"Every new agent starts with a blank slate. You paste context. You explain what happened. You re-describe the task that was already described in three other places."
— DevCD VISION.md
"The next agent repeats failed attempts because it cannot see what already went wrong. Context is not portable, not structured, and not policy-governed."
— DevCD VISION.md
"AI agents should know what you are trying to continue without you having to tell them every time."
— Core Conviction, DevCD
Action Packet
Structured warm-start context for every new agent session. Goal, last blocker, do-not-repeat, and one next action—in a single policy-filtered payload.
Learn more →Local-First Privacy
All context stays on your machine. No remote export, no telemetry, no cloud sync by default. You decide what gets shared and with whom.
Policy docs →Policy Receipts
Every observation and every withholding is logged with a policy receipt. You always see exactly what was included in a handoff—and what stayed out.
Read more →MCP Server
Native Model Context Protocol server for direct agent consumption. Any MCP-capable agent reads your local DevCD context without you pasting anything.
Agent docs →Works With All Agents
One command configures Claude, GitHub Copilot, Codex, and OpenClaw at once. All your agents start warm from the same local context source.
Agent landscape →CLI + HTTP API
Full-featured CLI backed by a local HTTP API. Use it interactively, in scripts, or from any tool that speaks HTTP on localhost.
Architecture →pip install devcd
devcd setup --yes
devcd agentic action-packet
Local-first defaults, explicit policy receipts, and visible withheld-context boundaries on every handoff.
The next agent reads the Action Packet instead of asking you to re-explain your last three hours of work.
Goal, blocker, do-not-repeat, one next action, and context-load hints—nothing raw, nothing that shouldn't travel.
DevCD is not another agent to run. It is the layer that lets the next agent resume instead of restart.