ai-journal-mcp

A local MCP server for journaling, organizing, recalling, and tracking your work. Capture what each session taught you, let it organize into a clean, queryable structure, then recall and analyze the whole archive — recurring patterns, past lessons, blog-post material. The tasks that fall out of the work live here too, linked to the entries that explain them. Plain markdown stays the source of truth, and nothing leaves your machine.

The problem

You journal the hard-won lessons — the debugging pattern, the process failure, the "this is a blog post" moment. Then they vanish. Not because you didn't write them down, but because a journal you can't interrogate is write-only memory: the insight is in there somewhere, in a file too big to reread, and the pattern recurs anyway because nothing surfaced it at the moment you needed it.

Naming a pattern in your journal doesn't prevent the next instance — but being able to recall it does. That recall is the whole point, and it's what plain markdown files alone can't give you.

What it feels like

Ask your journal a real question, mid-work, from the same LLM session you're already in:

You: "What do I keep relearning about AI-assisted development?"

Claude (via ai-journal-mcp): searches across months of entries, pulls the seven that recur on the theme, and synthesizes the through-line — "You've hit 'tests are an uneven safety net' three times since April; here are the entries and the common trigger…"

Or capture without breaking flow:

You: "Journal that — the bit about auditing every artifact when a hypothesis dies, not just the one with a test."

ai-journal-mcp writes entries/2026-04/30-when-a-hypothesis-dies.md with themes and blog angles, regenerates the index views, and rebuilds search — one call, everything consistent.

What it does

Four capabilities, one local MCP server:

1. Journal — capture however suits the moment: dump the whole session, jot a single lesson, or hand it a rough list to clean up. The add_entry tool takes freeform text and files it as a canonical entry (one per file, entries/YYYY-MM/DD-slug.md, with themes, tags, and blog_angles) — no format discipline required. Prefer to write entries by hand? ai-journal-mcp reads what's already there as-is.
2. Organize — themes are metadata, not folders, so one entry can carry several and no themed file grows without bound. The index and per-theme views are generated, never hand-edited, so the structure can't rot back into a megafile. Bringing a mess? scan reports what a migration would do; migrate --apply rewrites a sprawling journal into the clean layout — originals preserved in attic/, every dedup decision logged, no data loss, ever. (The first run absorbed 1,460 entries across 340 files spanning three format eras.)
3. Recall & analyze — full-text + structured search across one or many journals (search_journal, entries_over_time, list_themes, get_entry), filtered by theme, journal, or date range. Surface recurring patterns, find unused blog material, trace when a problem first appeared. This is the payoff: the journal as raw material for posts, talks, and not repeating old mistakes.
4. Track — the tasks that come out of the work, as a mutable list: status, priority, and what's blocked on what (add_task, update_task, list_tasks). Each task links to the entries that give it context, so resuming one surfaces the reasoning behind it. Tasks are journaling's mutable sibling — separate files, their own rules, never bending the append-only entries.

Your data stays yours

• Markdown is the source of truth. The SQLite + FTS5 index is disposable — delete it anytime, it rebuilds from your files. Nothing is locked in a database you don't control.
• It runs locally. An MCP stdio server; your journal never leaves your machine.
• It doesn't demand ownership of every source. Register a journal as indexed and ai-journal-mcp reads and searches it in place but never rewrites it — ideal for a journal that already has its own conventions. managed journals are the ones it maintains for you. Both are searchable together, so cross-domain patterns ("what was I learning in engineering the week I learned X in deal research?") stop being invisible.

Quickstart

Install from PyPI (Python 3.11+):

pip install "ai-journal-mcp[server]"   # MCP server + CLI; drop [server] for CLI-only

Register your journals in ~/.config/ai-journal-mcp/journals.toml:

[[journal]]
name = "technical"
path = "~/journal"
mode = "managed"            # ai-journal-mcp owns the layout

[[journal]]
name = "deal-research"
path = "~/research/deals"
mode = "indexed"            # read-only; searched but never rewritten

Wire it into Claude Code as an MCP server:

claude mcp add ai-journal-mcp -- ai-journal-mcp serve

Querying happens through the server (it builds and refreshes the index for you). The CLI handles intake and maintenance directly:

ai-journal-mcp scan ~/old-journal            # dry-run intake report
ai-journal-mcp migrate ~/old-journal --apply # rewrite into the managed layout
ai-journal-mcp refresh ~/journal             # regenerate JOURNAL.md + theme views

Documentation

Doc	Contents
docs/USE_CASES.md	What the product is for, case by case
docs/ARCHITECTURE.md	Components, data flow, journal modes, trust boundaries
docs/SPECIFICATION.md	Entry format, journals.toml, parser rules, tool/CLI contracts
docs/ARCHITECTURE_DECISIONS.md	The "why" behind each design choice
docs/DEVELOPMENT.md	Dev environment setup, make targets, tooling, troubleshooting

Development

./scripts/setup.sh   # Python 3.11+: creates .venv, installs the package + dev tools
make check           # lint + format-check + type-check + tests (the pre-push gate)

See docs/DEVELOPMENT.md for the full guide.

License

MIT © Solent Labs™