unerr

Stop babysitting your AI.

unerr is operational intelligence for your codebase — one local runtime, behind every MCP your agent speaks, that catches the refactor about to break 7 call sites, remembers what the team already decided, and keeps context lean. Joins no single-purpose memory or graph tool can make, because everything lives in one process.

It runs entirely on your machine as an MCP server: it builds a graph of your codebase and remembers what the codebase has been through — who changed each file last, why it drifted, what failed here before, and the rules you've fed it across sessions. Measured, not estimated: it removes 86–90% of the tokens an agent spends navigating code, with a fidelity gate that discards any "saving" that lost the answer.

The outcome is agents that behave like senior engineers — checking dependencies before editing, remembering project history, and refusing to thrash on a function they've already failed on.

Start here

• CLI Reference — every unerr command, generated directly from the binary.

What lives under the hood

These subsystems do most of the work and are documented as they mature:

• MCP tools — the graph-backed tools the agent calls (search_code, get_references, file_read, …) instead of grep/glob.
• MCP router — intent classification that routes a request to the right family of tools.
• Behavior automation — cascade guard, boundary checks, loop breaker, and incomplete-work detection that fire on edits.
• Reviews — a graph-aware code reviewer, distinct from a linter.
• File & graph intelligence — the CozoDB graph, AST extraction, and drift tracking.
• Shell-output compression — rolling long command output into compact, LLM-actionable summaries.
• Memory — anchored notes and persistent facts that survive across sessions.
• Cascade graph — the call/reference edges that power blast-radius signals.