- Category
- AI Agents
- Rank
- No. 1168Tools index
Previous survey · No. 1384 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Agent Building · Models: Train & Run · Model & Agent Evaluation
- Interfaces
- SDK
- Builder
- openbmb
- GitHub
- 833 stars
- Date
About
End-to-end infrastructure for training and evaluating LLM agents. Tooling for the full agent lifecycle from data to eval.
What it does
A research monorepo from Tsinghua's NLP lab, Renmin University and ModelBest that publishes small open-weight models together with the code used to build them. The centrepiece is a 4-billion-parameter search model, fine-tuned from Qwen3-4B-Thinking, that plans web lookups and tool calls over long sessions; the authors say it sustains more than 100 turns. Around it sit three projects: a Docker stack that hosts sandboxed tools behind MCP (Model Context Protocol, a standard way to expose tools to a model), a reinforcement learning stack that runs sampling and weight updates asynchronously, and a benchmark harness. A second, 8B report-writing model ships through a separate repository, UltraRAG.
Why it's ranked here
The case rests on openness at a small size. The code is Apache 2.0, the weights are on Hugging Face and ModelScope, there are two arXiv technical reports, and the release includes the pipeline behind the model rather than only a checkpoint. The authors' own table puts the 4B model at 63.9% on text-only GAIA, against 66.4% for an 8B open model and 70.9% for a 30B one. Those figures are self-reported. Shipping the tool sandbox and the training code alongside the weights is what makes this worth studying rather than only running.
What's good
The tool sandbox is a real deployment, not a sketch: one compose file brings up a management dashboard, a MongoDB store and two tool servers, one general purpose and one for search, crawling and code execution, each with explicit CPU and memory caps. The client that talks to it turns every failure, whether a timeout, an HTTP error or a dropped connection, into a structured error result instead of an exception, so one flaky tool does not kill a long run. The training stack pins every dependency to an exact version, and the quickstart refuses to start until a model name and a real API key are set.
Tradeoffs
This is research code with heavy hardware assumptions. The default tool containers are capped at 8 cores and 32GB for one node and 16GB for the other, and the training stack pins a CUDA 12.6 PyTorch build, flash attention and SGLang among about 340 packages, desktop automation and OCR libraries included. The Docker socket is mounted into two containers, which hands them control of the host's Docker daemon. Configuration means editing constants inside a Python script, which then writes your API key in plain text into a generated shell script. The demo scripts default to port 8000, while the compose file publishes the manager on 8080.
How to use it well
Pick it up to study or reproduce how a small open model is taught long tool-use sequences, or to borrow a self-hosted MCP tool farm for your own experiments. Start from the published evaluation Docker image rather than a local Python install, and point the manager address at the port your compose stack actually exposes. Run the sandbox on an isolated machine, given its access to the Docker daemon. If what you want is a private long-form research writer over your own documents, go to the Report model through UltraRAG instead; this repository is the place for the search model and its training and test rig.
Technical notes+
AgentCPM-Explore/AgentDock/docker-compose.yml defines agentdock-manager (port 8080, command '--workers 4 -t 600', health check on /alive), mongo:7 with 27017 published to the host, agentdock-node-full (8092 to 8088 for streamable-HTTP MCP, 8004 to 8000 for the API, limits of 8 CPUs and 32G) and agentdock-node-explore (8102 and 8014, 4 CPUs and 16G, JINA_API_KEY and GOOGLE_SERP_API_KEY injected); the manager and the explore node both bind-mount /var/run/docker.sock. AgentCPM-Explore/src/mcp_manager.py wraps httpx.AsyncClient (30s default timeout), reads /servers, /tools and /server/{id}/tools, rewrites tools into OpenAI function schema, and maps ReadTimeout, HTTPStatusError and RequestError to {status: error} dicts. AgentCPM-Explore/quickstart.py writes a JSONL task and a run script under AgentToLeaP/benchmarks/quickstart with MAX_INTERACTIONS=30, MAX_TOKENS=4096 and --skip-post-eval, exporting API_KEY inline; its MANAGER_URL default is http://localhost:8000, and AgentCPM-Explore/src/run_demo.py defaults to the same host with an /mcpapi suffix. AgentCPM-Explore/AgentRL/requirements.txt pins torch==2.8.0+cu126, flash-attn==2.8.3 and sglang==0.5.5.post3, and lists mcp==1.24.0 twice.
Observed
- License
- Apache-2.0
- Language
- Python
- Base model
- Qwen3-4B-Thinking-2507, 4B parameters
- Interfaces
- MCP tool servers over streamable HTTP, Python scripts, web management dashboard
- Packaging
- Docker Compose stacks, prebuilt Docker Hub images, model weights on Hugging Face and ModelScope
- Platform support
- Evaluation image for amd64 and arm64; training stack pinned to CUDA 12.6 PyTorch builds
- Stated requirements
- Docker and Docker Compose, Python 3.10+, 8GB RAM minimum with 16GB+ recommended
- Repository layout
- Model code, tool sandbox, RL training stack and evaluation harness in one monorepo; the report model deploys from a branch of the separate UltraRAG repository
Read from README.md, LICENSE, AgentCPM-Explore/README.md, AgentCPM-Explore/quickstart.py, AgentCPM-Explore/src/run_demo.py, AgentCPM-Explore/src/mcp_manager.py, AgentCPM-Explore/AgentDock/README.md, AgentCPM-Explore/AgentDock/docker-compose.yml, AgentCPM-Explore/AgentRL/README.md, AgentCPM-Explore/AgentRL/requirements.txt.
What it can do
Train LLM agents
Training data and model configuration → Trained LLM agent models
Evaluate agent performance
Trained agents and evaluation datasets → Performance metrics and benchmarks
Manage agent training data
Raw training datasets → Processed and formatted training data
Deploy trained agents
Trained agent models → Deployed agent instances
Monitor agent lifecycle
Agent instances and usage data → Lifecycle tracking reports
Execute agent evaluations
Agent models and test scenarios → Evaluation results and analytics
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.