DeerFlow

You will like this if you want the Manus experience on your own box: make setup walks you through it in a couple of minutes, a web UI comes up on localhost, and the agent researches, writes and runs code in a sandbox you control. The daily trait is the sandbox choice, local for a laptop, Docker for a server, so the same agent scales with you.

You will not like the toolchain: Node, pnpm, uv and nginx all have to be present before make dev works. Pick it for a team that self-hosts everything. Pick Manus if you want the result without the ops.

The risk is inheritance. Version 2.0 is described as a ground-up rewrite with no shared code from v1, which is kept on a 1.x branch. The star count, the issue history and the community answers all belong to the old code. What you install today has the maturity of its own commit log, not the project's, and the row on this board has not been verified against it.

The consequence: treat it as a new project that happens to have a famous name, and read the 2.0 issues only. What it does right is SkillScan, a deterministic scanner that checks a skill offline before a model ever sees it.

The design is legible because it is written down. 1. Context: long-term memory with progressive loading, so the agent reads summaries before it reads history. 2. Planning: a plan mode with a TodoList, documented in the backend docs. 3. Actions: skills, each declaring an allowed-tools policy, so capability is scoped per skill rather than per session. 4. Subagents for delegation over the graph.

Verification is absent as a first-class step; nothing in the documented loop checks an outcome except the next model call. No benchmark is published. The backend docs do include a sandbox memory profile, which is more measurement than most harnesses on this board offer.

There is no price and no plan to have one, which is fine when the parent is ByteDance. The tell is the model advice: the README recommends DeepSeek and Kimi, not the house Doubao, which means this is a research group's project rather than a distribution play for the company's own inference. The adoption signal is real, first place on GitHub Trending on 28 February 2026, but attention is not revenue.

Moat: none that survives a reorg. Likely path: no acquirer, since it is already owned; the pivot is either into a hosted product or into quiet maintenance. Position: use it, fork it, and do not assume the maintainers report to anyone who cares about your roadmap.

The demo is a research agent that writes a report and runs the code. Procurement finds more than usual for a free project: OIDC single sign-on is documented, authentication design covers per-user isolation and CSRF, and personal access tokens exist for non-interactive clients. Kubernetes deployment is supported through a provisioner, and the multi-worker gateway requires Postgres and Redis, which is an ops cost but a known one.

Sixty engineers cost nothing in licences and a real platform team in hosting. Onboarding is heavy because the stack is wide. No vendor to call. Approved with conditions: a platform owner assigned and an internal SLA before the first team depends on it.

MIT and split cleanly into files I can version: config.yaml for the runtime, extensions_config.json for MCP servers and runtime skills, .env for the secrets. vLLM is a first-class backend with reasoning tokens handled, so my own GPU serves the model, and if I want Claude the harness will ride a Claude Code OAuth login instead of a separate key.

The bit I did not expect is the API: POST /api/runs/wait gives me a synchronous run from a shell script, and Langfuse tracing means my traces stay on my box. It is a lot of machinery for one person. But every piece has a file, and every file is mine.