Evaluate

A benchmark score doesn’t tell you whether your server holds up in production. Evaluate lets you pin down the behaviors you care about — which tools fire, with what arguments, what the final answer looks like — and run them across the models your users actually use.

​

Requirements

Evals are available to both signed-in users and guests. Guests authenticate automatically using their guest session, so no sign-in is required to run or generate tests.

​

How it’s organized

​

Authoring a case

Click New case in the suite header (or in the empty-state screen when no cases exist yet) to open a blank case editor. The case is a draft until you press Save — backing out without saving leaves your suite unchanged. Each case has four moving parts:

• Scenario — short label so you can find it later (“Draw a rectangle”, “Refuses unsafe delete”).
• Prompt turns — one or more user messages. Multi-turn is supported; expected tools and checks attach per turn, so you can model “ask, get a result, follow up.”
• Expected tools — for each turn, the tool calls the model should make. Arguments can be exact values or typed placeholders ("string", "number") so you’re not chasing flaky literals.
• Expected output — optional free-text description of what a good final answer looks like. Used by the judge.

​

Recording widget interactions

If your MCP server uses MCP Apps (widgets rendered in a sandboxed iframe), you can record real in-widget interactions — clicks and typing — directly into a test case as interact steps. When you arm record mode on a test case, a recorder shim is injected into the widget sandbox. Every click and text input you make in the widget is captured as a locator-stable interact step and appended to the case. The locator is built from stable identifiers in priority order: data-testid → ARIA role + accessible name → visible text → CSS path. This means the recorded step resolves to the same element when the case runs headlessly.

​

Widget ui/message follow-ups

Some widget interactions send a ui/message back to the host (for example, a cart button that says “Show my cart”) rather than calling a tool directly. During eval execution, these follow-up messages automatically drive a continuation model turn — the same way Playground handles them live. The resulting tool call (e.g. view-cart) is attributed to the same interact step’s turn, so a Tool was called check on that turn sees the call. This means an eval case can faithfully verify a full widget interaction sequence: prompt → widget renders → user clicks → model reacts → tool fires → assertion passes.

​

What the trace shows

The Chat/Trace tab renders widget tool calls as the same app-attributed cards you see in Playground, so you can inspect exactly what the widget returned and what the model did with it. Checks are what actually decide pass/fail. They’re pure functions of the iteration transcript, so the verdict is the same every time you replay it — which is the property you want if you’re using Evaluate as a regression gate. Set defaults on the suite; override per case with inherit (use suite defaults), replace (use only the case’s list), or extend (suite defaults, then the case’s list). When a check is authored on an individual prompt turn it is evaluated against that turn’s slice of the transcript — tool calls, assistant message, tool errors, and widget observations for that turn only. Token budget under N is case-only because per-turn token usage is not reliably captured; all other checks are turn-scopable.

​

Harness system tools in assertion dropdowns

When a suite is configured to run under a harness host (such as Claude Code or Codex), the tool-name pickers in the case editor also offer the harness’s native system tools — for example bash, read, webSearch, WebFetch — alongside your MCP server tools. This lets you write assertions like Tool was never called for bash or Tool was called with… for read without having to type the wire name by hand. System tools appear only in assertion pickers. They are not available in pinned tool-call steps or widget-assertion “View (tool)” selects, because those require a tool that MCPJam can invoke directly. If an MCP tool and a system tool share the same name, the MCP tool takes precedence and the system tool is not added to the list.

​

Argument matching

Tool-call argument comparison runs in one of three modes, configured at the suite:

• partial (default) — every expected key must be present and match; extra keys in the actual call are ignored. Best for “I care about query and limit, not what else the model put in.”
• exact — actual args must equal expected args, key-for-key.
• ignore — only the tool name is checked.

Type placeholders ("string", "number", "boolean") match any value of that type, which lets you assert shape without locking in a literal.

​

Validator settings

Each suite has default validator settings controlling how tool calls are matched. Override them at three levels: A run override shows an override badge; click Reset in the popover to clear it.

​

LLM as judge

For cases where “did the right tools fire” isn’t enough — anything graded on the quality of the final answer — the judge grades the run against your expected output if you set one, and against the user prompt otherwise. It’s advisory: it produces a score and a rationale, but doesn’t gate the run unless you ask it to.

• On by default at the suite level. The cost is gated by an explicit Run judge click on the run-detail page — it won’t run for every iteration unless you turn auto-run on.
• Calibrate per suite. Judge scores aren’t comparable across domains; a 0.7 on one suite isn’t a 0.7 on another.
• When grading against the prompt rather than an expected output, scores are capped at 0.85 — you can’t get a “perfect” without saying what perfect means.

Open a completed run and find the LLM as Judge panel (labelled Goal judge when shown inside the suite results split). Pick a judge model (default: openai/gpt-5.4-mini) and a threshold (default 0.7), then click Run judge. Each case gets a score, an advisory verdict, a one-line reason, and rubric hits. The judge never runs automatically unless you enable Auto-run in suite settings.

​

Running

A run needs three things, all picked from the suite header:

1. Servers — one or more attached to the suite. Cases can attach their own subsets if they only need part of the surface.
2. Models — the multi-model picker is the whole point. Each model produces its own iteration per case, so you can see where Claude passes and ChatGPT trips.
3. Run all — kicks off every case × every model. Run one from a case row runs just that case.

​

Frozen execution snapshots

The first run of a suite saves the set of MCP servers used as a frozen snapshot; reruns reuse it, so connecting new servers can’t silently change what a suite runs against. Click Update snapshot in the suite header to re-save the current servers and start a new run.

​

Computer environment

In suite settings, the Computer environment select lets you pin a built Docker environment so every eval iteration boots a fresh sandbox from the same image. This keeps results comparable across runs and edits — each iteration gets an identical starting state.

• Choose an environment from the list. Environments that have not finished building are labelled (not built); selecting one will cause the run to fail immediately.
• Choose None (default image) to run without a pinned environment.
• The agent receives a bash tool inside the sandbox.

Once a run completes, the Run detail page shows an Environment row with the name of the environment used. If the environment has since been deleted, the row shows its raw identifier instead so the record remains durable. Hovering the label shows the image template, digest, and provider from the frozen snapshot taken at run time.

​

Suite execution config

Each eval suite has a Default Execution Config section that controls the model, system prompt, temperature, tool approval, connection settings, capabilities, and host context used when running the suite.

• Model and prompt — Set the default model ID, system prompt, and temperature for all runs in the suite. When running evals against MCP Apps that render widgets, the eval harness automatically enables browser interaction (screenshot, click, type) for any model that supports both vision and tool calling — not just Claude. Models that lack vision or tool calling run without browser interaction.
• Tool approval — Toggle requireToolApproval to pause before each tool call during a run.
• Server selection — Servers are not configured here. They come from the suite’s environment. The server picker is intentionally hidden in this editor.
• Save / Reset — Click Save config to persist changes. Click Reset to revert to the last saved state. Unsaved edits are preserved if the page refreshes the config from the server, but are discarded when you switch to a different suite.

Changes to the suite execution config apply to future runs only. Existing run snapshots are not affected.

​

How suite defaults apply to test cases

The suite-level system prompt and temperature are runtime defaults: when a test case does not set its own system prompt or temperature, the suite values are used for that iteration. A per-case override always wins — the suite default only fills the gap.

​

CI tab

The CI tab (when enabled for your project) shows only CI-active suites — suites created by the SDK or suites that CI has reported into at least once. Playground-only suites stay in the Evaluate tab. If a project has no CI evals yet, the CI tab shows a quickstart prompt. The commit rail in the CI tab groups runs by commit SHA and only includes CI (SDK-sourced) runs. Playground runs on mixed suites are excluded from the rail so they don’t appear as spurious manual-commit groups.

​

Pass Rate vs Accuracy

The metric label shown in run headers, the runs table, charts, and hero stats follows the run’s source, not the suite’s creation source:

• Pass Rate — shown for runs submitted by the SDK (CI runs). Reflects per-case pass/fail.
• Accuracy — shown for runs triggered manually from the UI, via the API, or on a schedule.

On a mixed suite (one that has received both CI and playground runs), the label updates to match the most recent run. Legacy runs without a recorded source fall back to the suite’s creation source.

​

CI-synced test cases

Test cases written by a CI report carry a CI chip next to their name in the cases list. Opening such a case in the editor shows a notice: “Synced from CI — the next CI report may overwrite manual edits.” Manual edits to these cases are preserved until the next CI ingest, which may overwrite them.

​

Suite deletion

The delete button in the suite switcher is hidden for any suite that is CI-active — either created by the SDK or one that CI has reported into. This protects CI run history; the next CI report would recreate the suite anyway.

​

Reading results

​

Suite navigation

Opening the Evaluate tab takes you directly into the most recent suite’s dashboard. To switch suites, use the suite switcher in the breadcrumb — it lists all your suites and lets you jump to one, create a new suite, or delete the current suite. If no suites exist yet, the empty-state create prompt appears instead.

​

Suite view

• Suite accuracy — pass rate of the most recent run, with the last three runs’ trend so you can see whether you’re improving or regressing.
• Run insights — an AI-written diff against your previous completed run: which cases moved, which tools changed behavior, which models diverged. Skim this first; it usually points at the right rabbit hole.
• Runs tab — every run with its summary metrics. Click in for the iteration-level breakdown.
• Cases tab — every case with its latest verdict and a quick replay button.
• Executions tab — a flat, filterable list of every individual test execution across all cases, sorted most-recent-first. Each row shows the case name, result (passed/failed/pending/cancelled), and timestamp. Click any row to open it in the compare view.

When you run a suite against multiple host configurations at once, the Runs list groups them into a single collapsible run group row showing mean accuracy and longest duration across all hosts. Expand it to inspect each host’s individual run. The Runs / Cases segmented control switches between the run history list and the test-cases overview. Performance by Model in the run summary appears only when more than one model was used. To compare two runs, check their boxes in the Runs list and click Compare.

​

Cross-host matrix

When two or more host configurations are attached to a suite, a By case / By host toggle appears. By host shows a matrix — one column per host, one row per case. Each cell shows pass/fail dots, pass rate, median latency, and token usage. A host detached after runs were recorded stays visible, labelled historical.

​

Run view

• Per-iteration row with case, model, pass/fail, tokens, duration, and the tool calls that actually happened.
• Expected vs actual tool calls side-by-side when an iteration fails. For failed iterations, a Categorized diff above the raw Expected/Actual grids groups discrepancies into four categories:
  • Missing — expected tool calls that were never made
  • Extra — actual calls that weren’t expected (reported but non-fatal by default)
  • Out of order — calls that happened in the wrong sequence (when order checking is enabled)
  • Arg mismatch — right tool name, wrong arguments (shown side-by-side)
• Full trace — every turn, every tool call, every token. This is the thing you couldn’t see before; spend time here. When a turn ran through a harness (e.g. Claude Code), the trace detail pane shows an Engine badge — “Claude Code” for harness turns, “Emulated” for turns that ran through MCPJam’s built-in chat engine — so you can tell at a glance which runtime handled each turn.
• Per-model breakdown to compare how the same case behaves across models.
• Suggested fixes — after a run completes, a Suggested fixes panel ranks tool-quality and workflow issues by impact. Each issue has a Copy button that copies a ready-to-use fix prompt (tool description + input schema) for a coding agent; Copy top N copies the top issues combined.
• Predicate Gate — when successPredicates are configured on a case, an expandable Predicate Gate section in the iteration detail lists each predicate with a PASS/FAIL verdict, a one-line summary, and the evaluator’s reason. Hidden when no predicates are configured.

​

Comparing two runs

Select any two completed runs and click Compare. The diff view shows per-case status changes: Passed, Still failing, Regressed (pass→fail), Fixed (fail→pass), New, Removed, and Changed (config differed). Summary metrics show deltas for tokens, cost, and duration.

​

Case view

• Pass rate across runs — is this case stable, flaky, or trending down?
• Pass rate by model — does this case only work on one model?
• Every past iteration with its trace, so you can A/B a regression against a working run.

​

Promoting sessions to cases

Any completed chat session can be turned into a test case without rewriting it from scratch. The full conversation is compiled into multi-turn prompt turns, and the dialog lets you pick a destination suite or create a new one. From the Playground — Open the Sessions rail, hover a session, and click Promote to test case. From the Swarms tab — Open a completed run, click a session row to select it, then click Promote to test case above the session viewer. The session must belong to a completed run attempt; sessions from in-progress runs show an error in the dialog instead of blocking at submit. In both cases the dialog shows the servers recorded for that session, lets you set a case title and suite name, and (when attachment pickers are enabled) lets you configure the server attachment and client host for the new suite.

​

Generating cases from your tools

The Generate button reads your attached servers’ tool catalog and drafts realistic cases — a mix of positive (“call search with a query”) and negative (“don’t call delete_user on a meta-question”). Treat it as a draft: skim, edit the prompts to match how your users actually talk, tighten the checks, then save. When the suite has a saved server group spanning two or more servers, the generator produces coverage for each server individually plus at least one cross-server case. Suites without a saved server group treat all available servers as a single pool.

​

Exporting traces

Open a completed run and click the Export traces button (download icon in the run header) to download the run’s traces as an OTLP JSON file. The export uses OpenInference semantic conventions plus mcp.app.* extensions, making it compatible with Arize Phoenix, Datadog, and any other OTLP-compatible observability backend.

​

Export options

The downloaded file is named mcpjam-traces-<scope>-<timestamp>.json.

​

What to author first

If you’re new to the surface, the shortest useful loop is:

1. Attach the server you’re shipping.
2. Generate a starting set of cases.
3. Delete the ones that don’t match real usage; tighten checks on the rest.
4. Add the two or three models your users will hit.
5. Run all, open the failures, and decide whether the bug is in your server, your prompt, or the model.

That last step is the one that benefits most from spending time in the trace view.