camel
Multi-agent societies and synthetic training data from a research collective — a laboratory, not a runtime
What is camel?
CAMEL's core idea is agents talking to each other: a RolePlaying session pairs an AI user that issues instructions with an AI assistant that carries them out, and a Workforce hands split-up subtasks to workers under a coordinator. Around that sit eighty-odd toolkits, around fifty model backends, benchmark harnesses such as GAIA and BrowseComp, and pipelines that turn the resulting conversations into fine-tuning data. That breadth is the point if you are running experiments; if you only want one dependable agent in production you will carry a great deal you never use, and the library has been public since 2023 yet is still on 0.2.x.
What can you do with camel?
- Start with one agent and a toolkit —
ChatAgent(model=..., tools=SearchToolkit().get_tools())followed byagent.step(...)runs a single turn with tool calling and conversation memory already wired in. Pass a Pydantic class asresponse_formatand you get parsed structured output back instead of prose. - Have two agents talk the task through —
RolePlayingpairs an AI user that issues instructions with an AI assistant that carries them out, and by default rewrites your task into a more specific one before the first turn. Staying in role is prompt-level: both system prompts open with "Never flip roles!", an instruction handed to the model rather than a rule the framework enforces. You drive the session withinit_chat()and repeatedstep()calls, and it ends on your own round limit, on theCAMEL_TASK_DONEmarker, or on either agent coming back terminated — which happens by itself once the conversation outgrows its token budget. - Grow it into a Workforce with a coordinator —
add_single_agent_workerandadd_role_playing_workerbuild the team, andadd_workforcenests one workforce inside another. A planner agent splits the task, a coordinator hands each subtask to a worker by reading its description, and a failed subtask is met with a retry, a replan, a further split, a reassignment or a brand-new worker.pause(),resume()andsave_snapshot()let a person step in mid-run. - Reach the outside world through toolkits and MCP — Eighty-odd toolkits cover web search, a browser, the terminal, GitHub, Notion, Slack, arXiv and Excel, and
FunctionToolturns any Python function into a tool by reading the description written above it.MCPToolkitconnects out to external MCP servers; publishing in the other direction isrun_mcp_server()for a toolkit andto_mcp()for a ChatAgent or an entire Workforce. - Turn agent runs into fine-tuning data —
CoTDataGenerator,SelfInstructPipelineandSelfImprovingCoTPipelinewrite reasoning traces and instruction sets out as JSON ready for fine-tuning. Verifiers for maths, physics and Python check whether the answers are actually right, and the single- and multi-step environments return a reward for each move — the score reinforcement learning trains against. - Score the result on published benchmarks —
camel.benchmarksships harnesses for GAIA, BrowseComp, APIBank, APIBench, Nexus and RAGBench, so a design can be compared against numbers other people have published rather than against a demo that happened to work.
Before you choose camel
- Its built-in Python interpreter restricts what code it accepts instead of isolating the process; a way past that was reported against 0.2.90 in July 2026, and the docs reserve it for trusted code and point to Docker.Reported in#4224
- Toolkits and model connections come as optional add-ons that each bring their own dependencies and API keys, and the option that installs everything at once was reported in February 2026 to fail to resolve.Reported in#3851
Star history
17 Aug to 30 Oct · +399
Frequently asked questions
Is camel free for commercial use?
camel is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.
How can camel be deployed?
camel is available as Self-hosted.