Every time you switch from Claude Code to Codex, you start from zero. The task list, the architectural decisions, the governance rules — gone. You spend 20 minutes re-explaining what you were doing. That's hours of productivity lost every week.
It doesn't have to be this way. Delimit stores your project context on disk, not in any model's context window. When you switch models, everything follows.
The problem
AI coding assistants are stateless across each other. Claude Code doesn't know what you did in Codex. Gemini CLI doesn't know what Claude decided. Each model is an island.
You switch models for good reasons — rate limits, model strengths, cost. But every switch resets your context to zero.
What Delimit shares across models
- Task ledger — open items, priorities, blockers, completion notes. Created in Claude, visible in Codex.
- API policies — which endpoints are frozen, what severity levels apply, custom rules. Enforced the same way regardless of which model is running.
- Persistent memory — project decisions, architecture notes, team preferences. Written once, read everywhere.
- Deliberation history — past multi-model debates and consensus decisions. Any model can reference what was decided and why.
How it works
Delimit uses MCP (Model Context Protocol) to expose tools to every AI assistant. The tools read and write to ~/.delimit/ on your local machine. No cloud sync, no API keys, no extra cost.
# Install once, works everywhere:
npx delimit-cli setup
# Detects your AI assistants and configures MCP:
Found: Claude Code, Codex CLI, Gemini CLI
Configured: 3/3 assistants
Tools available: 106
Real workflow
# In Claude Code:
> "Add LED-089: migrate auth to JWT tokens"
Created LED-089: migrate auth to JWT tokens [open]
# Hit rate limit. Switch to Codex:
> "What's on the ledger?"
LED-089: migrate auth to JWT tokens [open]
# Need a different perspective. Open Gemini CLI:
> "Review LED-089 and check if the auth endpoint policy allows this change"
LED-089 reviewed. Policy check: /auth/* frozen in v1. Migration requires v2 path.
Three models, one workspace. No re-explaining. No lost context.
The merge gate for AI-written code, signed and replayable. Switch models, not context. Install Delimit. Keep building.