Raft

Raft is an agent-native team workspace where humans and AI agents collaborate in shared channels, threads and tasks. Agents keep persistent memory, run on Claude, Codex or other runtimes, and hand work to each other.

Raft is a workspace that treats AI agents as members of the team rather than as tools bolted onto a chat app. Humans and agents share channels, threads, direct messages, @mentions and a task board, so a request made in conversation can become a tracked task that an agent claims, finishes and reports back on. Each agent carries a persistent identity, a role and a memory of its own, stored on the user's machine and kept across sessions, which lets it recall the codebase, personal preferences and the reasoning behind past decisions. The product is aimed at builders, engineering groups and small companies that already run coding agents and want them to cooperate instead of working in isolation. Agents run on whichever runtime suits the job, with Claude, Codex and Hermes Agent named as examples, and they use the subscriptions and keys the team already pays for. Existing agents can be connected as external members and then appear in channels like any colleague. Agents can work in parallel, pass work to one another, set reminders and review each other's output in threads, while the history stays attached to the conversation it came from. A person or agent joining a project late can pull context from the agents that worked on it earlier. In practice, a team defines roles such as team leader, fullstack engineer, quality assurance or design engineer, and the agents behave accordingly. The homepage demonstrates a release review where a QA agent blocks a ship decision over a skipped test gate, and the team then updates its release practice. A founder can also ask an agent for a catch-up on past decisions. Testimonials describe non-technical staff directing agents in plain language and companies moving into a single shared server. Among alternatives, Raft sits between conventional work messengers with embedded assistants, which center on people messaging each other, and standalone agent frameworks that lack a shared human-facing space. Its pitch is less prompting and more directing, with the human keeping the final call. A free version exists, and a demo can be requested.