Connect Claude Code to Knownbase.

claude mcp add --transport http knownbase https://knownbase.dev/mcp

Step by step

1. Create a workspace. Sign up free — 3 projects, 300 notes, 2 MCP keys, no card required.
2. Add the server. Run the command above in your project directory.
3. Verify. Run /mcp in Claude Code; knownbase should appear with its tools listed.
4. Authorize. Ask Claude Code to do something that needs memory (“search project memory for anything about deployments”). The first call opens a browser for sign-in and workspace selection.
5. Teach it the habit. Add the CLAUDE.md section below.

Using an API key instead

If you're in CI, on a machine with no browser, or on a client build without OAuth support, generate a key in Settings → Workspace & MCP (shown in full only once) and configure it directly:

{
  "mcpServers": {
    "knownbase": {
      "url": "https://knownbase.dev/mcp",
      "headers": {
        "Authorization": "Bearer kb_YOUR_KEY_HERE"
      }
    }
  }
}

Tell Claude Code when to use it

This is the step people skip, and it's the one that decides whether the integration is useful. Paste into your project's CLAUDE.md:

## Project memory

This project has persistent memory via the `knownbase` MCP server,
under project "<your-project>".

Before starting work on an unfamiliar area, `search_notes` for prior
decisions, constraints and debugging findings about it.

When you finish something durable, `upsert_note` it:
- an architecture decision, with the reasoning and the alternatives
- a non-obvious root cause (the explanation, not the patch)
- an environment or deployment constraint
- an approach that was tried and rejected, and why
- a handoff, if you're stopping mid-task

Code lives in git. Store the reasoning here. Don't store transcripts,
generated output, or anything unconfirmed.

What the agent can do once connected

• search_notes — find notes by query, project, tag, status or last-modified date. Returns snippets and metadata rather than full bodies, so searching is cheap; pass includeBody:true when you want everything inline.
• get_note / get_notes — read one note in full, or up to 50 by id in one call, with resolved [[links]] and backlinks.
• upsert_note — create or update. Updates are partial: omitted fields are left alone, so fixing a tag can't blank the body.
• list_projects — every project with note counts by status.
• list_note_revisions / get_note_revision — how a note changed, and what it used to say.
• delete_note — soft delete; restorable from the dashboard trash.
• backup_project — full content for one project, returned inline.

Every call is scoped to the workspace the credential belongs to. A key can additionally be marked read-only, or restricted to a single project — in which case notes in other projects are invisible to it, not merely refused.

Troubleshooting

• Server shows as failed in /mcp. Usually an older client without remote HTTP MCP support. Use the API-key configuration above.
• Authorization never opens. Make a request that actually requires a tool call — OAuth is triggered lazily on first use, not at add time.
• Writes are refused. The key is read-only. Check it in Settings → Workspace & MCP.
• Notes seem missing. The key may be scoped to one project. A project-scoped key sees only that project, by design.

Related

• Why Claude Code forgets, and what persists today
• Making Claude Code remember across sessions
• Connect Codex to the same workspace
• Full MCP tool reference

FAQ