MCP-Persona: Benchmarking LLM Agents on Real-World Personal Applications via Environment Simulation

Personas, identity, and agents2026arXivApproved editorial review

Authors: Wenhao Wang, Peizhi Niu, Gongyi Zou, Xiyuan Yang, Jingxing Wang, Haoting Shi, Yaxin Du, Jingyi Chai, Xianghe Pang, Shuo Tang, Yanfeng Wang, Siheng Chen

Keywords: Personalized tool-use benchmark, Stateful MCP simulation, Tool-using agents, Synthetic application contexts, Benchmark reproducibility risks

Source: Open primary source (opens in a new tab)

12
Authors
9
Findings
15
Limitations
4
Evidence

Editorial summary

English

MCP-Persona is a benchmark of tool-using agents in stateful local environments inspired by personal applications. Persona here does not mean human personality or a psychometric profile: it is a context tree containing users, chats, calendars, posts, files, and relations that tools can read or modify. Construction has three stages. Tool-Traverse executes human-authored valid calls and LLM-generated invalid calls against real MCP servers; another LLM summarizes the traces and writes Python kernels intended to reproduce responses and errors. Context-Tree derives entity hierarchies from schemas and traces, fills fields through enumeration, free-form text, constrained random generation, or sanitized authentic text, and links entities. Persona-Gen first samples tool chains, generates instructions that require them, injects context values, hides parameters a user might omit, and submits instruction-context-chain triples to human review. The released English file contains 173 unique task IDs and instructions, 959 call positions, 139 tools, and 18 server prefixes; mean chain length is 5.54 and the maximum is 18. Agents receive necessary and distractor context and can act for up to twenty rounds. GPT-4o scores checkpoints from zero to one; Acc averages checkpoints per task, SR@0.8 counts tasks above .8, and Exec-Acc evaluates searches and state changes observed in the sandbox. The main table compares thirteen models. None exceeds 50 percent Acc or Exec-Acc. Claude Sonnet 4.5 leads with 38.66 Acc, 10.40 SR@0.8, and 41.50 Exec-Acc; GPT-5 obtains 36.99, 6.94, and 41.45. Qwen3-Max-Latest is the strongest listed open model on Acc and Exec-Acc at 27.54 and 29.23. Performance varies sharply by family: o4-mini reaches 53.83 on Email, while content management and many cross-server combinations are difficult. Tool-Traverse validation uses only fifty Lark traces, twenty-five valid and twenty-five failed, over fourteen tools, reconstructing each exact precondition. It reports 94 percent accuracy and 93.8 F1 versus 58 percent and 53.3 for a documentation-only simulator; serialized-JSON similarities are .739 TF-IDF, .737 ROUGE, .741 BLEU, and .870 METEOR. This is promising evidence for that slice, not behavioral equivalence across all servers. The so-called human-LLM correlation revisits only GPT-5 decisions and reports 82 disagreements among 970 checkpoints, or 91.5 percent agreement, but provides no coefficient, kappa, annotator count, independence, uncertainty, or adjudication. Ablations are also mixed. On Lark, the authors' guide raises GPT-5 from 37.50/64.29 to 45.00/80.36 Acc/Exec-Acc; on Rednote some guides reduce scores. Restricting tools to ground-truth servers improves many execution cells but uses oracle information and hurts others; adding distractors lowers GPT-5 Acc from 41.04 to 36.99 while raising Exec-Acc from 29.25 to 39.15. The release audit finds discrepancies that prevent table reproduction. The paper refers to twenty-four servers with twelve personalized servers, the conclusion to twelve simulated servers, the README to eighteen and 139 tools, and an ablation to 140 tools. Released tasks do contain eighteen prefixes and 139 tools, but the repository ships only seventy-nine kernels in eight simulator directories; universal_email appears in forty-two chain positions and has no released simulator. More critically, the released evaluator constructs 724 checkpoints, not the 970 in the human study. Thirty-five tasks store the whole plan as one string and the code scores it as one checkpoint. English data are not fully localized either: 172 of 173 ground-truth arrays are identical to the Chinese version; sixty-five retain Chinese in execution targets and 149 in checkpoint annotations. Task 1 literally requests the English title '2025 Q4 Team Review Meeting' while its checkpoint requires '2025Q4团队复盘会议', so faithful execution can conflict with the hidden target. The repository is not runnable end to end. The README clone URL returns Repository not found; one source file fails compilation with two indentation errors; evaluators import agentoolkit and prompts/configuration from private /data/JohnDoe paths; the runner that generates results and sandboxes is absent; configuration names nonexistent task files; and matplotlib plus agentoolkit are missing from dependencies. API and judge arguments are required but not wired into the pipeline. When an agent result or sandbox context is missing, evaluation skips it rather than assigning zero and never verifies final coverage, potentially inflating means. Outputs, item scores, human labels, the fifty fidelity traces, cost logs, authors' skills, and aggregation scripts are not published; there is also no actual license, test suite, CI, lockfile, or container. Rankings have no repeated runs, intervals, or tests, and privacy, authorization, phishing, and destructive-action risks are discussed but not evaluated. The defensible contribution is an inspectable set of complex tasks, contexts, chains, and many simulators, plus limited evidence that traversal improves a Lark simulator. It does not demonstrate synthetic humans, user realism, psychometric validity, safety, empirical privacy, equivalence with complete real systems, or end-to-end reproducibility.

Español

MCP-Persona es un benchmark de agentes que usan herramientas en entornos locales con estado inspirados en aplicaciones personales. En este trabajo, persona no significa personalidad humana ni perfil psicométrico: es un árbol de contexto con usuarios, chats, calendarios, publicaciones, ficheros y relaciones que las herramientas pueden consultar o modificar. La construcción tiene tres etapas. Tool-Traverse ejecuta llamadas válidas creadas por anotadores y llamadas inválidas generadas por LLM contra servidores MCP reales; otro LLM resume las trazas y genera kernels Python que pretenden reproducir respuestas y errores. Context-Tree deriva jerarquías de entidades de esquemas y trazas, rellena campos mediante enumeración, texto libre, generación aleatoria o texto auténtico saneado y enlaza entidades. Persona-Gen muestrea primero cadenas de herramientas, genera instrucciones que las requieren, inserta valores del contexto, oculta parámetros que un usuario podría omitir y somete las ternas instrucción-contexto-cadena a revisión humana. El fichero inglés publicado contiene 173 tareas con identificadores e instrucciones únicos, 959 posiciones de llamada, 139 herramientas y 18 prefijos de servidor; la cadena media tiene 5,54 pasos y el máximo es 18. Los agentes reciben contexto necesario y distractores y pueden actuar hasta 20 rondas. GPT-4o puntúa checkpoints de 0 a 1; Acc promedia esos checkpoints por tarea, SR@0.8 cuenta tareas por encima de 0,8 y Exec-Acc evalúa búsquedas y cambios observados en el sandbox. La tabla principal compara trece modelos. Ninguno supera 50% en Acc o Exec-Acc. Claude Sonnet 4.5 lidera con 38,66 Acc, 10,40 SR@0.8 y 41,50 Exec-Acc; GPT-5 obtiene 36,99, 6,94 y 41,45. Qwen3-Max-Latest es el abierto mejor situado en Acc y Exec-Acc, 27,54 y 29,23. El rendimiento varía mucho por familia: o4-mini alcanza 53,83 en Email, mientras gestión de contenido y muchas combinaciones cross-server son difíciles. La validación de Tool-Traverse usa solo 50 trazas Lark, 25 válidas y 25 fallidas, sobre 14 herramientas, reconstruyendo para cada una el estado previo exacto. Informa 94% de exactitud y F1 93,8 frente a 58% y 53,3 para un simulador basado solo en documentación; las similitudes de JSON son .739 TF-IDF, .737 ROUGE, .741 BLEU y .870 METEOR. Es evidencia prometedora para ese bloque, no equivalencia conductual de todos los servidores. La llamada «correlación» humano-LLM revisa únicamente decisiones de GPT-5 y reporta 82 discrepancias entre 970 checkpoints, 91,5% de coincidencia, pero no da coeficiente, kappa, número de anotadores, independencia, incertidumbre o adjudicación. Las ablations también son mixtas. En Lark, la guía propia eleva GPT-5 de 37,50/64,29 a 45,00/80,36 Acc/Exec-Acc; en Rednote algunas guías reducen resultados. Restringir a los servidores del ground truth mejora muchas celdas de ejecución, pero usa información oráculo y perjudica otras; añadir distractores baja Acc de GPT-5 de 41,04 a 36,99 y a la vez eleva Exec-Acc de 29,25 a 39,15. La auditoría del release encuentra divergencias que impiden reproducir las tablas. El paper habla de 24 servidores con 12 personalizados, la conclusión de 12 simulados, el README de 18 y 139 herramientas, y una ablación de 140. Los datos publicados sí contienen 18 prefijos y 139 herramientas, pero el repositorio solo libera 79 kernels en ocho directorios; universal_email aparece 42 veces en las cadenas y no tiene simulador publicado. Más crítico: el evaluador publicado construye 724 checkpoints, no los 970 del estudio humano. En 35 tareas el plan completo está guardado como una cadena y el código lo trata como un único checkpoint. Los datos ingleses tampoco están plenamente localizados: 172/173 arrays de ground truth son idénticos a los chinos; 65 conservan chino en el objetivo de ejecución y 149 en anotaciones. La tarea 1 pide literalmente el título inglés «2025 Q4 Team Review Meeting» y su checkpoint exige «2025Q4团队复盘会议», de modo que una ejecución fiel puede chocar con el objetivo oculto. El repositorio no es ejecutable end-to-end. El URL de clonación del README devuelve Repository not found; un fichero falla al compilar por dos errores de indentación; los evaluadores importan agentoolkit y prompts/configuración desde rutas privadas /data/JohnDoe; falta el runner que genera resultados y sandboxes; la configuración referencia tareas inexistentes; faltan matplotlib y agentoolkit en dependencias. Los argumentos de API y juez son obligatorios pero no se conectan al pipeline. Si falta una salida del agente o un contexto sandbox, la evaluación la omite en vez de puntuar cero, sin comprobar cobertura final, lo que puede inflar medias. No se publican salidas, puntuaciones por ítem, etiquetas humanas, 50 trazas de fidelidad, costes, skills propios o scripts de agregación; tampoco hay licencia real, tests, CI, lockfile o contenedor. No hay repeticiones, intervalos o pruebas para los rankings, y los riesgos de privacidad, autorización, phishing y acciones destructivas solo se discuten, no se evalúan. La contribución defendible es un conjunto inspeccionable de tareas complejas, contextos, cadenas y muchos simuladores, más evidencia limitada de que el traversal mejora un simulador Lark. No demuestra personas humanas sintéticas, realismo de usuario, validez psicométrica, seguridad, privacidad empírica, equivalencia con sistemas reales completos ni reproducibilidad end-to-end.

Research question

Can a reproducible benchmark of personal tasks with simulated MCP servers be built from real traces, and how well do current LLM agents resolve chains with implicit context, persistent state, and coordination across tools?

Method

Synthetic and comparative benchmark. Tool-Traverse collects valid and failed real calls and generates Python kernels; Context-Tree constructs and fills state hierarchies; Persona-Gen derives fuzzy instructions from tool chains and reviews them manually. Thirteen leading models execute 173 tasks up to 20 rounds. GPT-4o scores 0/0.5/1 per checkpoint and a second protocol judges searches and sandbox changes. Validation of 50 Lark traces, human review of GPT-5 decisions, and ablations of skills, tool selection, and distractors are added. The independent audit reviews the 20 pages, TeX, both JSON, counts, kernels, evaluators, configuration, and official commit.

Sample: The main unit is 173 synthetic and manually reviewed tasks, not users. The chains have between 1 and 18 calls, mean 5.54. The simulation validation uses 50 balanced Lark traces; the human-judge review declares 970 decisions of GPT-5, although released data and code form 724 checkpoints.

Findings

  • Claude Sonnet 4.5 leads the main average with 38.66 Acc and 41.50 Exec-Acc; no model exceeds 50% globally.
  • Claude Sonnet 4.5 obtains the highest SR@0.8, 10.40, versus 6.94 for GPT-5; completing almost an entire task remains rare.
  • Qwen3-Max-Latest is the open model with the best published Acc and Exec-Acc, 27.54 and 29.23.
  • Tool-Traverse reaches 94% accuracy and 93.8 F1 on 50 Lark traces, with much more textual similarity than Vanilla.
  • Interface-aligned guides help markedly in several Lark cells, but have mixed effects on Rednote.
  • Oracle selection of servers improves many long executions but not all metrics or models.
  • Distractors do not produce a uniform effect: they can lower Acc and raise Exec-Acc simultaneously.
  • The declared human-judge agreement is 91.5% for GPT-5, not a general statistical correlation.
  • The release does not reproduce its own main counts and omits a substantial part of the runtime and the results evidence.

Limitations

  • Persona is synthetic state of applications and not personality, psychometrics, or human behavior.
  • Tasks are generated from target chains and may penalize valid alternative solutions.
  • Fidelity is validated only on 50 traces of one server and with reconstructed prior state.
  • JSON similarity metrics do not prove functional equivalence outside the observed calls.
  • Human-judge validation covers one model and omits methodology, kappa, intervals, and adjudication.
  • The paper does not reconcile 24/18/12 servers, 140/139 tools, or 970/724 checkpoints.
  • English ground truths retain Chinese targets that may contradict translated literals.
  • There are no repetitions, seeds, intervals, or significance for rankings and ablations.
  • Comparisons do not match models, context, price, tokens, retries, or tools.
  • Runner, sandboxes, agentoolkit, prompts, private configuration, and several simulators/schemas are missing.
  • The evaluator omits failures and missing contexts without checking total coverage.
  • Outputs, per-item scores, human labels, fidelity traces, costs, or skills are not published.
  • The code does not compile completely, the quick start clones a nonexistent repository, and dependencies are missing.
  • There is no real license, tests, CI, lockfile, container, or versioned manifest.
  • Privacy, authorization, injection, destructive actions, and harm are only discussed, not measured.

What the study does not establish

  • It does not demonstrate synthetic personality, psychometric profile, or cognitive simulation of users.
  • It does not demonstrate that generated tasks represent frequency, intention, or behavior of real users.
  • It does not demonstrate fidelity of twelve servers from fifty Lark traces.
  • It does not demonstrate that a reference chain is the only valid solution.
  • It does not demonstrate robust rankings without uncertainty and public outputs.
  • It does not demonstrate privacy, authorization security, or resistance to abuse.
  • It does not demonstrate that the English benchmark is aligned with its ground truth.
  • It does not allow reproducing end-to-end results, ablations, costs, or validations with the current public checkout.

Traceability

Scope: Full text

Version: arXiv:2606.02470v1

Consulted source: https://arxiv.org/abs/2606.02470v1

Review: Codex twenty-page full-text visual, TeX, bilingual-task, checkpoint-count, simulation-validity, human-judge, repository-coverage, compilation and evaluator-code audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Claude Sonnet 4.5
  • GPT-5
  • Claude Opus 4.1
  • o4-mini
  • o3
  • GPT-4o como agente y juez
  • Grok-4
  • Gemini 3 Pro
  • Gemini 2.5 Pro
  • Qwen3-Max-Latest
  • Qwen3-235B-A22B
  • DeepSeek-V3
  • Qwen3-Coder
  • Kimi-K2.5, MiniMax-M2.5 y Qwen3.5-Plus en ablations

Instruments and metrics

  • Checkpoint Accuracy, puntuación GPT-4o de 0, 0,5 o 1
  • Success Rate at 0.8
  • Execution Accuracy sobre búsquedas y estado CRUD
  • Matriz TP/TN/FP/FN de fidelidad Lark
  • Accuracy, precision, recall y F1 de comportamiento simulado
  • TF-IDF, ROUGE, BLEU y METEOR sobre JSON serializado
  • Coincidencia humano-GPT-4o en checkpoints de GPT-5
  • Coste, tokens y pasos medios por tarea

Data used

  • MCP-Persona inglés: 173 tareas, 959 posiciones de llamada, 139 herramientas y 18 servidores
  • MCP-Persona chino: 173 traducciones con cadenas, contextos y casi todos los ground truths compartidos
  • Release de 340 checkpoints de ejecución: 118 personalized_search y 222 operate
  • Release evaluable de 724 checkpoints narrativos, frente a 970 declarados en el paper
  • 50 trazas Lark reales/simuladas declaradas, no publicadas
  • 79 kernels Python publicados en ocho directorios de simulador

Evidence and location

  • Metadata, version, acceptance, and extension: Official arXiv record 2606.02470v1, checked 2026-07-17
  • Method, results, tables, ablations, impact, guide, and prompts: arXiv v1, all twenty PDF pages and complete TeX source
  • Tasks, ground truth, kernels, evaluators, and reproducibility: wwh0411/MCP-Persona commit b510f5a, English and Chinese JSON, checked 2026-07-17
  • Audit of persona scope, localization, counts, omissions, compilation, and claims: reports/verification/article-308-mcp-persona-persona-scope-bilingual-ground-truth-checkpoint-count-simulation-validity-and-repository-audit.json