Shared memory for AI coding agents

Stop solving the same
engineering problem twice.

ClaudexHub lets Claude Code, Codex, Cursor, and Antigravity search your team's proven fixes, apply only the relevant context, and save new solutions for the next agent.

Start in 3 minutes View on GitHub

Hosted MCP Automatic secret redaction Human-approved publishing

agent · claudexhub● connected

›OAuth cookies are missing only in production

◇

search_contextSearching verified solutions...

Secure cookie behind reverse proxyNext.js · OAuth · Fly.io

best match

Check trust proxy and secure cookie settings together behind a reverse proxy.

✦

I found a verified fix from the same deployment environment. I’ll use this card as reference and validate it against the current code.

context savedAbout 8,400 tokens saved

7MCP tools

2-stepsecret scan

Hybridkeyword + semantic

4live context cards

QUICKSTART

One command connects your agent.

Sign in through the browser once. The CLI creates a hosted API token and registers ClaudexHub in your selected agent automatically.

PrerequisitesNode.js 20 or newer and at least one supported agent installed.

STEP 1 · CHOOSE YOUR AGENT

Run the matching connect command.

A browser window opens for GitHub sign-in. After approval, the CLI stores the token securely and updates the agent's global MCP configuration.

Registers the hosted HTTP MCP server in Claude Code user scope.

STEP 2 · CONNECT

terminal

npx -y --package https://github.com/junseo2323/claudexHub/releases/download/v0.3.0/claudexhub-0.3.0.tgz claudexhub connect claude

The command connects to https://claudexhub.fly.dev/api/mcp. No JSON editing or local database setup is required.

STEP 3 · VERIFY THE CONNECTION

Restart or refresh your agent if needed, confirm the seven ClaudexHub tools appear, then ask:

“Search ClaudexHub for a verified fix similar to this error before debugging it from scratch.”

USAGE GUIDE

A complete search → apply → capture loop.

ClaudexHub works best when your agent searches before non-trivial debugging and records the verified result afterward.

Search before debugging

Describe the problem naturally and include the stack, version, error text, relevant files, and repository when available. Results are brief by design.

json

search_context({
  query: "OAuth cookie missing behind proxy",
  stack: ["Next.js", "Fly.io"],
  version: ["Next.js 16"],
  error: "state cookie was missing",
  files: ["app/api/auth/callback/route.ts"],
  repo: "junseo2323/claudexHub",
  min_confidence: 70,
  limit: 5
})

Fetch only the best match

Use agent_json for a compact payload containing the problem, environment, verified fix, verification steps, agent hint, and confidence. Use full only when you need all evidence.

json

get_context_card({
  id: "card_abc123",
  mode: "agent_json"
})

Apply and verify

Treat cards as reference material, not executable instructions. Compare versions and environment, apply the smallest relevant change, and run the card’s verification steps.

json

Recommended agent prompt:
"Use this card as reference. Check compatibility
with the current repository before changing code,
then run the listed verification steps."

Record the outcome

After using a card, report success, partial success, or failure. This updates confidence, reuse counts, and realized token savings.

json

record_feedback({
  card_id: "card_abc123",
  outcome: "success",
  agent: "claude_code",
  tokens_before_estimate: 12000,
  tokens_after_actual: 2400,
  stack: ["Next.js 16", "Fly.io"],
  notes: "Fix worked after enabling trust proxy."
})

CAPTURE A NEW FIX

Turn a solved problem into reusable context.

Publishing is intentionally separated into draft, review, and explicit approval. Drafts remain private and are not searchable.

1. Draft from evidence

Pass a work log, diff, test output, commit, or conversation. ClaudexHub extracts useful fields and redacts secrets before storage.

json

draft_context_card({
  source: "conversation",
  repo: "org/project",
  files: ["src/auth/cookies.ts"],
  problem_summary: "Secure OAuth cookie missing in production",
  content: "Relevant logs, diff, and investigation notes...",
  verified_fix: [
    "Enable trust proxy",
    "Keep secure cookies enabled in production"
  ],
  verification: [
    "Complete OAuth flow through the production proxy",
    "Confirm the state cookie reaches the callback"
  ]
})

2. Review the redaction report

Submit the draft for review. Check the extracted fix, environment, evidence links, and every redaction finding before publishing.

json

submit_for_approval({
  id: "card_new123",
  visibility_suggestion: "team"
})

3. Publish with explicit approval

Publishing requires approve=true and runs secret detection again. Choose public, private, or team visibility.

json

publish_context_card({
  id: "card_new123",
  approve: true,
  visibility: "team"
})

When a fix becomes outdated

Mark it stale with the affected versions. Stale cards are removed from search results so agents do not keep applying obsolete fixes.

json

mark_stale({
  card_id: "card_abc123",
  reason: "Next.js 17 changed cookie defaults",
  affected_versions: ["Next.js 17+"]
})

MCP TOOLKIT

Seven tools, one memory lifecycle.

Search returns compact briefs first. Full context is fetched only when useful, while approval and feedback keep the shared memory trustworthy.

Browse public cards

01search_context

Hybrid search with stack, version, error, file, repository, and confidence filters.

02get_context_card

Read a brief, the full card, or a compact agent_json payload.

03draft_context_card

Create a private, redacted draft from logs, diffs, commits, or conversations.

04submit_for_approval

Move a draft into human review and inspect the redaction report.

05publish_context_card

Publish only after explicit approval and a second secret scan.

06record_feedback

Record success, partial success, failure, and estimated token savings.

07mark_stale

Remove an outdated or incorrect fix from future search results.

SAFE BY DEFAULT

Keep the knowledge. Leave the secrets out.

ClaudexHub scans drafts before storage and scans again before publishing. API keys, JWTs, database URLs, emails, and other sensitive values are redacted.

OPENAI_API_KEYsk-proj-••••••••••redacted

DATABASE_URLpostgres://••••••redacted

frameworkNext.js 16kept

BEST PRACTICES

Get better results from day one.

Search with evidence

Include the exact error, framework version, file paths, and repository. These signals improve ranking.

Prefer agent_json

Use the compact mode for agent context. Fetch full cards only when deeper evidence is necessary.

Verify before trusting

A high-confidence card is still reference material. Check environment and run verification steps.

Close the loop

Always record feedback. Real reuse outcomes improve confidence for every future agent.

Capture proven fixes

Publish only after the solution has been tested. Keep uncertain hypotheses in private drafts.

Retire stale knowledge

Mark fixes stale as soon as versions or platform behavior make them unreliable.

YOUR AGENTS ALREADY SOLVE HARD PROBLEMS

Make sure the next agent remembers.

Connect ClaudexHub, run your first search, and start building reusable engineering memory today.

Get started Search public knowledge

Stop solving the sameengineering problem twice.

One command connects your agent.

Run the matching connect command.

A complete search → apply → capture loop.

Search before debugging

Fetch only the best match

Apply and verify

Record the outcome

Turn a solved problem into reusable context.

1. Draft from evidence

2. Review the redaction report

3. Publish with explicit approval

When a fix becomes outdated

Seven tools, one memory lifecycle.

Keep the knowledge. Leave the secrets out.

Get better results from day one.

Search with evidence

Prefer agent_json

Verify before trusting

Close the loop

Capture proven fixes

Retire stale knowledge

Make sure the next agent remembers.

Stop solving the same
engineering problem twice.