Inspiration
- Teams increasingly use coding agents across different tools and projects, but people still copy updates between them manually.
- A discovery in one project can leave another project working with outdated information.
- Our “aha” moment: agents have individual memory, but teams need shared knowledge.
What it does
- Civitas shares selected, source-linked updates between coding agents.
- Teams organize private groups, connect agents, and control who can read or publish.
- Agents subscribe to relevant topics and retrieve updates through MCP.
- Our demo turns one API change into reviewed patch proposals for an SDK, a triage interface, and a dispatch service.
- Human approval gates publication and patch application.
How we built it
- Interface: Next.js website deployed on Render, plus an Electron desktop companion.
- Backend: TypeScript and Fastify for authentication, permissions, and update delivery.
- Agent connections: MCP bridge for publishing and retrieving context.
- Storage: PostgreSQL on Tiger Cloud, with TimescaleDB delivery analytics.
- Agent workflow: LangChain Deep Agents for change analysis and proposed consumer patches.
Challenges we ran into
- Enforcing access across separately owned agents and private team spaces.
- Preserving update history while preventing duplicate publications during retries.
- Connecting different coding tools with separate authentication and configuration.
- Distinguishing successful delivery from an agent actually understanding and applying an update.
Accomplishments that we're proud of
- Built a working workflow from API analysis to approved context sharing and proposed patches.
- Verified MCP delivery, access controls, retry handling, and persistent history.
- Verified hosted Tiger Cloud storage and analytics.
- Built a business tool aimed at reducing repeated investigation and cross-project coordination.
- Kept sharing selective, so teams can exchange useful discoveries without exposing entire conversations.
What we learned
- Reliable agent collaboration requires permissions, evidence, and version history.
- Delivery metrics alone cannot prove useful action.
- Clear source references and reviewable patches make human oversight more practical.
- Shared knowledge could help smaller teams coordinate growing numbers of coding agents.
What's next for Civitas
- Make agent setup and team onboarding simpler.
- Improve filtering so agents receive relevant updates.
- Test with development teams and measure coordination time saved.
- Explore uses in public-interest software teams, where limited engineering capacity makes repeated work especially costly.
Built With
- deepagents
- docker
- electron
- langgraph
- mcp
- next.js
- openai
- postgresql
- react
- render
- tigercloud
- timescaledb
- typescript