The Rillsoft MCP server

What do I need?

#

• Rillsoft Project version 10 on Windows; earlier versions have no MCP server.
• An AI client or a script that speaks the Model Context Protocol over Streamable HTTP in protocol revision 2025-06-18. Older revisions are not accepted by the server.
• For the first steps a shipped industry example; all data in it is synthetic.

Set up in three steps

#

Start the server, point the client at the endpoint, check one read-only call – the first contact needs nothing more. Every step can be reverted on its own, and no step changes a project.

1Enable the server

#

The start flag overrides the settings for the session and suppresses the start screen and the auto-recovery prompt so that an unattended start does not hang on a dialog. The status line on the settings page shows whether the server is running, running via flag, blocked (port in use, probably a second instance) or inactive.

2Connect a client

#

The Claude guide covers Code and Desktop separately, including the connection check.

The endpoint is http://127.0.0.1:<port>/mcp – the path is exact, even /mcp/ is a 404. Configuration of an MCP client:

{
  "mcpServers": {
    "rillsoft-project": {
      "type": "http",
      "url": "http://127.0.0.1:3928/mcp"
    }
  }
}

With an API key set, one header is added:

{
  "mcpServers": {
    "rillsoft-project": {
      "type": "http",
      "url": "http://127.0.0.1:3928/mcp",
      "headers": { "Authorization": "Bearer <KEY>" }
    }
  }
}

Both versions are available as files in the public repository: client configurations on GitHub.

3First safe test

#

No call works without initialisation. Three messages are enough, and none of them changes the project:

1. initialize – the server assigns a session; the id is in the response header Mcp-Session-Id.
2. notifications/initialized – the confirmation unlocks the tools. From now on every request carries Mcp-Session-Id and MCP-Protocol-Version.
3. tools/list – the contract: every tool with title, description, input and output schema. A first read-only call is rillsoft_project_tree_get, the complete computed project tree.

Every tool answer delivers the result twice: as plain text in content[0].text and machine-readable in structuredContent. At the end a DELETE /mcp with the session id ends the session.

# 1. initialize – the session id is in the response header
MCP_SESSION=$(curl -s -i -X POST http://127.0.0.1:3928/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":0,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1.0"}}}' \
  | tr -d '\r' | awk -F': ' 'tolower($1)=="mcp-session-id"{print $2}')

# 2. from now on every call carries session and protocol revision
mcp() { curl -s -X POST http://127.0.0.1:3928/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Mcp-Session-Id: $MCP_SESSION" \
  -H "MCP-Protocol-Version: 2025-06-18" -d "$1"; }

# 3. the confirmation unlocks the tools
mcp '{"jsonrpc":"2.0","method":"notifications/initialized"}'

# the contract: every tool with title, description, input and output schema
mcp '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# a first read-only call: the complete computed project tree
mcp '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"rillsoft_project_tree_get","arguments":{}}}'

MCP server security: how is access protected?

#

The API key is optional. Empty means: no authentication. If a key is set in the settings, every request – regardless of how the server was started – requires Authorization: Bearer <KEY>, otherwise the server answers 401. The comparison runs in constant time. The key is needed on multi-user and terminal-server machines, because 127.0.0.1 is machine-wide, not session-wide.

What do read-only mode and licence cover?

#

In read-only mode tools/list shows exactly the tools of the access class read-only: no task, no assignment, no date, no file, no undo step. Mutating calls – including file, pool and undo tools – return a clear tool error without touching the project.

Some read-only tools visibly change something: the window tools only the window, the display tools the view state that is saved with the project file (columns, outline, selection). The project counts as modified afterwards. The description of each of these tools states this effect.

Most tools check no licence; they are as open as the document model. Four follow the licence like their counterpart in the user interface: the candidate list, the assistants “assign staff” and “remove staff”, and the role derivation. If the function is missing from the licence, they return an error instead of working silently.

What you see is what the AI gets

#

The capacity report computes nothing itself. It serialises the visible, already computed view: period from the time filter, grid from the calendar scale, metric from the selected display. Every perspective the AI sets is on screen at the same time – and one Ctrl+Z restores view, structure, time filter and display together. The same applies to every mutation: if Rillsoft Project detects a conflict on completion, it rolls back itself and the tool reports the reason with the project unchanged.

Without AI: scripts and integrations

#

Not a chat plug-in but access to Rillsoft Project as a whole – with or without AI, by script, with sample data included. The same endpoint works without an AI assistant. The mcp function from the curl example above already is a script; in PowerShell a call looks like this:

$body = @'
{"jsonrpc":"2.0","id":1,"method":"tools/list"}
'@
curl.exe -s -X POST http://127.0.0.1:3928/mcp `
  -H "Content-Type: application/json" `
  -H "Mcp-Session-Id: $env:MCP_SESSION" `
  -H "MCP-Protocol-Version: 2025-06-18" -d $body

Typical uses: mirror absences from the leave system into the resource pool on a rolling basis, pull capacity and plan/actual reports on a schedule – with a scheduled agent also the status report –, generate projects from templates, set a view reproducibly for a screenshot. Every tool describes input and output as a complete JSON schema; both are enforced at runtime. Field names are English and identical in both directions, so an answer can be passed on to the next call without renaming.

Two pitfalls when starting automatically: start Rillsoft Project from PowerShell or cmd, because a POSIX shell such as Git Bash rewrites the /mcp flag into a path. And the server listens only a few seconds after the program has started.

How fast does the server respond?

#

Measured on the largest shipped sample project of the plant engineering package with 650 elements, 637 dependencies and about 1,500 assignments, in the debug build on the reference machine:

Working with a finished project is uncritical. Only the scripted construction of large projects is slow, because every mutating call creates a complete undo snapshot; the whole plant engineering package with twelve projects takes about 16,700 calls.

What about several instances?

#

Rillsoft Project is a single-document application: one instance, one project, one port. The server binds exclusively to its port; a second instance on the same port fails at start and reports this in the status line. For parallel projects every instance gets its own port (/mcp:3928, /mcp:3929, …) and its own server entry in the client. There is deliberately no automatic port search – the client must know which instance it addresses. There is no project directory and no addressing by project id; a project is opened by its file path.

How does the server report errors?

#

The errors, undo and safe writes reference explains responses and the checks required before retrying.

Protocol errors and business rejections are separate: the caller cannot correct a protocol error, but builds the second attempt from a business rejection.

The error text is an agent’s self-correction channel. It names the field path, the permitted values or the tool that establishes the missing state. Next to it stands a stable English error code that is the same in the German, English and Russian language versions – a retry rule is built on the code, not on the wording. At most 32 sessions run concurrently; JSON-RPC batches are rejected.

If no port opens after a hard program end: the auto-recovery dialog waits for an answer when the server was started via the setting rather than the flag. Answer the recovery files *.mbk under %TEMP%\Rillsoft or remove them beforehand for scripted runs.

Reference: the ten tool areas

#

The six cards above summarise; in full, the tools cover ten areas – the grouping of the product documentation. The contract per tool – schemas, access classes, error codes – is described on the developers page; what an area looks like in practice is shown by the linked examples.

Next step

#

The prompt library shows 39 questions with prerequisites, tool chain and limits. The contract with schemas and error codes is on the developers page. If you want to feel it first, begin with getting started.

Product version

Rillsoft Project 10

Verification

Details from the Rillsoft Project 10 product documentation (as of 2026-09-08); reconciliation with the released build is pending.