LemmavsQwenPaw
Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.
Lemma
Lemma · Agent harness
OSS
Panel
6.82 spec wins
- Reliability
- 6.2
- Usefulness
- 6.8
- Cost
- 7.7
- Longevity
- 6.3
“You install the platform of the future with uv, which is at least honest about which decade the tooling comes from.”
QwenPaw
AgentScope (Alibaba) · Agent harness
OSSMCP
Panel
6.84 spec wins
- Reliability
- 6.3
- Usefulness
- 6.7
- Cost
- 7.3
- Longevity
- 6.7
“The cloud quick start includes a reminder to set the Studio to non-public, so strangers cannot control your assistant.”
Spec by spec
| Spec | Lemma | QwenPaw |
|---|---|---|
| Architecture | ||
| Category | Agent harness | Agent harness |
| Runs | local, cloud | local, cloud |
| Platforms | macos, linux, windows, web | macos, linux, windows |
| Context window | not documented | not documented |
| Protocols | ||
| MCP client | No | Yes |
| MCP server | No | No |
| Capabilities | ||
| Runs terminal commands | Yes | Yes |
| Multi-file edits | YesA pod is authored by the coding agent you already use, and pod runs dispatched through Agent Host let that agent change files; Lemma itself edits pod definitions. | No |
| Git operations | No | No |
| Browser control | No | Yes |
| Sandboxed execution | No | Yes |
| Multi-agent orchestration | Yes | Yes |
| Headless / CI mode | YesWorkflows are triggered by schedules, webhooks, table events, chat or the API, with optional human approval steps. | No |
| Models | ||
| Backbone | Claude Code, Codex, OpenCode, Cursor, any Anthropic-compatible endpoint, any OpenAI-compatible endpoint | Qwen, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, OpenRouter, QwenPaw-Flash, Ollama, LM Studio |
| Bring your own model | Yes | Yes |
| Local models | YesServer-run agents accept any Anthropic- or OpenAI-compatible endpoint, which covers a self-hosted gateway or a local model. | Yes |
| Cost | ||
| Pricing model | byok | byok |
| Starts at | $0/mo | $0/mo |
| Free tier | Yes | Yes |
| Bring your own key | Yes | Yes |
| Openness | ||
| Open source | Yes | Yes |
| License | AGPL-3.0 | Apache-2.0 |
| GitHub stars | 475 | 35,413 |
Which one would each critic pick
| Critic | Lemma | QwenPaw | Pick |
|---|---|---|---|
| El Juez | — | — | not enough reviews |
| El Amigo | 6.8 | 7.3 | QwenPaw — Pick QwenPaw if you want a personal agent in DingTalk, Lark or iMessage that runs on its own small models without a key; pick NanoClaw if you would rather have Docker walls and Claude. |
| El Crítico | 6.3 | 6.0 | Lemma — Agent Host dispatches to a paired machine, so a run succeeds or fails on whether that machine is awake, online and still authenticated. |
| El Profesor | 7.0 | 6.5 | Lemma — The system is authored by a coding agent and then verified by the CLI, which puts generation and validation on opposite sides of a readable file boundary. |
| La Inversora | 6.3 | 6.8 | QwenPaw — Alibaba's AgentScope team, with one-click deployment to Alibaba Cloud ECS, ModelScope Studio and a free always-on AgentScope Platform: the product is a funnel into DashScope and the cloud. |
| La Jefa | 6.5 | 6.0 | Lemma — Work starts from a schedule, a webhook or a table event with approval steps in the middle, which is the first question procurement asks about automation. |
| El Hacker | 7.8 | 8.0 | QwenPaw — Apache-2.0, Ollama at 32k context or LM Studio, a YAML Tool Guard from STRICT to OFF that inspects every call, MCP plus A2A plus ACP drivers, and a persona I edit as SOUL and PROFILE files. |
Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.