Distillery 0.4.0 Stabilizes Its MCP API

What Changed — The API Hardening Release

Distillery 0.4.0 shipped on April 19, 2026, with one clear mission: turn the MCP tool surface into a public contract. This isn't about new features—it's about stability. The release hardens tool names, parameter shapes, error codes, and response envelopes. Breaking changes now require a major version bump, with deprecation warnings first.

This follows Distillery's initial launch as a solution to Claude Code's "evaporating knowledge" problem. While Claude Managed Agents launched with memory labeled "research preview," the community ecosystem (including Distillery, memsearch, Honcho, Hippo, and others) rushed to fill the gap. Teams need production-ready memory layers, not preview features.

What It Means For Your Claude Code Workflow

Stable Foundations for Agent Development

If you're building agents with Claude Code, the memory layer sits under everything. Planners read from it, tools write to it, and evaluations depend on it being deterministic. Before 0.4.0, drifting tool names or changing response shapes could break downstream agents. Now you can pin against min_server_version=0.4.0 with confidence.

Practical Improvements That Matter Today

• distillery_list defaults to output_mode="summary": A typical limit=50 GitHub sync response shrinks from ~300KB to a few kilobytes of titles, tags, and previews. This directly impacts your context window usage.
• Async feed syncing: gh-sync now runs via server-side background jobs. Long syncs don't block your Claude Code session—poll distillery_sync_status for progress instead.
• Better scheduling: /setup and /watch now configure Claude Code routines instead of CronCreate jobs or GitHub Actions. Three routines ship: hourly feed poll, daily stale check, weekly maintenance.
• Cleaner interest profiles: Feed entries are excluded from the interest profile, so /radar won't drift toward whatever feed happens to be loudest that week.

Try It Now — Commands That Work Differently

Update Your Installation

# Update to the stable release
npm update -g @distillery/cli

# Or install fresh
npm install -g @distillery/cli

Configure Claude Code Routines

After updating, run:

claude code /setup

This configures three Claude Code routines instead of external cron jobs:

• Hourly feed polling
• Daily stale entry checking
• Weekly maintenance

The deprecated webhook endpoints (/hooks/poll, /hooks/rescore, /hooks/classify-batch) will log warnings if hit.

Use the New Defaults

When querying your knowledge base:

# Old way (could flood context window)
claude code /recall "authentication patterns"

# New default behavior - automatically uses summary mode
claude code /recall "authentication patterns"
# Returns: titles, tags, previews instead of full content

# If you need full content, explicitly request it
claude code /recall "authentication patterns" output_mode="full"

Monitor Async Operations

For long-running sync operations:

# Start a sync
claude code /sync gh-sync

# Check status without blocking
claude code /recall sync_status

The Bigger Picture: Karpathy's LLM Wiki Pattern

Distillery implements what Andrej Karpathy calls the "LLM Wiki" pattern: raw sources → synthesized wiki → query layer. This aligns with the broader trend in agent memory systems, where three-tier architectures (fast index, episodes, raw transcripts) are becoming standard.

The community has seen rapid iteration here—within a week of Karpathy's post, projects like Knowledge Raven, Memoriki, OpenTrace KG-MCP, and OptiVault emerged. Distillery takes a pragmatic approach: single DuckDB table with hybrid BM25 plus vector search, rather than complex L0/L1/L2 access tiers.

What makes Distillery operational is its focus on provenance, correction, and expiration. Every entry carries author and session ID. Entries can be corrected without losing history. They can be marked expired without deletion. These primitives let a shared knowledge base admit when it was wrong—essential for a living memory layer.

What's Next: Dashboards and Integrations

With the MCP surface stabilized, the dashboard/ directory in the repo is becoming a SvelteKit dashboard. Community plugins can now build on Distillery with the same confidence as any public SDK. LangChain orchestrators, Letta-style frameworks, and MCP-native runtimes can treat Distillery as a durable backend.

The team is tracking LongMemEval benchmarks and transcript-mining integration in issue #233—this is how they'll measure against three-tier systems like Vektori and MemPalace.

For now, the message is clear: if you're using Claude Code with a team, your knowledge base just graduated from experimental to production-ready.