Guard the agent's persisted approvals: scan Claude Code settings files for secrets saved inside allow rules, check git history, redact them, and install a pre-commit gate.
Free · Open source · On GitHub
Open-source dev tools
Small, inspectable tools you can run yourself — free on GitHub, with human deployment help when you need it.
- FREE
- OPEN-SOURCE
- INSPECTABLE CODE
- HUMAN HELP AVAILABLE
Featured tools
31 tools · all free · all on GitHub
Run one agent task in a fresh, disposable cell - scrubbed environment, network off by default, destroyed on exit - and sanitize tool/MCP/web output for instruction injection before it enters agent context.
Verify an agent's report against its own transcript: every claim gets a verdict (VERIFIED / UNVERIFIED / CONTRADICTED) with a transcript line pointer, so it works as a CI gate on real work.
Prove the instruction your agent is running is the instruction you approved — SHA-256 lockfile, pre-run drift diff, and reference-existence checks for AI agent skills.
Find which images and containers run a vulnerable Alpine + OpenSSH combination (CVE-2026-123158) — five-condition check, zero network, first-hour playbook.
Snapshot your dev-tool trust surface and diff it across updates: plugins, hooks, MCP servers, remote endpoints, git hooks, shell startup files — one snapshot, one diff, one exit code.
Verify-before-install for agent extensions. One command, one verdict (ALLOW, DENY, or REVIEW) before you install an extension into your editor — after a trojanized 'Claude Code' extension reached thousands of users before takedown.
Pico CSS was archived — v2.1.1 is the final release ever. picoexit audits your real Pico footprint, flags the selectors most likely to rot, and plans your exit: stay pinned, migrate (codemod to Tailwind v4 / Bulma v1 / plain CSS + Open Props), or follow a community fork.
Measure where your Rust build time goes, rank the cheapest fixes first, verify the delta. Non-invasive: reports only, never changes your toolchain.
Per-tool, per-model token & USD attribution from real agent session transcripts - measure where your AI bill burns, verify the cut paid off. 100% local.
Before you rent the GPUs, check the map: hardware inventory, memory math, and dated engine-support matrix to pick the right LLM serving engine.
The 2-second pre-merge gate for AI-assisted diffs: complexity, duplication and growth checks on changed code only. No model, no account, no network.
sessguard: local, read-only, zero-network CLI that inventories AI-assistant session and credential material on your machine and gives you a per-app revocation checklist. Never prints token values.
Measure your app across installed JS runtimes and get an upgrade verdict with reasons.
A local, deterministic, read-only behavioral detector for the Gheist planted-workflow attack shape: one command, one verdict per repository, one rotate checklist. No model, no account, no secret values ever printed.
A local, deterministic pre-execution gate for software an agent is about to install.
A deterministic diet for agent context: shrink logs, outputs, and repeated turns with an exact manifest. No model, no network, nothing invented.
Find the first real error in a failed CI log, tell flaky tests apart from real breakage, and cluster repeat failures. Zero dependencies.
Vendor-neutral policy, evidence and credential layer for AI agents: deterministic policy decisions outside the model process, a tamper-evident action ledger, and coverage accounting that fails loudly when an action leaves no record.
Open vendor-neutral DevSecOps operating layer: reference architecture, policy-as-code packs, CI security gates and audit-evidence automation for regulated organisations.
Audit .env files: drift vs .env.example, placeholders, risky secrets, unused vars. Zero dependencies.
Audit Agent Skills packs: duplicates, overlapping triggers, missing metadata. Zero dependencies.
Trim verbose logs and tool output to a token budget — keep the errors, drop the noise. Zero dependencies.
Catch tests that lie: fake passes, retry loops, swallowed assertions, deleted tests — Python + JS/TS, zero dependencies.
Scan a project for things that will break or leak: deprecated deps, EOL runtimes, unpinned CI actions, over-broad permissions, unpinned Docker bases, removed APIs. Zero dependencies.
Map where your API keys actually live — local, read-only auditor for agent-tool configs. Never prints key material, never sends anything anywhere.
Decision memory with provenance: record decisions, facts and constraints as a graph, trace why anything was decided, and catch contradictions and stale dependencies.
Will this GGUF model actually fit your GPU? Offline, zero-dependency VRAM check — weights + KV cache vs your card, in one command.
Approval leases for agent-run commands — approvals that expire by construction.
Check the numbers in AI-generated prose against their cited sources — zero-dependency Python CLI.
Shrink JSON for LLM prompts by folding long keys into short aliases — lossless, reversible, with a receipt.
Why these tools
Inspect everything
Every tool's code is on GitHub under MIT. Read it before you run it — no black boxes.
Run it yourself
Clone, run, done. Docs show the real commands and expected output — no accounts, no paywalls.
No lock-in
Tools run on your machine, and help deploys to your infrastructure and your accounts. You stay in control.
Human help when stuck
If you would rather not deploy it yourself, paid human deployment help is available — scoped plan and fixed quote in writing first.
