What shipped.

Bring your assistant's memory to every agent

ChatGPT, Grok, Claude, and the rest each learn how you work, then keep it to themselves. ctxfile now imports that memory, at the scope it belongs to, and nothing that could steer an agent goes live until you approve it. Free core.

• New tool ingest_memory and prompt ctx-import-memory: an assistant exports what it knows as one fact per entry, in batches of up to 100, against a strict schema with field-by-field errors. Two scopes: global (about you, every project sees it) and project.
• Review built in: instructions and identity entries wait for your approval. Duplicates from different assistants merge into one entry; a rejected entry stays rejected on re-import.
• get_context carries active memory in a new memory block, attached at read time so approvals show immediately, capped at 4,000 tokens.
• ctxfile memory prompt|import|list|approve|reject|export, plus an import and review section in the dashboard Memory view. export prints dated lines under category headers, ready to paste into another assistant or an AGENTS.md.
• Sessions: user_directives on save_session and ingest_context turns rules you state mid-session into pending instructions, and ctxfile ingest import takes a pasted session from chats without MCP.
• Sync and relay: memory syncs end-to-end encrypted, global memory rides every vault, and the relay serves ingest_memory so ChatGPT, Grok, and claude.ai can export straight into your vault. Handoff grants and org federation never see memory.

A new instrument: the dashboard and ctxfile.dev, redesigned

Same engine, same privacy, a new surface. The local dashboard and this site now share one “glass instrument” design system: a deep canvas, translucent panels, and LED-grade status, in dark and light.

• Dashboard shell: a grouped, collapsible sidebar with digit shortcuts, a top bar that shows the snapshot recording live from any view, the R key everywhere, a fuzzy command palette, and the view kept in the URL so refresh and back keep your place.
• Reading captured context is now pleasant: syntax-coloured files with line numbers and wrap, plan and session text rendered as Markdown, a collapsible JSON view, and a filterable tree that stays smooth with thousands of key files.
• Overview gains a token-budget gauge, connector timing bars, a usage chart over recent snapshots, and the heaviest files at a glance. Git shows a commit timeline and diffstat bars. Sessions group by day with source filters; Memory groups by agent with search; Playbooks highlight the placeholders to fill; Consult streams Markdown with a Stop button.
• ctxfile.dev: new hero, install card, docs rail grouped by topic, code windows with a copy key, and a rewritten design-system page that renders the live tokens.
• No behaviour, API, or privacy change: the dashboard still makes zero network calls.

Your Obsidian vault, readable by every agent

Point ctxfile at a vault — any folder of markdown — and every agent you use loads your notes as context. Read-only, fully local, redacted like everything else.

• Obsidian vault connector: notes join snapshots as a notes section, ranked by pin (ctxfile: pin frontmatter), relevance to your recent threads, and recency — with one-hop [[wikilink]] previews for graph context.
• ctxfile init detects a nearby vault and offers to connect it, with PARA-aware defaults (Archive and Resources stay out unless you opt in).
• Privacy holds: vaults never sync, repo-safe exports never include notes, and every note passes the same redaction as project files.
• Agents now say which notes informed a context load — visibility over silence, per the behavior pack.

One key for every model, one skill for every agent

Wider reach and a docs pass that checked every setup instruction against each tool's current documentation.

• OpenRouter consult provider: name a model slug, set OPENROUTER_API_KEY, and consult fans one question across hundreds of models on a single key.
• Per-client setup tabs in the docs: copy-paste config for Claude Code, Cursor, Codex CLI, OpenCode, Gemini CLI, OpenClaw, Hermes, Claude Desktop, and Aider.
• The behavior-layer skill now renders for nine harnesses: one portable SKILL.md for Claude Code, OpenCode, OpenClaw, and Hermes, a Cursor rule, and managed GEMINI.md / AGENTS.md blocks.
• Live-docs audit fixes, including Hermes tool naming (mcp_ctxfile_*) and the Claude Desktop extension manifest.

Playbooks, and easy to find

A new Pro surface, and a site any search or answer engine can read.

• Playbooks (Pro): ask any agent to distill a reusable prompt from your own saved sessions. Stored encrypted with provenance, served back as native MCP prompts.
• Discovery layer for the site: robots and sitemap, an llms.txt, and JSON-LD so search engines and AI answer engines can read and cite ctxfile.
• A glass-case hero and a brand kit (wordmark, marks, manifest), for dark and light.

Your context, in the chat tabs

The same context your CLI and editor agents load now travels to the web chatbots.

• Web-chatbot connector surface on the relay: search and fetch tools built for custom connectors, plus a tokened URL for claude.ai. Connect Grok, ChatGPT, Claude web, or Perplexity to your vault.
• The Sync relay is published to npm as @ctxfile/relay and self-hostable: encrypted vaults, bearer-token auth, one Docker image, deployable to Fly.
• Opt-in full-conversation transcripts: keep the whole chat alongside the digest, redacted, retrieved only on demand.
• Thread-naming hints, so a thread-less save can be re-filed and resumed by name.

Subscription licensing

Activation accepts a subscription credential that auto-refreshes a short-lived, offline-verified license. A Pro subscription stays valid without any phone-home while you work.

Launch: local core, threads, and Sync

The free core is Apache-2.0. The paid tiers verify their license offline and run locally.

• Snapshot engine: plan, key files, git state, opt-in Notion pages.
• Secret redaction on everything ingested; denied paths never read.
• SQLite snapshot cache; opt-in anonymous telemetry, default off.
• MCP server over stdio and Streamable HTTP: get_context, save_session, continue_thread, list_threads, ingest_context.
• Cross-provider threads: save in one agent, continue in any other; handoff packages for clean handoffs.
• Sync: end-to-end encrypted vault on a relay you can self-host, with a recovery code for a lost passphrase.
• Session connectors: Claude Code and Cursor session digests flow into the snapshot (Pro).
• Encrypted memory: cross-session memory, AES-256-GCM at rest, key in the OS keychain, provenance on every entry (Pro).
• Multi-provider consult with answer diffing: Anthropic, OpenAI-compatible, or Ollama (Pro).
• Voice notes via whisper.cpp with repo-aware vocabulary (Pro).
• Ed25519 offline licensing: signed key, local verification, 14-day grace, no phone-home.
• Local dashboard server: ctxfile ui, bound to 127.0.0.1.