Credence — Epistemic Guard
Run an epistemic guard in the agent loop so uncertain facts are tracked and file writes stay blocked until you confirm them.
Overview
Credence is a MCP server for the Ship phase that tracks uncertain agent-proposed values and blocks writes until they are verified.
What is this MCP server?
- Tracks uncertain values the model proposes during coding sessions
- Blocks write operations until values are explicitly verified
- PyPI package credence-guard v1.2.5 with stdio MCP transport
- Epistemic guard pattern aimed at AI-assisted editing workflows
- Open-source Credence project on GitHub (credence-ai)
- Published MCP server version 1.2.5
- Package identifier credence-guard on PyPI
- Transport type stdio
What problem does it solve?
Coding agents often commit plausible-but-wrong literals, and nothing in the default tool loop forces you to confirm them before write.
Who is it for?
Solo builders using high-autonomy agents on sensitive repos who want a verification gate before filesystem changes land.
Skip if: Pure read-only research chats, teams that only use completion without write tools, or workflows that already block all writes manually.
What do I get? / Deliverables
Writes stay gated until uncertain values are resolved, reducing bad config and false-confidence merges from agent sessions.
- Active epistemic guard in the MCP tool chain
- Blocked writes for unverified uncertain values
- Clearer audit trail of what the agent was unsure about before save
Recommended MCP Servers
Journey fit
How it compares
Epistemic write-guard MCP server, not a linter, test runner, or secret scanner alone.
Common Questions / FAQ
Who is Credence for?
Developers using Claude Code, Cursor, or similar agents who need a guardrail when the model edits files that must reflect verified facts.
When should I use Credence?
Use it in Ship review before merging agent-generated changes, especially for env vars, URLs, IDs, and security-sensitive configuration.
How do I add Credence to my agent?
Install credence-guard from PyPI, add the stdio MCP server to your agent configuration, and keep it enabled during sessions where the agent has write access.