june-mcp

Give your agent a memory. june-mcp is the official MCP server for Junê — it connects any MCP host (Claude Desktop, Claude Code, and friends) to a June knowledge graph, so your agent can ask, search, and remember against a shared, cited, tenant-isolated memory.

This package is a thin, zero-logic connector: all retrieval, graph assembly, and answering happen on the June endpoint you point it at. No engine code lives here — which is why it's small enough to read in one sitting.

Claude Desktop / Claude Code  ──stdio──▶  june-mcp  ──HTTPS──▶  your June endpoint
                                                                 (graph · retrieval · answers)

Install

pip install june-mcp          # just the connector   (or: pipx install june-mcp)
pip install june-ai           # umbrella: june-mcp + june-bench (the benchmark suite)
pip install "june-bench[mcp]" # the bench, with the connector as an extra

Point it at a June endpoint

june-mcp speaks to any June service. Three ways to have one:

1. Junê desktop app (local-first). Run the Junê app and connect to its local engine — your files, graph, and keys stay on your machine.
2. Your own June service. Pro/Team customers running the june-local engine package point JUNE_BASE_URL at their own server.
3. Hosted (Team). Point at your hosted June workspace endpoint with the API key from your console.

Configure

The server is fail-closed: it refuses to start unless it knows where to connect and as whom, and tells you everything that's missing in one message (not one error at a time).

Check it before your agent does

JUNE_BASE_URL=http://localhost:8000 JUNE_API_KEY=... JUNE_CANVAS=work june-mcp --doctor

The doctor verifies, in order: config → service reachable → canvas resolution (your canvas name → its id, e.g. name "work" → 9147bee6-…) → search seam healthy → tool manifest, and prints PASS/FAIL per check with a mapped hint (e.g. a missing name lists the canvases that DO exist and points at JUNE_CANVAS_CREATE=1). The doctor exits 0 only when every check passes (1 otherwise); the server itself exits 2 on a config error instead of starting half-wired. Run the doctor first; it catches every common misconfiguration before your agent ever sees the server.

Wire it into Claude

Claude Desktop — merge into claude_desktop_config.json (Settings → Developer):

{
  "mcpServers": {
    "june": {
      "command": "june-mcp",
      "env": {
        "JUNE_BASE_URL": "http://localhost:8000",
        "JUNE_API_KEY": "your-key",
        "JUNE_CANVAS": "work",
        "JUNE_LLM_KEY": "your-llm-provider-key"
      }
    }
  }
}

Claude Code:

claude mcp add june -e JUNE_BASE_URL=http://localhost:8000 \
  -e JUNE_API_KEY=your-key -e JUNE_CANVAS=work \
  -e JUNE_LLM_KEY=your-llm-provider-key -- june-mcp

Fully restart the host (Cmd+Q on macOS), then check the server shows 20 tools — the compact surface, the default since 0.4.2. JUNE_TOOL_PROFILE=full lists the same capabilities as 30 individually named tools instead (31 when you opt into june_ingest_file via JUNE_FILES_ROOT).

The tools

The default surface is compact: 20 tools, seven of which group related operations behind an op argument. JUNE_TOOL_PROFILE=full lists the 30 members under their own names instead — same capabilities, same gates, same behaviour.

Everything else keeps its own name: june_answer, june_search, june_enumerate, june_context, june_usage, june_remember, june_ingest, june_page_write, june_page_delete, june_canvas_create, june_doc_save, june_doc_delete, june_learn. A verb that can remove something is never folded in with one that cannot — so june_page_write and june_page_delete stay separate from june_page_edit, and each family carries one honest destructiveHint.

Old names keep working in your saved agent docs: the standing-docs digest carries the old-name → new-name map, and a call to a folded name is refused with the exact replacement (june_page_get is not a tool on this surface (compact): call june_page_read with op='get').

What each operation does:

Receipts on every read

When the engine runs with JUNE_USAGE=1 (desktop: Settings → Usage receipts), every june_answer / june_context / june_search result also carries receipt and a one-line receipt_footer:

receipt r_7f…: served 812 tokens (exact, tiktoken:cl100k_base) from 3 blocks across 2 docs
· 1 doc this session already had — june_usage(receipt_id="r_7f…") shows it in full

The connector sends X-June-Source: mcp and one X-June-Session id per server process, so the engine can record which documents this agent session already had (re-reads it avoided). The footer never says "saved": that word exists only on a receipt that holds a measured pair. An engine without receipts sends no footer, and june_usage answers plainly that they are off.

Descriptions are written for the agent (what → when → returns), and every clamped input is visibly noted back to the agent instead of silently truncated.

Agent memory — docs, skills, and the anti-forgetting digest

Long sessions forget: instructions an agent read at session start (its CLAUDE.md, your conventions) lose force thousands of tokens later. june-mcp fixes this structurally.

Agents save standing docs into June — kind='doc' for durable instructions (pinned=true = always in effect), kind='skill' for named procedures with a one-line when_to_use trigger (bodies load lazily, like skills should), kind='learnings' for an append-only dated log written via june_learn. Each doc is an ordinary June page in the docs canvas (JUNE_DOCS_CANVAS, default agent_docs), marked by a small metadata block — so you can open your agent's memory in the Junê app, read it, and edit it; the agent picks your edits up on its next refresh.

The anti-forgetting half: on the first tool call of every session, and then every 12 calls or 10 minutes (tunable), the connector attaches a compact standing_docs digest to an ordinary tool result — pinned bodies in full, skill trigger lines, doc one-liners. Tool results always re-enter the model's fresh context, so the instructions can't decay the way a system prompt does, in any MCP host, with no host cooperation. A digest that can't be built (service busy, canvas missing) is silently skipped — it never costs the carrying call anything. Set JUNE_DOCS_REFRESH=0 to turn the digest off; the doc tools keep working.

June teaches agents how to use it — from inside itself. The first save creates the docs canvas and seeds agent-memory-guide: the operating manual (what belongs in the system canvas vs a workstream canvas, the three kinds and when to use each, naming, what to pin, revision discipline, repo sync). It's listed in every registry and digest, agents read it with june_doc_get('agent-memory-guide') whenever unsure — and it's an ordinary page, so edit it and your agents follow your version. Before anything is saved, empty states return a setup walkthrough instead of a shrug, and the june_memory_setup prompt has the agent interview you and save your conventions as the first docs.

Making June automatic — the agent depends on it without being told

"Use June" should never need saying. Three mechanisms stack to make usage automatic, each covering the previous one's blind spot:

1. The host hook (closes the cold start). A server can't speak until the agent's first call — so install June's standing instructions into the file your host loads natively every session:

   JUNE_EXPORT_ROOT=/path/to/project june-mcp --install-instructions            # → CLAUDE.md
   JUNE_EXPORT_ROOT=/path/to/project june-mcp --install-instructions AGENTS.md  # other agents
   

   It's written as a managed section (your own content is never touched; re-runs update it in place), and it puts the june-first posture — check June before claiming ignorance, remember facts unprompted, learn lessons as they happen — into the system prompt itself.

2. Proactive tool descriptions (never decay). The core verbs' descriptions tell the model when to reach for them unasked — and descriptions are re-read on every single turn, in every MCP host, with no cooperation needed.

3. The pinned june-first doc (re-asserts all session). Seeded alongside the guide, it rides every standing_docs digest, so the posture is repeated mid-session exactly where long-context drift would otherwise erode it. Like everything seeded, it's an ordinary page — edit it and your agents follow your version.

What no MCP server can do — honestly — is force a host to act: an agent whose host hides SERVER_INSTRUCTIONS and has no instruction file and never makes one June call stays cold. Mechanism 1 exists precisely so that case never occurs in practice.

Repo sync — the repo stays current with what June knows

Opt in with JUNE_EXPORT_ROOT=<your repo> and three more tools appear:

Safety rules, all enforced in code and pinned by tests: every path is fenced inside the root (lexical .. check and symlink resolution); a file not written by june-mcp is never overwritten; nothing is ever deleted; and with JUNE_EXPORT_GIT=1 each export commits exactly the files it wrote — pathspec-limited, so your staged work is never swept in, and push never happens. Exported files carry frontmatter and are byte-deterministic, so an unchanged doc re-exports to an identical file and git stays quiet.

The manifest (.june-export.json) makes currency checkable — two CLI modes for CI:

june-mcp --export         # sync agent docs + every managed page/section, commit if enabled
june-mcp --export-check   # write NOTHING; exit 1 if the repo has drifted from June

--export-check in CI turns "are the docs up to date?" from a hope into a failing build.

Free vs Pro — the june-pro tag

june-mcp is one package for everyone; there is no separate "pro build". Pro is a property of the endpoint, not the connector: connect to a Pro-activated June (a Pro license in the app, a Pro key on a hosted workspace) and the same tools carry Pro-grade results: every june_remember and june_ingest_file write runs the richer entity/edge engines automatically (the result reports which engine ran), june_resolve upgrades to semantic matching, and june_enrich backfills memories that were written on the free floor before you upgraded. The terminal shows which world you're in: --doctor prints an edition line and the server's startup banner tags the connection —

june-mcp: connected http://localhost:8000 canvas name "work" → 11d2… [june-pro]

The tag is read from the service's own /v1/whoami (the same entitlement state that gates Pro routes server-side), so it can't disagree with what you actually get — and it's display-only: entitlements are enforced on the service no matter what any client prints. Older services without /v1/whoami simply show no tag.

Security model

The tool surface exposes no canvas/workspace parameter — the workspace is bound server-side from your connection's context, fail-closed. A cross-tenant read isn't a permission check that could fail open; it's unrepresentable from the client. JUNE_READONLY=1 adds a second fence for read-only deployments. Your BYO LLM key rides each answer request as a header and is never persisted or logged by the service.

Errors

Every upstream failure maps to a typed, redacted error payload (built from exception type + HTTP status only — never from response bodies), so the server survives anything the endpoint throws and your agent sees a clean, actionable message.

License

MIT. The Junê engine itself is a separate, closed-source product — this connector is the open part, by design.