PSA XML, Galileo DFT, RePyability JSON or Reliafy’s Excel RBD template, with notes on anything approximated; preview first to check the structure.
  • import_excel — A sheet as a dataset, an FMEA / RCM worksheet into a new or existing RCM study, or the RBD template as a diagram (as a copy of an existing one, unless you ask to replace it).
rcmRCM studies as structured data.
  • create / update — Save an FMEA / RCM worksheet given as data — the same study the RCM editor or an Excel import makes, with the same checks and plan limit — or add to and change one by node id, replace its worksheet, or set its name, system and description.
sharingRead-only pages anyone with the link can open, no account needed.
  • create_link — Publish one of your own items — a model, dataset, diagram, strategy analysis, RCM study or fleet — optionally password-protected (a passphrase shown once, for you to send separately) and expiring.
  • create_report — Several of your items on one print-ready page, with a title, an author line, an introduction and a note per item.
  • revoke_link — End a link at once.
fleetFleet alerts (Pro).
  • create_alert — Email alerts on a fleet’s expected failures — checked each time usage arrives through the ingest API.
upgrade_linkA Stripe checkout link for Reliafy Pro, for you to open and pay yourself — nothing is charged until you complete it. Works before you have Pro.
  • Censoring follows Reliafy’s convention: 0 = failed, 1 = still running. Data coded the other way round can be fitted with the inverted-flags option.
  • Reading and calculating tools (guide, get, list, the four analysis tools and upgrade_link) are marked read-only; the tools that save, change, share or delete in your workspace are marked as writes, so Claude asks before running them, and deletes only what you ask it to, after confirming by name.
  • Errors: 401 (not signed in, or a missing, expired or revoked token), 403 (an API token whose owner isn’t on Pro — tokens are Pro-only; a signed-in Claude connection without Pro connects, and each tool says it needs Pro), 429 (over 120 requests/minute), 503 (the MCP connector isn’t switched on — it’s available on request).
PreviousFiles and photosNextOriginal routes