PromptWise is the governance & intelligence layer for AI coding agents — model routing, cost control, security scanning, and a hash-chained audit trail — compiled down to MCP, SKILL.md, and AGENTS.md: the formats every agent already reads. Runs the same for a solo developer, a small team, or a regulated enterprise.
No tier, no paywall, no separate "enterprise edition" — the same MIT-licensed core scales from one laptop to a regulated org, just by how many people point their agent at it.
One pip install -e ., one MCP entry, and it's live inside Claude Code, Cursor,
or whatever you already run. No new UI to learn, no team to onboard — every check fails
open, never blocks your flow.
/promptwiseAdvisory quality gates, policy-as-code, and a governed agile method (PRD → stories → TDD → review) run automatically — PASS / CONCERNS / FAIL with a reason attached, not a silent checkbox nobody reads.
A hash-chained audit trail that breaks visibly on tampering, mapped to GDPR, HIPAA, and OWASP controls, exported as a signed, self-contained bundle — no vendor dashboard, no seat license, data never leaves your infrastructure.
Not a settings page. Each capability writes into the same hash-chained trail, in this order, every time a hook fires.
Haiku, Sonnet, or Opus, picked by task type and budget — not by habit. Reasoning effort (low / medium / high) is routed independently, with an outcome-learning adapter that fails open to the static pick when it doesn't yet have enough evidence.
Secrets, prompt injection, PII, and OWASP patterns — scanned air-gapped by default, with a red-team regression harness diffed against a stored baseline so fixes don't quietly reopen old holes.
An advisory quality gate and policy-as-code check run before a change merges — PASS / CONCERNS / FAIL, with a reason attached, not a silent pass.
Standing exceptions expire on their own — 60 minutes by default, 8 hours max — and decay back to the normal permission prompt without anyone remembering to revoke them.
Every governed action links to the one before it. A tampered or missing entry breaks
the chain visibly — export and verify it any time with promptwise:audit.
A correction becomes a durable, searchable rule, replayed before similar work — local first, air-gapped safe, no round-trip to a vendor to remember what you told it once.
The properties a security review or a procurement checklist actually asks about — not marketing claims.
A broken check never blocks your work — advisory by default, hard-deny only for destructive commands.
Scanning and routing run locally — no unconditional network call, no data leaving your machine.
Hash-chained records, Ed25519-signed exports — a missing or altered entry is visible, not silent.
A single MCP server — no per-feature plugins to track, update, or lose track of.
Write your policy once — it compiles to CLAUDE.md, AGENTS.md, and every native rules format.
No seat license, no vendor lock-in, no dashboard you don't control. It's your repo, your data.
The governance core writes its rules once and compiles them down to every format an agent already reads — no proprietary config, no separate plugin per host.
Not a marketing simplification — the real handler categories the tools are organized into.
One governance core, compiled to whatever your agent already reads.
A gap analysis, not a certification — every mapped control names the real tool that evidences it; an unevidenced control is reported absent, never guessed.
/mcp# clone and install git clone https://github.com/u2anilgit/PromptWise.git cd PromptWise pip install -e . # restart your agent, then in Claude Code /mcp /promptwise