selvedge

What is Selvedge?

Selvedge is a local MCP server that gives AI coding agents (Claude Code, Cursor, Copilot, Cline, Windsurf, Continue) long-term memory. As an agent works, it calls Selvedge's log_change tool to record what it changed and why — captured live, in the same context window that produced the change, not inferred afterward from a diff by a second LLM. Everything is stored in a plain SQLite file under .selvedge/ next to your code, with zero external dependencies beyond MCP, Click, and Rich — no accounts, no servers, no telemetry.

Why it matters

git blame tells you what changed and when, but human-written intent — commit messages, PR descriptions, Slack threads — simply doesn't exist for most AI-generated changes. Six months later, a column like user_tier_v2 shows up in git blame with a generic message like “Update schema,” and the reasoning is gone. Selvedge closes that gap by capturing the agent's own reasoning at the moment of the change.

Key features

• Entity-level tracking: attribution is tied to things you actually search for — users.email, env/STRIPE_SECRET_KEY, api/v1/checkout, deps/stripe — not just line ranges.
• Changesets: tag related events across many entities under one feature or task slug, so a multi-PR rollout can be pulled back as a single scope.
• Active memory: the prior_attempts tool lets an agent check what was already tried and reverted before editing an entity again, and stale_decisions surfaces dated decisions (via revisit_after) that have come due — but only for entities still in active use.
• Agent Trace export: selvedge export --format agent-trace emits Agent Trace v0.1.0 records, the open AI code-attribution wire format from Cursor and Cognition AI, so captured history can travel into other tools that read the standard.
• CLI and MCP parity: query history with commands like selvedge blame, selvedge diff, selvedge history --since 30d, selvedge changeset, and selvedge search.

Who it's for

Teams running long-term, AI-coded codebases that need an audit trail for schema decisions, migrations, and dependency changes that survives after the original agent session and prompt are gone.

Solo developers using Claude Code on side projects and internal tools who just want to remember why they (or their agent) made a change last week, without any enterprise governance overhead.

Getting started

Install with pip install selvedge, then run selvedge setup inside your project. The interactive wizard detects which AI tools are installed, wires the MCP server into each one's config, adds the agent-instructions block to your project's prompt file, and installs a post-commit hook — re-running it is a no-op.