Horizun Project MCP

An MCP server for Microsoft Project that needs neither Java nor Microsoft Project — and tells you the truth about what it wrote.

Español

Point any MCP client at a .mpp, a Primavera .xer, or an MSPDI .xml and ask real questions: where the critical path runs, which resources are overbooked, whether the schedule would survive a DCMA audit, what a two-week slip actually does to the finish date. Then write changes back — and know which ones landed, because every one is re-read from the model before it is reported.

dotnet tool install -g HorizunMsProjectMcp

That is the whole install. No JVM. No Microsoft Project. No licence. It reads and schedules on its own.

Connect a client

Full per-client instructions, including the plugin route and uninstall, are in docs/INSTALL.md.

Why this one

There are a handful of Microsoft Project MCP servers. Every one of them requires Java, or a licensed Microsoft Project install, or a paid JDBC driver — and on the machine this was built on, none of them would start. Eight things here are not available anywhere else:

The 25 tools

Session — project_health · project_open · project_save

project_health is a doctor, not a ping. It detects Microsoft Project, tests whether its COM server actually starts, and when it does not, hands back the HRESULT diagnosis and the repair steps. That is not hypothetical: this machine hit CO_E_SERVER_EXEC_FAILURE with a perfectly valid registration, and step one of the repair path it emits is what fixed it.

Reading — project_info · tasks_query · links_query · resources_query · timephased_query

Filtered, paged, field-selectable. Tasks are addressed by their stable uid; the row id is display only, because it shifts the moment a task is inserted and addressing by it edits the wrong task.

Analysis — schedule_analyze · schedule_qa · baseline_compare

Computed server-side, so the agent asks a question instead of pulling two thousand tasks into context. Critical path, float distribution, driving path, day-by-day overallocation, DCMA-14, and full earned value (BCWS/BCWP/ACWP, SPI, CPI, EAC, TCPI) — denominated in cost where the schedule carries costs, in work hours where it carries hours, and weighted by duration where it carries neither, which is most of them. The report says which.

Writing — tasks_write · links_write · resources_write · calendars_write · schedule_update

Batched, typed, verified. Cycles are refused before they are applied, with the offending chain named. Two things this backend cannot do are not offered: reordering a task within the outline, and editing a calendar's weekly working-hours pattern. Asking for either gets a refusal that names it and says where to do it instead — an operation that half-works is worse than one that is absent.

Planning — schedule_recovery · schedule_target · schedule_sequence · schedule_learn · schedule_generate

Reprogramming, measured rather than asserted. schedule_recovery finds what is late, ranks it by how much of the schedule sits behind it, then tries each recovery lever — removing lag on the driving chain, overlapping hand-offs, compressing the longest critical tasks — on a throwaway copy and reports the finish date each one genuinely produces. schedule_target tests a date you have been handed and names the work the network does not hold in place. schedule_sequence recovers the logic a schedule is missing by reading the order its own dates already state — the planner laid the work out correctly and never linked it, and that decision is recoverable. schedule_learn mines finished schedules for how long each activity actually takes and what usually comes before it; schedule_generate turns that into a first draft, one task per apartment or floor, sequenced the way the history says the trades follow each other.

Feed schedule_learn a model export alongside the schedules and it measures productivity — what a crew actually got through in a day — by joining quantities to tasks on the shared code. schedule_generate then sizes durations from the quantities of the new project rather than copying a remembered duration, because the rate is what carries between projects and the quantity is what changes. Give one export per schedule, in the same order: rates are measured per project, and quantities totalled across projects would inflate every one of them.

A draft can only be as well sequenced as the schedules it learned from. Where the sources link each activity to itself unit after unit but never to the trades around it, both tools say so and name the number: the library reports how many activities learned a predecessor other than themselves, and the draft reports how many trades it left with nothing scheduled before them.

Interop — project_export · project_import · bim_link · bim_sync

CSV, JSON, MSPDI, Primavera XER and PMXML, native .mpp, and a shaped Power BI dataset. Imports plan before they write.

What it costs to have loaded

The 25 tools present about 8,400 tokens of schema, in every prompt, for as long as the server is connected. That is the honest price of the surface and it is worth knowing before choosing to carry it. It is also why the surface stayed at 25: the largest alternative ships 79 tools, and past a point an agent cannot hold the surface in its head well enough to choose correctly within it.

The two contracts

Nothing is applied until it is verified.

{
  "applied": 12,              // operations that fully succeeded, re-read from the model
  "fieldsVerified": 31,
  "rejected": [{
    "uid": 45, "field": "finish",
    "requested": "2026-09-10", "actual": "2026-09-14",
    "reason": "'finish' is calculated from the task's duration and its predecessors. Change the
               duration or the logic, or set a constraint, rather than writing the date."
  }],
  "impact": {
    "tasksMoved": 312,
    "projectFinishBefore": "2027-03-14", "projectFinishAfter": "2027-03-28",
    "criticalPathChanged": true, "newNegativeFloat": 8
  },
  "verifiedBy": "reread"
}

A dry run is a real simulation. dryRun: true deep-copies the schedule, applies the batch, reschedules it with the critical-path engine, measures the difference, and throws the copy away. The impact numbers are observed, not predicted.

Capability honesty

project_health publishes a capability matrix, and the tools honour it. Two things stay false on the file backend and refuse rather than approximate: