Hi all,
I run a small web agency and I am looking for people who have tried something similar, or who think this is a bad idea.
My situation
I maintain around 40 git repositories (client websites, WordPress plugins, internal tooling). Most of the actual coding is done by Claude Code, Anthropic’s agentic coding tool. Until recently that meant up to 8 VS Code windows, one repo and one Claude session each. I have since moved to the Claude Code desktop app, which puts all sessions in one window with a sidebar. That is a real improvement, but the core problem is unchanged: I still have 6 to 8 parallel sessions, each one a chat I have to babysit. Which one is waiting for my answer? Which one finished? What did it change, and in which repo? The answers live inside eight chat transcripts, not in one place I can look at. A chat window is not a work management tool. Fibery is.
Fibery is already my system of record. I have an Operations space with Tasks, Initiatives and Routines, an Agents space with an agent catalog and an “Agent Run” database for telemetry, and every repo has a small “anchor” block in its config that tells the agent which Fibery entities it belongs to. The agent already reads and writes Fibery through the MCP server (and a small CLI I built on the API). Task states are used as a handoff protocol: Backlog, To Do, In Progress, Waiting (for my review), Done (only me).
What I want to get to
Fibery as the single interface for delegating work to coding agents, roughly:
- I create or move a Task to “To Do” and mark it as “for the agent”.
- Something (Fibery automation + webhook, or a scheduled job) starts a cloud Claude Code session in the right repository with the task description as the brief.
- The agent works, pushes a branch or PR, writes its report as a comment on the Task, logs an Agent Run and moves the Task to “Waiting”.
- I review in one Fibery board instead of scrolling through eight chat transcripts. Follow-up instructions go as comments on the Task and get forwarded to the running session.
- Code review itself stays on GitHub, Fibery only holds the links (branch, PR, session).
The missing pieces on the Fibery side are a Repo database linked to Projects/Websites/Initiatives, a few fields on Task (branch, PR link, session link), and the outbound trigger. The missing piece on the agent side is a reliable “write back to Fibery when done” step.
My questions
- Has anyone built a “Fibery as agent inbox / kanban for AI workers” setup? With Claude Code, Codex, Copilot coding agent, Devin, or anything else? What worked, what broke?
- Several tools now ship kanban-style agent collaboration out of the box (assign an issue to an agent, it opens a PR): GitHub’s Copilot coding agent, Linear’s agent integrations, Vibe Kanban, Jira/Rovo. Has anyone compared that with doing it in Fibery? Is there anything they do that Fibery structurally cannot?
- Fibery’s GitHub integration: is anyone using it to show PR and branch status next to Tasks? Is it good enough for this, or did you end up syncing via n8n/Make instead?
- Comments as a two-way channel between human and agent: has anyone done this with automations + webhooks? Any pitfalls with notification noise or automation loops?
- For the Fibery team: is agent-facing tooling (e.g. a native “assign to AI agent” concept, or webhooks on comment events that carry the full context) something on the roadmap?
Happy to share my schema and conventions if there is interest. I would rather learn from someone else’s mistakes before I build the whole thing.
Thanks,
Manuel