Add
Authto your appin minutes

Describe it once and your coding agent wires up sign-in with Descope's MCP server and Skills.

Add MCP to your agentSee the results
56s
average time to working auth with Descope's agent tools For each eval, framework and model, we take the highest-scoring run with Descope's Docs MCP, Skills, or both, and average its active time across 63 tests that also have an agent-alone run.
31%
faster than the same agent alone Total active time of those best-scoring runs compared with the same agent on the same tasks without Descope's tools (1m 21s average). Runs are picked by score, not speed.
97%
of evals got an A, up from 68% without Descope's agent tools Share of those same 63 best-scoring runs graded A (an overall score of 90 or higher), compared with the same agent on the same tasks without Descope's tools.
Recent runs · Click to open the full breakdown
MigrationsNext.jsQwen3.8 Max 0902A · 94.6MigrationsNext.jsQwen3.8 Max 0902A · 96.2MigrationsNext.jsQwen3.8 Max 0902A · 99.6MigrationsNext.jsQwen3.8 Max 0902A · 97.9QuickstartsAngularClaude Opus 5A · 92.7QuickstartsAngularClaude Opus 5B · 88.5QuickstartsAngularClaude Opus 5A · 92.7QuickstartsAngularClaude Opus 5A · 94.3QuickstartsVanilla JSClaude Sonnet 5A · 95.6QuickstartsVanilla JSClaude Sonnet 5A · 92.9QuickstartsVanilla JSClaude Sonnet 5A · 98.3MigrationsNext.jsGemini 3.8 FlashA · 96.2MigrationsNext.jsQwen3.8 Max 0902A · 94.6MigrationsNext.jsQwen3.8 Max 0902A · 96.2MigrationsNext.jsQwen3.8 Max 0902A · 99.6MigrationsNext.jsQwen3.8 Max 0902A · 97.9QuickstartsAngularClaude Opus 5A · 92.7QuickstartsAngularClaude Opus 5B · 88.5QuickstartsAngularClaude Opus 5A · 92.7QuickstartsAngularClaude Opus 5A · 94.3QuickstartsVanilla JSClaude Sonnet 5A · 95.6QuickstartsVanilla JSClaude Sonnet 5A · 92.9QuickstartsVanilla JSClaude Sonnet 5A · 98.3MigrationsNext.jsGemini 3.8 FlashA · 96.2

LEADERBOARD

LEADERBOARD

Every run, every model, every framework

Compare agent-only runs against the same task with Descope MCP and Skills attached. Open a row to see the full breakdown, inline.

1
Qwen3.8 Max 0902
Next.js
Agent alone
90.6%
Agent + Descope Tools
98.7%8.1%Skills scored 98.7%, 8.1 points higher than Agent alone (90.6%).Skills
Time
16m 25s
Cost
$5.83
A98.7%success
Cost$5.83
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.16m 25s
Active timeHow long the agent spent actively working, not counting idle waits.48s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…
2
Qwen3.8 Flash
Next.js
Agent alone
96.1%
Agent + Descope Tools
97.7%1.6%MCP & Skills scored 97.7%, 1.6 points higher than Agent alone (96.1%).MCP & Skills
Time
6m 37s
Cost
$0.29
A97.7%success
Cost$0.29
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.6m 37s
Active timeHow long the agent spent actively working, not counting idle waits.33s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…
3
Gemini 3.1 Pro Preview
Next.js
Agent alone
84.7%
Agent + Descope Tools
97.1%12.4%MCP & Skills scored 97.1%, 12.4 points higher than Agent alone (84.7%).MCP & Skills
Time
2m 8s
Cost
$0.51
A97.1%success
Cost$0.51
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.2m 8s
Active timeHow long the agent spent actively working, not counting idle waits.1m 22s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…
4
Claude Sonnet 5
Next.js
Agent alone
95.6%
Agent + Descope Tools
96.9%1.3%MCP scored 96.9%, 1.3 points higher than Agent alone (95.6%).MCP
Time
3m 10s
Cost
$0.49
A96.9%success
Cost$0.49
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.3m 10s
Active timeHow long the agent spent actively working, not counting idle waits.1m 26s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded10/16/2026
Loading grader breakdown…
5
GPT 5.3 Codex
Next.js
Agent alone
87.8%
Agent + Descope Tools
96.8%9.0%MCP & Skills scored 96.8%, 9.0 points higher than Agent alone (87.8%).MCP & Skills
Time
1m 47s
Cost
$0.59
A96.8%success
Cost$0.59
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.1m 47s
Active timeHow long the agent spent actively working, not counting idle waits.21s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded10/16/2026
Loading grader breakdown…
6
Gemini 3.8 Flash
Next.js
Agent alone
95.7%
Agent + Descope Tools
96.0%0.3%Skills scored 96.0%, 0.3 points higher than Agent alone (95.7%).Skills
Time
2m 2s
Cost
$0.38
A96.0%success
Cost$0.38
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.2m 2s
Active timeHow long the agent spent actively working, not counting idle waits.34s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…
7
Gemini 3.7 Flash
Next.js
Agent alone
95.5%
Agent + Descope Tools
93.7%1.8%MCP scored 93.7%, 1.8 points lower than Agent alone (95.5%).MCP
Time
2m 42s
Cost
$0.29
A95.5%success
Cost$0.29
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.2m 42s
Active timeHow long the agent spent actively working, not counting idle waits.54s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…
8
Claude Opus 5
Next.js
Agent alone
88.7%
Agent + Descope Tools
92.1%3.4%MCP scored 92.1%, 3.4 points higher than Agent alone (88.7%).MCP
Time
7m 5s
Cost
$2.42
A92.1%success
Cost$2.42
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.7m 5s
Active timeHow long the agent spent actively working, not counting idle waits.2m 15s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…
9
Claude Haiku 4.5
Next.js
Agent alone
68.2%
Agent + Descope Tools
90.0%21.8%Skills scored 90.0%, 21.8 points higher than Agent alone (68.2%).Skills
Time
2m 11s
Cost
$0.23
A90.0%success
Cost$0.23
Wall timeTotal elapsed time from the start of the run to the finish. It includes time spent waiting on model responses, tools and builds.2m 11s
Active timeHow long the agent spent actively working, not counting idle waits.30s
Run
EvalNext.js Add Auth
CategoryQuickstarts
Run recorded11/4/2026
Loading grader breakdown…

PERFORMANCE

PERFORMANCE

What MCP and Skills actually change

Same model, same framework, only the tools change. Watch each model's score move as it picks up Descope's MCP server, then Skills, then both.

Hover or select a model to see how each setup changes its score

METHODOLOGY

METHODOLOGY

How a run earns its score

Every run is graded the same way, no matter the model or framework. Some checks just read the code. Others actually run it, click through a real sign-in, and ask a model to judge what's left.

Weighted dimensions

  • Correctness25%

    Does the code actually work? Hallucination and security are graded separately, so nothing gets counted twice.

  • Hallucination15%

    Did the agent only use real packages, APIs, and SDK methods, nothing invented?

  • Setup Friction12%

    How smoothly setup went. Interruptions and provider or SDK errors drag the score down.

  • Setup Speed12%

    How quickly the agent finished, measured against a 60-second baseline.

  • Efficiency12%

    How much of the agent's work was actually useful, versus duplicate reads, repeated failed attempts, or overwritten files.

  • Security10%

    Were secrets handled safely? No hardcoded credentials, no insecure storage.

  • Error Recovery7%

    How cleanly the agent recovered from errors. Each one costs 20 points.

  • Docs Quality7%

    How well the agent used documentation: trustworthy sources, clean fetches, no needless rewrites of code that already worked.

Grader types

  • Code checkscontainsmatches

    Checks the files the agent wrote for a required string or pattern, like the real SDK import or the right flow ID.

  • Forbidden patternsnot_containsnot_contains_in_source

    Fails the run if something forbidden shows up, like a hallucinated package or a hardcoded project ID.

  • Buildcompile

    Actually builds or runs the project. Code that merely looks right doesn't get credit.

  • Live API requestapi_check

    Sends a real request to the running app and checks the response, like an unauthenticated call actually getting rejected.

  • Real browser sign-inbrowser_checkbrowser_flow_check

    Loads the app in a real browser and signs in with a real test user to confirm a session actually gets created.

  • Model judgejudgebrowser_judge

    A model-graded call for the questions a regex can't answer, used on code or on a screenshot.

FOR YOUR AGENTS

FOR YOUR AGENTS

The tools your agents can use

Descope ships an MCP server and a Skills repo so your coding agent gets auth right too.

MCP ServerDocs and project access for agents

Gives agents direct access to Descope's docs, project config, and session management.

View MCP docs
SkillsImplementation guardrails for agents

Teaches agents the exact Descope integration patterns and guardrails to follow.

View Skills docs