claudex turns each developer's claude code memory, hooks, mcp servers, and skills into a governed, shared organizational asset — so productivity compounds at the team level, not the seat.
claude code, cursor and copilot deliver 25–55% individual task speed-ups. but at the team level, gains decay to 15–25% without shared knowledge processes.
the leakage lives in private ~/.claude directories — memory, hooks,
mcp servers, agent capabilities, hard-won skills — none of it visible to teammates,
none of it governed, none of it durable across attrition.
mckinsey state of ai 2024 · github research (peng et al.) · atlassian state of dev experience 2024
claudex is a local-first rust cli + a multi-tenant control plane.
every feature ships behind the same trust boundary: memory contents never leave
a developer's laptop until they explicitly run claudex propose.
useful local entries are converted into a codeowners-routed draft pr against your repo's
.claude/ tree — with a capability diff in the body. nothing auto-merges.
nothing auto-writes.
inventory the executable surface across the org — hooks in .claude/settings.json,
mcp servers in .mcp.json, agents in .claude/agents/.
drift detection. one-click revoke.
.claude/skills/ is the source of truth for both claude code (developer) and
the agent sdk (server). edit once. both surfaces see it.
four postgres schemas — core, tenant, audit, analytics. rls everywhere. pgcrypto + supabase vault for at-rest encryption. enterprise-ready by construction, not by patch.
rust heuristics handle ~80% of entries with no network call.
the remaining 20% can route to a haiku-class agent — opt-in via --cloud.
personal entries never reach the cloud, by design.
a read-only managed mcp server distributes governed org knowledge back into every developer's claude code session — never writes locally, never leaks across tenants.
a four-step loop. it runs continuously on the developer's laptop, opportunistically in the cloud, and durably across the org.
a local rust daemon watches ~/.claude/projects/<dir>/memory/ and picks up new
entries the moment they're written. zero config. zero network. opt-out per project.
heuristics first — four classes (personal, team, project, reference). entries flagged personal never propose, never leave the laptop, never reach the cloud. anything ambiguous can be routed to a haiku-class cloud agent on explicit opt-in.
~80% local accuracy · <$0.005/entry cloud costthe control plane builds a live capability inventory across the org — every hook, mcp server, and agent capability is visible, diff-able, and revocable. drift against policy is surfaced before it lands in production.
rls-isolated · append-only audit · soc 2 type ii (in progress)
approved entries flow back into the team via three surfaces: (a) propose-prs into the
repo's .claude/ tree, (b) skills sync visible to claude code & the agent sdk,
(c) the read-only managed mcp server. the loop closes.
25–250 engineers · 3–10 active repos · series b–d
"our memory is fragmented. each engineer's claude knows a different version of how we ship."
250–2,000 engineers · fintech · health · gov-tech
"the eu ai act + soc 2 require capability inventory for every ai tool we let touch the codebase."
2,000+ engineers · mature idp · multi-team
"we standardize 40 teams. ai tooling is the only surface we haven't put under the platform."
we charge for the surface we govern — your repos — not for the people typing. a 50-engineer team buys 3 projects ($99/mo), not 50 seats. inspired by supabase, vercel, render.
individuals · oss · evaluators
ai-native scale-ups
mid-market eng orgs
regulated · platform-eng
per-seat ai tools (copilot, cursor) read as "monitor my developers" to procurement and to the developers themselves. claudex meters on the surface we govern — repos, not heads. a 50-engineer team buys 3 projects ($99/mo), not 50 seats. it lands faster and feels less surveillance-y. fair-use contributor caps prevent monorepo gaming.
not without an explicit claudex propose or --cloud flag. the rust cli
handles ~80% of classification locally. anything labeled personal never proposes
and never reaches the cloud — even via override. this is enforced in the schema, not just policy.
claude code is first-class; cursor and github copilot enterprise are in compatibility preview. mcp servers and the claude agent sdk are governed natively. multi-vendor capability ingestion for cursor + copilot is on the 2027 roadmap.
yes — on the enterprise tier. a kubernetes helm chart ships the control plane, postgres, and the managed mcp server in your own cloud. air-gapped license available for fintech, health, and gov-tech. fedramp moderate path is on the q1 2027 roadmap.
when you run claudex propose, the cli (1) re-classifies any new entries,
(2) drops anything personal, (3) opens a draft pr against your repo's
.claude/ tree via the github app, (4) routes it via codeowners,
(5) writes a capability diff into the pr body so reviewers see exactly what hooks /
mcp servers / agent capabilities are being added or changed. nothing auto-merges.
soc 2 type ii is in progress (target q4 2026). the platform was designed against common criteria 2.0's ai-assisted-access controls from day one. eu ai act article 9 (high-risk) and article 50 (transparency) reporting is built into the audit-log export. iso 27001 is on the 2027 roadmap. enterprise customers receive the soc 2 report under nda.
cody is search-first; glean searches existing docs; so4t is a wiki. none of them are ai-tool-aware — they don't know what a hook, mcp server, skill, or agent capability is. claudex sits in the gap: an ai-tool-native, provider-agnostic governance plane that creates governed knowledge from the artifacts your team is already producing inside claude code.
install the cli with brew install claudex, run claudex init in any
repo, and you're in the hobby tier — free forever. for teams: sign up at the start-free
link below; you'll be on pro within five minutes, no credit card for the first 30 days.
five minutes to install. zero credit card to start. zero memory leaves your laptop until you say so.
$ brew install claudex && claudex init