Busabase Was #1 on Product Hunt: An Open-Source Database Where Agents Leave Their Work Behind
Busabase took the top spot on Product Hunt on Thursday with 354 upvotes, calling itself the general system of record for AI agents. The problem statement is one every Claude Code or Codex user recognizes: every session starts from zero, your CLAUDE.md lives in one repo, the best output dies in a chat window, and switching tools means explaining the project all over again. Busabase's answer is a shared workspace, open source under MIT, where agents and people read the same typed records, docs, files, skills and small apps, and where every agent write comes back as a change request with a message, a field-level diff, an author and a history.
The mechanics matter. You run npx busabase server and get a local workspace on an embedded Postgres with no account and no config. You paste one URL into your agent, Claude Code, Codex, Cursor, Gemini CLI or anything that can read a web page, and the agent reads a SETUP_SKILL.md that teaches it the workspace. Connections also work through MCP, OpenAPI, a CLI, or an ACP chat session. Skills and playbooks live as files in the workspace so every connected agent can look them up before acting. Permissions decide whether an agent's change merges immediately or waits for a human review. There is a desktop app, a Docker image and a hosted cloud version on the same engine. The repo has 445 stars and started in June, so this is a launch of a product that has been quietly built for four months.
What makes it more than another notes app is the review gate on writes. The memory-layer conversation in agent research this month has turned to write-time curation: what to admit into persistent memory and how to present it, as one paper this week put it. Busabase is that idea as a product, where the admission control is a diff a person can approve, and the stored object is a typed record rather than a blob of prose. It is also deliberately model-agnostic, with no built-in LLM, which is the right call for a layer that is supposed to outlive whichever agent wrote to it.
The open question is whether teams will accept one more place their data lives. The pitch is that it replaces the sprawl of chat transcripts, scattered skill files and per-tool memory with a single reviewed store, and the Product Hunt win says builders want that. The harder test is six months from now, when the question is whether agents actually write back to it without being told.
Repo: https://github.com/busabase/busabase
Site: https://busabase.com
← Back to all articles
The mechanics matter. You run npx busabase server and get a local workspace on an embedded Postgres with no account and no config. You paste one URL into your agent, Claude Code, Codex, Cursor, Gemini CLI or anything that can read a web page, and the agent reads a SETUP_SKILL.md that teaches it the workspace. Connections also work through MCP, OpenAPI, a CLI, or an ACP chat session. Skills and playbooks live as files in the workspace so every connected agent can look them up before acting. Permissions decide whether an agent's change merges immediately or waits for a human review. There is a desktop app, a Docker image and a hosted cloud version on the same engine. The repo has 445 stars and started in June, so this is a launch of a product that has been quietly built for four months.
What makes it more than another notes app is the review gate on writes. The memory-layer conversation in agent research this month has turned to write-time curation: what to admit into persistent memory and how to present it, as one paper this week put it. Busabase is that idea as a product, where the admission control is a diff a person can approve, and the stored object is a typed record rather than a blob of prose. It is also deliberately model-agnostic, with no built-in LLM, which is the right call for a layer that is supposed to outlive whichever agent wrote to it.
The open question is whether teams will accept one more place their data lives. The pitch is that it replaces the sprawl of chat transcripts, scattered skill files and per-tool memory with a single reviewed store, and the Product Hunt win says builders want that. The harder test is six months from now, when the question is whether agents actually write back to it without being told.
Repo: https://github.com/busabase/busabase
Site: https://busabase.com
Comments