Skip to content

MCP servers

An MCP server in a project is a server whose tools the project tests. Plune keeps its contract — what the server itself answers about its tools — as numbered snapshots, says what changed between two of them and whether a client that worked may now fail, and counts which tools your cases test.

The catalog keeps which server it is and how it is reached (stdio or HTTP), never where: there is no field for an address, a command, a header or an environment variable, and a request that sends one is refused with 400.

Catalog in the project has a switch API | MCP. MCP lists the tools of every MCP server of the project, by name: the server and the tool’s title under its name, where it was learnt (Snapshot), how far it is trusted, when a snapshot last listed it and in which environment, and how many cases test it. A tool no case tests says No cases, never 0%. Under the table, MCP servers: links each server.

Catalog with MCP chosen and Add MCP server: the Tools tab, Drift · 1 and Without cases · 2, then six tools by name — cancel_order, get_order, list_orders, ping marked Not listed by snapshot #2, refund_order and search_articles — each with its server, Snapshot, Confirmed, Nov 20, 2026 · staging and its cases or No cases; under them 6 tools · by name and MCP servers: orders-mcp · help-search-mcp. Catalog with MCP chosen and Add MCP server: the Tools tab, Drift · 1 and Without cases · 2, then six tools by name — cancel_order, get_order, list_orders, ping marked Not listed by snapshot #2, refund_order and search_articles — each with its server, Snapshot, Confirmed, Nov 20, 2026 · staging and its cases or No cases; under them 6 tools · by name and MCP servers: orders-mcp · help-search-mcp.
  • Without cases · N shows only the tools the newest snapshot lists that no case tests.
  • Drift · N lists the tools the latest snapshot no longer lists — The server stopped listing it. The tool stays, with its cases and its history.

Add MCP server asks for a Name (up to 100 characters, once per project whatever its case), a Transport — HTTP or stdio — and an optional Description. A read-only member does not see the button. A project holds at most 50 servers, and a server cannot be removed yet.

A server also comes in from a mission: an MCP server source added to a mission puts it in the catalog at once, marked Derived; a server you add is Confirmed.

A snapshot is what plune mcp snapshot reads from a running server — its name and version, its capabilities and instructions, and its tools, prompts, resources and resource templates (Snapshot and diff an MCP server). --push sends it to the project of the token:

Terminal window
export PLUNE_TOKEN=plune_EXAMPLE-not-a-real-token-0000000000000
npx -y @plune-ai/cli mcp snapshot --push --server-name orders-mcp --environment staging -- node orders-server.mjs

The first push adds the server and its first snapshot; a later contract is the next number, and the same contract again records nothing. With @plune-ai/cli 0.39.5:

Added server "orders-mcp" to the platform.
Pushed snapshot 1 of server "orders-mcp": a new contract (efe88b6db1c3).
Pushed snapshot 2 of server "orders-mcp": a new contract (07143539a3f0).
Snapshot 2 of server "orders-mcp" already holds this contract (07143539a3f0); nothing new was recorded.
  • --server-name finds the server by name, whatever its case, and adds it on the first push; --server <id> names one the project already has. One of the two is required.
  • --environment says where the server was running (ci, staging). It is kept with the snapshot and is not part of the contract.
  • The same contract as the latest snapshot adds none. The platform answers with the latest one again, and the CLI says nothing new was recorded. Any other contract is the next number.
  • Without --push nothing is sent: the snapshot goes to stdout, or to --out.

The token must be one that may write — a person’s session, or a write token without a session, as a CI step holds. An agent is refused with 403 agent_proposes. A server holds at most 200 snapshots.

What is refused, with 400: a field the snapshot does not name, and a key that carries a credential or a connection — authorization, headers, env, command, args, cookie, token, password, secret, apikey and any key ending in token, secret, password or apikey — at any depth of capabilities and of a tool’s annotations and execution. The answer names the path, never the value. The text a server wrote (description, instructions, title) and its schemas are kept as sent: a secret a server writes there is stored.

A server’s page names who added it and when, its Transport, its Tools and its Snapshots, and has three tabs: Tools, Contract history and Coverage.

The server orders-mcp, Confirmed, Added by John Doe · Nov 20, 2026; Transport stdio, Tools 4, Snapshots 2; the tabs Tools (5), Contract history (2) and Coverage; the five tools with ping marked Not listed by snapshot #2. The server orders-mcp, Confirmed, Added by John Doe · Nov 20, 2026; Transport stdio, Tools 4, Snapshots 2; the tabs Tools (5), Contract history (2) and Coverage; the five tools with ping marked Not listed by snapshot #2.

Contract history lists the snapshots, newest first and the newest marked Latest: the date, the environment, how many tools, who took it, and the first 16 characters of its sha256. The hash is taken from the contract with its keys sorted, so the same contract has the same hash whatever order it came in. With two snapshots or more, From, To and Compare open two of them side by side.

Plune compares the tools of two snapshots of one server, tool by tool, and gives each change a class by what a client that worked with the older contract meets with the newer one:

ClassWhen
BreakingA client may fail or misread: a tool gone, a parameter gone or newly required, a type changed, a value no longer accepted, an output field gone or optional, a tool that is no longer read-only or may now destroy data.
Non-breakingIt keeps working: a tool added, an optional parameter added, a parameter no longer required, a field added to the result.
InformationalA text or a hint changed (a title, a description), or no rule knew the change. A change no rule knows is never called safe.

A changed description is Informational and marked Read by a model: it is what a model is told about the tool, so it is a reason to look whatever its class.

#1 → #2 of orders-mcp, Breaking, with both hashes and the counts 2 breaking, 1 non-breaking, 1 informational; Changed tools (4): cancel_order added, Non-breaking; list_orders changed, Informational, Read by a model, its description; ping removed, Breaking; refund_order changed, Breaking, inputSchema.amount — The parameter became required.; under Other sections, server. #1 → #2 of orders-mcp, Breaking, with both hashes and the counts 2 breaking, 1 non-breaking, 1 informational; Changed tools (4): cancel_order added, Non-breaking; list_orders changed, Informational, Read by a model, its description; ping removed, Breaking; refund_order changed, Breaking, inputSchema.amount — The parameter became required.; under Other sections, server.

The comparison says one of three things, never one for another: the tools that changed, each with its class and one sentence per field; Unchanged — both snapshots list the same tools; or Not measured — a snapshot lists no tool, so there is nothing to compare, which is not the same as nothing having changed. Sections other than the tools — the server’s name and version, its capabilities, its instructions, prompts, resources and resource templates — are named under Other sections when they differ, and not classed.

A tool’s card shows its description, where it was learnt, when it was last seen, the server and snapshot it is from, its Input schema, Output schema and Annotations as the server sent them, its Changes — each pair of neighbouring snapshots in which it changed, linked to the comparison — and its Cases, each with its newest result.

The card of refund_order, Added by John Doe: its description, Source Snapshot, Last seen Nov 20, 2026 · staging, Server orders-mcp · snapshot #2; the input schema with amount and orderId required, The server sent no output schema., the annotation destructiveHint; Changes (1): #1 → #2, Breaking, inputSchema.amount; Cases (1) with Link cases, and Refund part of an order marked May be out of date and failed. The card of refund_order, Added by John Doe: its description, Source Snapshot, Last seen Nov 20, 2026 · staging, Server orders-mcp · snapshot #2; the input schema with amount and orderId required, The server sent no output schema., the annotation destructiveHint; Changes (1): #1 → #2, Breaking, inputSchema.amount; Cases (1) with Link cases, and Refund part of an order marked May be out of date and failed.

Link cases picks the cases that test the tool, searched by title, up to 200; the set is saved whole. A read-only member does not see it. A tool is linked to a case only this way, never by a name. A tool the server stops listing stays, with its cases and its history.

Coverage counts the tools the newest snapshot lists: how many have cases and how many have none, how many have cases with no result yet, and how many cases may be out of date. It gives counts, never a percentage: a server nothing tests reads Nothing on this server is tested yet: coverage is 0 of N, never 100%. A tool no longer listed is neither shown nor counted here.

The Coverage tab of orders-mcp: 3 of 4 tools have cases · 1 has none · 1 has no result yet · 1 case may be out of date; cancel_order Not run yet, get_order 2 passed, list_orders No cases, and refund_order 1 failed with May be out of date — Breaking change in snapshot #1 → #2 — See the change and Dismiss. The Coverage tab of orders-mcp: 3 of 4 tools have cases · 1 has none · 1 has no result yet · 1 case may be out of date; cancel_order Not run yet, get_order 2 passed, list_orders No cases, and refund_order 1 failed with May be out of date — Breaking change in snapshot #1 → #2 — See the change and Dismiss.

Each tool shows its cases counted by their newest result — No cases and Not run yet stand apart from a pass — and the date of the latest result.

When the newest pair of snapshots changes a tool — a Breaking change, or any change to a tool that is still there — its cases are marked May be out of date, with the change that marked them and See the change. The mark is advice: no gate reads it, and the case itself is not changed. Dismiss takes it off for that snapshot; it comes back when a later snapshot changes the tool again. A read-only member sees the mark without Dismiss.

Results reach the cases as any other run’s do — for example a JUnit report your MCP tests wrote, with plune run import:

Terminal window
npx -y @plune-ai/cli run import ./junit.xml --name "orders-mcp tools"

An eval run of several models, as plune sync sends it, has a Models tab beside Results: each case on each model.

The Models tab of an eval run: Population, 3 cases with its fingerprint; each case on model-small and model-large with its passed share, cost and latency, 1 broken counted apart, and All cases 20/29 · 69% against 28/29 · 97%; under Trend · same population an earlier run of the same day and this run. The Models tab of an eval run: Population, 3 cases with its fingerprint; each case on model-small and model-large with its passed share, cost and latency, 1 broken counted apart, and All cases 20/29 · 69% against 28/29 · 97%; under Trend · same population an earlier run of the same day and this run.
  • A cell gives the attempts that got a verdict and the share that passed, the cost and the median latency. Attempts without a verdict are counted apart, as N broken, and never in the share.
  • A model that never ran a case reads not run; a cell where nothing got a verdict reads no verdict — neither is 0% or 100%.
  • A cell with fewer than 5 judged attempts claims no difference: Too few attempts to compare, and only the counts are shown.
  • Trend · same population compares this run with the earlier synced runs of the project (up to 20 looked at) whose cases, data, checks and repeats are the same. A run that does not say what it measured is compared with no other: Comparison unavailable.
  • A result that names no model is left out of the table and counted under it.

At 375 px the MCP tab and a server’s page are one column.

The MCP tab at 375 px wide: the top bar of Orders, Catalog with Add MCP server, API and MCP, Tools and Drift · 1, Without cases · 2, and the tools as cards — cancel_order, get_order, list_orders, then ping, Not listed by snapshot #2. The MCP tab at 375 px wide: the top bar of Orders, Catalog with Add MCP server, API and MCP, Tools and Drift · 1, Without cases · 2, and the tools as cards — cancel_order, get_order, list_orders, then ping, Not listed by snapshot #2.
The server orders-mcp at 375 px wide: Confirmed under its name, Added by John Doe, Transport stdio, Tools 4, Snapshots 2, the tabs Tools (5), Contract history (2) and Coverage, and the tools as cards. The server orders-mcp at 375 px wide: Confirmed under its name, Added by John Doe, Transport stdio, Tools 4, Snapshots 2, the tabs Tools (5), Contract history (2) and Coverage, and the tools as cards.

Every call reads the token’s project. Anyone in the project may read; writing is for a person or a write token without a session.

Terminal window
export PLUNE_TOKEN=plune_EXAMPLE-not-a-real-token-0000000000000
API=https://beta-api.plune.ai
curl -H "Authorization: Bearer $PLUNE_TOKEN" "$API/v1/mcp-servers"
curl -H "Authorization: Bearer $PLUNE_TOKEN" "$API/v1/mcp-servers/$SERVER_ID/snapshots"
curl -H "Authorization: Bearer $PLUNE_TOKEN" "$API/v1/mcp-servers/$SERVER_ID/snapshots/diff?from=1&to=2"
curl -H "Authorization: Bearer $PLUNE_TOKEN" "$API/v1/mcp-servers/$SERVER_ID/coverage"

The comparison of the two snapshots above, shortened to one of its four tools:

{
"state": "changed",
"changes": [
{
"tool": "refund_order",
"kind": "changed",
"class": "breaking",
"fields": [
{
"field": "inputSchema.amount",
"class": "breaking",
"reason": "The parameter became required."
}
]
}
],
"summary": { "breaking": 2, "non-breaking": 1, "informational": 1, "overall": "breaking" },
"otherSections": ["server"]
}

summary.mayBeOutOfDate in the coverage counts the marked cases over all tools, not the tools: a tool with two marked cases counts 2.

CallWhat it does
GET /v1/mcp-serversThe servers, in the order they were added.
POST /v1/mcp-serversAdds a server: name, transport (stdio or http), optional description. 409 server_exists for a name the project has.
GET /v1/mcp-servers/{id}One server.
GET /v1/mcp-servers/{id}/toolsEvery tool a snapshot has listed, with listed and how many live cases test it.
GET /v1/mcp-servers/{id}/snapshotsThe snapshots by number, without their content.
POST /v1/mcp-servers/{id}/snapshotsTakes a snapshot: 201 a new one, 200 the latest again when the contract is the same.
GET /v1/mcp-servers/{id}/snapshots/{number}One snapshot with the contract it holds.
GET /v1/mcp-servers/{id}/snapshots/diff?from=&to=The classed changes between two snapshots.
GET /v1/mcp-servers/{id}/coverageThe coverage of the listed tools, with the marks and the requirements their cases are linked to.
PUT /v1/mcp-tools/{id}/casesSets the cases that test a tool, whole: {"testCaseIds": [...]}; [] unlinks every case.
PUT /v1/mcp-tools/{id}/cases/{caseId}/dismissalTakes the mark off a case for a snapshot: {"snapshotNumber": 2}; 409 not_marked when the newest snapshot does not mark it.
GET /v1/runs/{id}/modelsThe model × case matrix of a synced run and the trend of each model.

Another project’s server and one that does not exist are the same 404. Every call, with examples, is in the API reference.

ShotWhat it showsWidth
catalogThe MCP tab of the catalog1440
phone-catalogThe MCP tab375
serverA server with its tools1440
phone-serverA server375
compareTwo snapshots compared1440
toolA tool’s card1440
coverageA server’s coverage1440
modelsThe Models tab of an eval run1440

Each is in a dark and a light version; the page shows the one your theme is in.