coldteadotai/abide

What it solves

Coding agents often ignore project-specific rules (e.g., "don't create premature abstractions") because these rules are too nuanced for standard linters to catch. Abide ensures agents follow these instructions by automatically checking every edit and turn against a project's specific rubric, forcing the agent to repair violations immediately without human intervention.

How it works

Abide integrates into coding agents (like Claude Code, Codex, OpenCode, and Pi) via hooks or plugins. It compiles project instructions (from files like AGENTS.md or CLAUDE.md) into a machine-readable rubric. On every edit or turn, it sends the diff and the relevant rule to Jev (a specialized decision model by TypeSafe) which returns a calibrated probability of whether the rule was broken. If the probability exceeds a threshold, Abide notifies the agent to fix the error before proceeding.

Who it’s for

Developers using AI coding agents who want to enforce strict architectural, stylistic, or logic rules that cannot be scripted via traditional linting tools.

Highlights

• Low Latency & Cost: Checks take approximately 300ms and cost a fraction of a cent per turn.
• Agent Integration: Native support for Claude Code, Codex, OpenCode, and Pi.
• Automated Repair: Agents are told exactly which rule they broke and are required to fix it in the same turn.
• Audit Capabilities: Includes an audit command to check existing codebases against the project rubric.
• Privacy Focused: Data is sent directly to TypeSafe via the user's API key; Abide does not operate its own server in the data path.