---
title: SimLab HR — AI Recruiting & People Management Agent Environment
emoji: 👔
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
short_description: "MCP gym for benchmarking & training AI HR agents"
tags:
 - openenv
 - hr
 - human-resources
 - recruiting
 - hrms
 - agent-evaluation
 - agent-benchmark
 - simlab
 - reinforcement-learning
 - rl-environment
 - ai-agent
 - tool-use
 - function-calling
 - enterprise
 - multi-tool
 - gymnasium
 - gym
 - benchmark
 - mcp
 - model-context-protocol
 - reward-model
 - verifier
 - collinear
 - langchain
 - openai
 - sandbox
 - docker
 - toolbench
 - swe-bench
 - bfcl
pinned: true
license: apache-2.0
---

# SimLab HR — MCP-Powered Gym for AI HR Agent Evaluation

A fully-functional HR simulation gym for training, evaluating, and benchmarking AI recruiting and people management agents. Test your agent's tool-use and function-calling abilities across 4 MCP tool servers — HRMS, email, calendar, and team chat — with automated rubric-based evaluation.

Built on [OpenEnv](https://github.com/meta-pytorch/OpenEnv) and powered by [SimLab](https://github.com/collinear-ai/simlab).

Unlike single-API function-calling benchmarks like [BFCL](https://github.com/ShishirPatil/gorilla) or [ToolBench](https://github.com/OpenBMB/ToolBench), SimLab HR gives your agent a full workplace — HRMS, email, calendar, and team chat — and asks it to complete real multi-step HR workflows end-to-end.

## 4 MCP Tool Servers, 1 Environment

| Tool Server | Port | What it does |
|---|---|---|
| **HRMS** (Frappe) | 8030 | Employee records, leave management, attendance, payroll |
| **Email** (MailHog) | 8040 | Send and read emails, inbox management |
| **Calendar** (Baikal/Chronos) | 8050 | Schedule meetings, check availability, manage events |
| **RocketChat** | 8060 | Team messaging, channels, direct messages |

Agents must reason and coordinate tool calls across all four MCP servers to complete real HR workflows — the kind of multi-step function calling that separates real tool-use from single-API benchmarks.

## Quickstart

```python
from simlab_hr import HRAction
from simlab_hr.client import HREnv

client = HREnv(base_url="http://localhost:8000")

with client:
    obs = client.reset()
    print(obs.observation.task_instruction)
    print(obs.observation.tools_available)  # {'hrms': [...], 'email': [...], ...}

    # Check leave balance in HRMS
    result = client.step(HRAction(
        tool_server="hrms",
        tool_name="get_leave_balance",
        parameters={"employee_id": "EMP-0042"}
    ))

    # Send an email notification
    result = client.step(HRAction(
        tool_server="email",
        tool_name="send_email",
        parameters={"to": "manager@company.com", "subject": "Leave approved", "body": "..."}
    ))
```

## Benchmark Tasks

**8 sample tasks** covering real HR workflows across three difficulty levels:

| Difficulty | Example Tasks |
|---|---|
| Easy | Approve a leave request, update an employee's designation |
| Medium | Schedule a phone screen + send confirmation, run an attendance report |
| Hard | Multi-person panel interview scheduling, full new-hire onboarding flow |

Every task requires the agent to coordinate function calls across multiple MCP tool servers — this is what makes it hard.

## Automated Evaluation

SimLab HR includes a rubric-based LLM judge that evaluates agent performance after each episode:

- **0.8–1.0**: All requirements fully met with clear evidence
- **0.6–0.8**: Core requirements met with minor gaps (0.6 = PASS threshold)
- **0.4–0.6**: Partial completion, significant gaps remain
- **0.0–0.4**: Minimal or no meaningful progress

Configure the verifier model:

```bash
export VERIFIER_MODEL="gpt-4o"
export VERIFIER_API_KEY="sk-..."
```

## Run Locally

```bash
git clone https://github.com/collinear-ai/simlab.git
cd simlab/envs/simlab_hr

# Start all services (HRMS, Email, Calendar, RocketChat, OpenEnv wrapper)
docker compose up

# First run pulls ~10 images and takes a few minutes for HRMS to initialize
```

Or run from Hugging Face:

```python
from simlab_hr.client import HREnv

client = HREnv.from_hub("collinear/simlab-hr")
```

## Unlock 14+ Tasks from the API

This environment ships with 8 sample tasks. Want more?

Set your Collinear API key to unlock the full task set with real HR scenarios:

```bash
export COLLINEAR_API_KEY="your-key-here"
```

Get a free API key at **[platform.collinear.ai](https://platform.collinear.ai)** (Developer Resources → API Keys).

With the API key, every `reset()` pulls a fresh task from Collinear's Scenario Manager — recruiting workflows, people management scenarios, compliance tasks, and more.

## Use with TRL / GRPOTrainer

Compatible with Hugging Face TRL for reinforcement learning fine-tuning:

```python
from simlab_hr import HRAction
from simlab_hr.client import HREnv

env = HREnv.from_hub("collinear/simlab-hr")
with env:
    obs = env.reset()
    # ... your RL training loop
```

## How SimLab HR Compares

Most tool-use benchmarks evaluate function calling in isolation — single API calls with predefined schemas. SimLab HR tests whether your agent can actually get work done across a real enterprise environment.

| | SimLab HR | [BFCL](https://github.com/ShishirPatil/gorilla) | [ToolBench](https://github.com/OpenBMB/ToolBench) | [EnterpriseOps-Gym](https://github.com/ServiceNow/EnterpriseOps-Gym) | [tau-bench](https://github.com/sierra-research/tau2-bench) |
|---|---|---|---|---|---|
| What it tests | Multi-tool HR workflows end-to-end | Function call accuracy (single/parallel) | API discovery & chaining across 16k APIs | Enterprise planning across 8 domains | Customer service policy compliance |
| Real backing services | ✅ Frappe HRMS, MailHog, CalDAV, RocketChat | ❌ Schema validation only | ❌ API simulation | ❌ Mock APIs | ❌ Simulated |
| MCP tool servers | ✅ 4 servers | ❌ | ❌ | ❌ REST APIs | ❌ |
| Multi-step workflows | ✅ 10+ steps, cross-system | ❌ Single/parallel calls | ✅ Multi-hop chains | ✅ Avg 9 steps | ✅ Multi-turn |
| HR-specific | ✅ Dedicated | ❌ | ❌ | ✅ 1 of 8 domains | ❌ |
| Automated evaluation | ✅ Rubric-based LLM judge | ✅ AST matching | ✅ Pass rate + win rate | ✅ Expert-curated | ✅ Policy checks |
| RL / Gymnasium support | ✅ OpenEnv-compatible | ❌ | ❌ | ❌ | ❌ |
| Task generation | ✅ API pipeline | ❌ | ❌ | ❌ | ❌ |

## More Environments

SimLab includes **5 enterprise simulation scenarios** with **14 MCP tool servers**:

| Scenario | MCP Tool Servers |
|---|---|
| **Human Resources** | HRMS, email, calendar, team chat ← *you are here* |
| **Customer Service** | Helpdesk ticketing, team chat, email |
| **Finance** | SEC filings, market data, Google Workspace |
| **Coding** | Sandboxed IDE, browser automation, team chat |
| **CRM** | Contacts, deals, pipelines, activities |

Install the full toolkit:

```bash
pip install simulationlab
simlab templates list
```

Learn more: [github.com/collinear-ai/simlab](https://github.com/collinear-ai/simlab) | [docs.collinear.ai](https://docs.collinear.ai)

## License

Apache 2.0 — [Collinear AI](https://collinear.ai)