Most tooling around coding agents helps them do more. These mostly help you find out what they already did: what context they needed, what they changed, what they broke, and what they quietly got wrong.
Every one ships a NOVELTY.md: the prior-art search run before any code, including a table
of claims the project is not allowed to make. Several were renamed or cut down by that
gate. It is the most useful file in each repo.
| rollcall | Which worktree does each running agent own, and which work does nobody own? Joins the OS process table, provider session data and git. Anthropic's docs say to remove leftover worktrees by hand; this finds them. |
| sincelast | Your agent read that file four minutes ago. Another agent has changed it since. Read-dependency invalidation for independently-launched sessions. |
| claude-memory-atlas | Renders a directory of Claude Code auto-memory files as one offline interactive knowledge graph. |
| stopless | Which words does your prompt actually need? Learned stopwords, measured per model. |
| stopbench | How much language does an LLM need? A contributable linguistic-ablation benchmark. |
| mincontext | Delta debugging for AI failures. Shrink a failing agent context to the part that actually causes it. |
| toolsweep | Which tool-schema decision is costing you accuracy? Controlled ablation over function definitions. |
| errorbars | The statistics your LLM eval does not do. Power analysis, paired bootstrap, minimum detectable effect. |
| inconclusive | Your agent test says FAIL. Did the property fail, or did the trace just end? Three-valued verdicts. |
| knowwhen | Every agent benchmark asks whether it finished. This one asks whether it should have started. |
| assumptionledger | Make an agent state what it is assuming, in a form you can execute against reality. |
| validwhile | Evidence-bound agent memory. Every fact carries an executable check for whether it is still true. |
| worker-king | Always-on AI desktop assistant for Windows. Listens, and delegates real work to Claude Code. |
| local-transcription | Fully offline Windows transcription with speaker memory, and an optional live meeting AI. No cloud, ever. |
| claude-roundtable | Two Claude personas argue a topic adversarially, streamed live in your terminal. |
Everything here is Apache-2.0 or MIT, local-first, and runs without an account.