MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers

MCP-Atlas evaluates how well AI agents use tools to complete real-world tasks, across 36 Model Context Protocol (MCP) servers in a reproducible Docker sandbox, scored with an LLM-as-judge.

• Paper: arxiv.org/abs/2602.00933 (local copy)

• Leaderboard: scale.com/leaderboard/mcp_atlas

• Dataset: huggingface.co/datasets/ScaleAI/MCP-Atlas

Overview

• 36 real MCP servers spanning search, code execution, databases, APIs, and productivity tools — 20 need no setup, 11 require API keys, and 5 require API keys plus data setup (see data_exports/README.md). All are open-source and version-pinned for reproducibility.

• 500 tasks with ground-truth expected tool calls and answers.

• LLM-as-judge scoring reporting pass rate and coverage, plus per-task failure-mode diagnostics.

Server definitions are in mcp_server_template.json; a full list of the 36 servers and 307 tools is here.

Related MCP server: Forage MCP Server

Quick Start

Requires docker, jq, and Python 3.10+.

git clone git@github.com:scaleapi/mcp-atlas.git && cd mcp-atlas

1. Configure

cp env.template .env

Set in .env:

• LLM_API_KEY — key for the model under evaluation (comma-separated keys are rotated per request).

• LLM_BASE_URL — any OpenAI Chat-Completions-compatible endpoint (a LiteLLM proxy, OpenAI, Anthropic-via-LiteLLM, Azure, or a self-hosted vLLM/TGI server).

• EVAL_LLM_API_KEY / EVAL_LLM_BASE_URL / EVAL_LLM_MODEL — optional judge settings for scoring and diagnostics; fall back to LLM_*, with the judge defaulting to gemini/gemini-3.1-pro-preview.

• MCP_SANDBOX_URL — optional, defaults to http://localhost:1984.

The agent harness was rewritten from Python to TypeScript in v2.0.0 — see CHANGELOG.md.

2. Start the MCP servers

Allocate at least 8 GB (10 GB+ recommended) to Docker.

Option A — prebuilt image (recommended):

docker pull ghcr.io/scaleapi/mcp-atlas:1.2.7
docker tag ghcr.io/scaleapi/mcp-atlas:1.2.7 agent-environment:latest
make run-docker

Option B — build from source (only if you're modifying the server set, pinned versions, or baked-in data):

make build && make run-docker

Neither bakes in API keys — both inject them at runtime from .env. Startup takes 1+ minute; wait for Uvicorn running on http://0.0.0.0:1984. The 20 no-key servers are enabled by default; key-gated servers turn on when their keys are present. Verify:

curl -s http://localhost:1984/enabled-servers | jq -c

3. Start the agent harness (new terminal)

make install-harness
make run-harness

Starts the TypeScript harness on port 3001, exposing /v2/mcp_eval/run_agent — the multi-turn agent loop that runs the model against the sandbox until it finishes or hits a limit.

4. Smoke-test one task (new terminal)

Expected answer: "Customer".

curl -X POST http://localhost:3001/v2/mcp_eval/run_agent \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-4o",
    "messages": [{"role": "user", "content": "What is the first word of the file at /data/Barber Shop.csv?"}],
    "enabledTools": ["filesystem_read_text_file"],
    "image": "ghcr.io/scaleapi/mcp-atlas:1.2.7"
  }' | jq

5. Run the full eval

make install-python   # one-time: deps for run_eval, scoring, diagnostics
python run_eval.py --model "openai/gpt-4o" --output outputs.csv

Pulls the 500-task dataset from HuggingFace by default; pass --input tasks.csv for a local CSV (TASK, PROMPT, ENABLED_TOOLS columns). Reruns skip already-completed task_ids, so an interrupted run resumes by rerunning the same command. Output columns: task_id, raw_conversation_history, response.

Keep each run's artifacts together by writing --output into a per-run directory and pointing the scoring/diagnosis steps at the same directory.

Configuration

Override any default per run:

• --extra-llm-params sets reasoning/provider-specific options, e.g. --extra-llm-params '{"reasoning_effort": "high"}' (use whatever key your provider expects; default is the provider's own).

• Harness request timeouts are env-configurable for slow models: TOOL_CALL_TIMEOUT_MS (60000), LIST_TOOLS_TIMEOUT_MS (180000), LLM_TIMEOUT_MS (600000).

• Each run writes a run_config.json beside the output CSV; the scorer embeds it into coverage_stats_*.json so every result is traceable to its configuration.

6. Score

python services/scoring/score_claims.py \
  --groundtruth-file path/to/groundtruth.csv \
  --model-file outputs.csv \
  --model-name your-model \
  --output-dir results/your-model

LLM-as-judge claim-coverage scoring (default judge gemini/gemini-3.1-pro-preview). The ground-truth file is the HuggingFace dataset exported to CSV (columns TASK, PROMPT, GTFA_CLAIMS), or the same --input CSV if you ran locally. Outputs scored_<model>.csv, coverage_stats_<model>_*.json (pass rates at 0.50 and 0.75 coverage thresholds), and a coverage histogram. --concurrency auto-tunes per judge model.

6b. Diagnose failures (optional)

python services/diagnostics/single_model_diagnostic.py --scored-file scored_<model>.csv --verbose

Classifies each failing task into one of 11 failure modes (4 tool-call + 7 cognitive) over an enriched trajectory, and writes a diagnosis_*.csv plus a model-level narrative.

7. Evaluate another model

Change LLM_API_KEY / LLM_BASE_URL in .env, restart the harness, and rerun with a different --model. See LiteLLM providers for model names.

Scaling throughput

A single sandbox handles concurrent tasks comfortably, and you can run several evals in parallel against it. The agent loop is I/O-bound — most of each task's time is spent waiting on the model, not calling tools — so one sandbox stays well under capacity at typical concurrency. Raise --concurrency or launch multiple runs as needed; reach for the scale-out options below only when the sandbox itself becomes the bottleneck (very high concurrency, or tool-heavy workloads where some MCP servers degrade under load):

Shard across independent stacks (simplest). Run several sandbox + harness pairs on different ports, point run_eval.py at a slice of the tasks for each, then concatenate the output CSVs. Each task runs end-to-end on one stack, so within-task state (filesystem, memory, git) stays consistent. The harness's .env does not override variables already set in the environment, so per-stack PORT / MCP_SANDBOX_URL overrides just work:

# Stack A — sandbox on 1984, harness on 3001
docker run -d -p 1984:1984 --env-file .env ghcr.io/scaleapi/mcp-atlas:1.2.7
PORT=3001 MCP_SANDBOX_URL=http://localhost:1984 make run-harness

# Stack B — sandbox on 1985, harness on 3002
docker run -d -p 1985:1984 --env-file .env ghcr.io/scaleapi/mcp-atlas:1.2.7
PORT=3002 MCP_SANDBOX_URL=http://localhost:1985 make run-harness

# Run each half of the dataset against its own harness, then concatenate
HARNESS_URL=http://localhost:3001 python run_eval.py --input tasks_part_a.csv --output out_a.csv --model "<model>"
HARNESS_URL=http://localhost:3002 python run_eval.py --input tasks_part_b.csv --output out_b.csv --model "<model>"

Point at an orchestrator (scales furthest). Because the harness reaches the sandbox solely through MCP_SANDBOX_URL, you can point it at a service that provisions an ephemeral sandbox per task — no harness changes; any HTTP endpoint implementing the agent-environment API works.

One rule when adding sandboxes: keep all of a task's tool calls on the same sandbox. Per-call load-balancing across replicas breaks stateful tools (filesystem, memory, git, MongoDB), which assume a consistent view within a task.

What's Included

• Agent harness (services/agent-harness/, TypeScript) — multi-turn agent loop, talks to the sandbox via MCP_SANDBOX_URL.

• Agent environment (services/agent-environment/, Python) — Dockerized sandbox serving the 36 MCP servers over HTTP.

• Scoring (services/scoring/, Python) — LLM-as-judge claim-coverage scoring.

• Diagnostics (services/diagnostics/, Python) — failure-mode classification across an 11-mode taxonomy.

Citation

If you use MCP-Atlas in your research, please cite:

@misc{bandi2026mcpatlas,
  title         = {MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers},
  author        = {Bandi, Chaithanya and Dumitru, Razvan-Gabriel and Hertzberg, Ben and Agarwal, Divyansh and Boo, Geobio and Polakam, Tejas and Hassaan, Sami and Da, Jeff and Kim, HiJae and Gupta, Vipul and Sharma, Manasi and Park, Andrew and Dimakis, Martin and Hernandez Montoya, Ernesto Gabriel and Rambado, Dan and Salazar, Ivan and Cruz, Rafael and Rezaei, MohammadHossein and Rane, Chetan and Levin, Ben and Zhang, Daniel Yue and Kenstler, Brad and Liu, Bing},
  year          = {2026},
  eprint        = {2602.00933},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2602.00933}
}