GNO.

• 01

  Conceptual misses

  Keyword search returns nothing the moment you paraphrase your own note.

• 02

  Unverifiable context

  A ranked list tells you what scored well. It cannot tell you what it failed to find.

• 03

  Confident hallucination

  Assistants fill gaps with plausible prose instead of admitting the evidence ran out.

• 04

  Data leaves

  Cloud search ships your vault to someone else’s vector database.

GNO is built for the part after the demo.

• Fast

  gno search

  Exact strings, identifiers, error codes, quoted phrases.

  Document-level BM25 over SQLite FTS5 with Snowball stemming. Starts without loading a model.

• Semantic

  gno vsearch

  Conceptual queries and half-remembered phrasing.

  Vector search via sqlite-vec, about 70 ms on a 17,600-chunk index. Chunks are embedded with their document title for context.

• Recommended

  gno query

  The daily driver.

  BM25 + vectors + optional graph expansion, fused per document with RRF and reranked by a cross-encoder. --fast answers without expansion or rerank; --explain shows the score components.

• Evidence

  gno context build

  Multi-document work an agent or a reviewer will reason over.

  Compiles a deterministic Capsule under one global budget, with gaps and omissions stated.

• Enforced

  gno ask --verify

  Answers where being wrong costs more than being slow.

  Generates against one closed Capsule and abstains below full substantive-claim support or on any misquoted source.

Every mode takes the same hard filters: one collection or several (repeat -c), tags, dates, authors, and your own typed fields. Nest project, status or confidence under gno.metadata in frontmatter, and --filter can ask for approved notes with a confidence of at least 0.8. The same predicate works without a search term: gno ls --filter lists every document that matches it.

• 01

  Split-view editor

  Live markdown and preview, auto-save, wiki-link completion.

• 02

  Keyboard-first

  ⌘K palette, ⌘N capture, ⌘B/I formatting, ⌘S save.

• 03

  Cross-collection browse

  Tree sidebar, folder detail panes, per-tab browse context.

• 04

  Health centre

  Per-folder lexical proof, the exact failed stage, and remediation.

• 05

  Connectors

  Install Skill/MCP integrations and run an explicit read-only retrieval check.

• 06

  Graph view

  Wiki, markdown, similarity and typed edges, navigable.

• 07

  Sessions page

  Manage archive sources, preview and run imports, and search turns with Human and Assistant badges.

• 08

  Desktop app (beta)

  A native window that starts and supervises gno serve and opens gno:// links. Signed macOS DMG; Windows x64 zip.

• 09

  GNO Recall for Omarchy

  A bar widget, a health panel, and a Super+R overlay for search and collection browsing.

The desktop app is a beta and can trail the CLI release. GNO Recall runs on its own verified GNO runtime, so a global upgrade does not update it.

• 01

  Context Capsules

  One bounded evidence bundle per goal: exact spans, content hashes, one global token budget, collapsed duplicates, declared gaps.

• 02

  Verified answers

  Generate against one closed Capsule, classify every substantive claim, withhold the draft below 100% support.

• 03

  Freshness receipts

  Watch a saved Capsule. When its sources move, the resident runtime reverifies the exact spans and records a receipt.

• 04

  Proof-based setup

  gno setup returns only after lexical search finds a real corpus-derived hit. No green checkmark over a broken index.

Measured on 48 paired tasks in the agentic retrieval benchmark shipped in the repository. The raw receipts, pinned corpus and stated limitations are published alongside, so anyone can replay the result.

• 01

  Scoped facts

  gno remember stores one fact per Markdown file with 1 to 8 explicit scopes. gno recall returns at most 8 current facts under 512 tokens, each with a gno:// cite.

• 02

  Supersede, never overwrite

  A replacement names its predecessor by URI and content hash. Two agents racing to replace the same fact get one successor and one conflict. Search, get, ask and Capsules mark a replaced fact with its successor.

• 03

  Session archive

  gno sessions imports selected Codex, Claude Code, OpenClaw and Hermes conversations into a separate archive with its own config and index.

• 04

  Who said what

  Each archived turn is labelled human or assistant, carries a provenance line, and has credentials redacted before it is written.

• 05

  Manual until enabled

  A fresh install imports, watches and schedules nothing. Per profile you can switch on a Claude Code SessionEnd hook or a schedule that gno daemon runs.

• 06

  Retry-safe writes

  Send a request ID with a capture, a remembered fact or a document save. A retry after a lost response replays or finishes that write once.

gno remember "Prod deploys from main only" --scope project:api --add
gno recall "deploy branch" --scope project:api --json

gno sessions discover                 # preview; imports nothing
gno --config ~/gno-sessions/archive.yml --index sessions \
  sessions import --source codex --dry-run

gno request-status "$id"              # committed? pending? then retry

Imported turns are evidence. Nothing from a session becomes a remembered fact until someone runs gno remember and points its source at the turn. Redaction is best effort, so keep a new archive local_only until you have read through it.

A partial list. gno serve and gno daemon also expose the same surface as stateful Streamable HTTP at 127.0.0.1:3000/mcp, so several clients share one warm store and one model lifecycle instead of each paying a cold start. An agent can reindex just the files it changed; requests that arrive while a job runs queue and merge into one pass.

MCP clients

Claude Desktop

gno mcp install

Claude Code

gno mcp install --target claude-code

Cursor

gno mcp install --target cursor

Codex

gno mcp install --target codex

Zed

gno mcp install --target zed

Windsurf

gno mcp install --target windsurf

OpenCode

gno mcp install --target opencode

Amp

gno mcp install --target amp

LM Studio

gno mcp install --target lmstudio

LibreChat

gno mcp install --target librechat --scope project

Agent skills · no protocol overhead

Claude Code

gno skill install --target claude --scope user

OpenAI Codex

gno skill install --target codex

OpenCode

gno skill install --target opencode

OpenClaw

gno skill install --target openclaw

Hermes

gno skill install --target hermes

Retrieval protocol · harness instructions

Install

gno agents install --target claude

Verify

gno agents verify --target claude

All detected

gno agents install

The skill teaches the commands. The protocol block tells the harness when to retrieve, which search path to take, and how to cite and file what it learns. gno agents manages that block in the global instruction files of Claude Code, Codex, Cursor Agent, OpenCode, Hermes and OpenClaw, and leaves the rest of each file alone.

• 01

  Three policy levels

  local_only, lan, or remote per collection. Existing collections migrated fail-closed.

• 02

  Policy follows content

  Capsules, traces, embeddings and exports inherit the most restrictive contributing policy.

• 03

  Relaxation is a handshake

  Loosening needs the exact current revision, single-use. Stale and replayed confirmations fail closed.

• 04

  Independent gates

  Authentication and write permission never grant egress. A valid token cannot pull local_only content off the box.

• 05

  Content-free receipts

  Bounded local audit records the action and destination class, never the passage.

• 06

  Conservative classification

  DNS rebinding, redirects, mixed evidence and stale confirmations all fail closed.

gno collection policy set client-work local_only
gno collection policy get notes            # note the revision
gno collection policy set notes lan --confirm-relaxation 3
gno egress-audit list                      # content-free receipts

gno changes --since 2026-07-20T00:00:00Z   # lifecycle journal
gno diff    gno://notes/plan.md            # structural delta
gno impact  gno://notes/plan.md --max-depth 3
gno context watches                        # which evidence went stale
gno audit   links --collection notes       # broken or ambiguous links

Journal entries and notifications are bounded metadata. No source bodies are retained, so the diff is structural: headings, links, typed relations. When retention expires a cursor, the response says so and points at the earliest available one.

gno audit runs one offline, read-only report over links, declared provenance, and source/index freshness. Findings carry stable IDs, so two runs can be diffed. It reports what it saw and repairs nothing. A collection can opt into its own rules: which other collections it may link to, whether external URLs in its notes are reported, and which folder layout its documents must follow.

gno trace label <id> --label missing-expected \
  --target gno://notes/postmortem-q3.md
gno trace export <id> --format qrels --output recall.json
gno trace replay <export-id> --candidate hybrid --md

Recording is off by default. Only judgments you state explicitly become qrels, a missing click is never a negative, exported fixtures contain no source text, and replay is read-only: it can recommend a promotion, it always returns applied: false. The most valuable label is missing-expected, the recall failure a click-based system can never see.

• 01

  Export, then upload

  gno publish export writes a self-contained artefact. Nothing outside it reaches the server.

• 02

  Four visibility modes

  Public URL, secret link, invite-only space, or encrypted locally before upload. The Web UI asks for the mode explicitly.

• 03

  Reader-grade typography

  Editorial serif type, drop caps, outline nav, reading time, j/k and / hotkeys, Shift+Y to copy Markdown.

• 04

  Filtered projection

  Backlinks and related notes resolve only within the published subset.

• 05

  Stable URLs

  Republishing updates the same link. Each publish is a snapshot; rotate or revoke when needed.

• 06

  Agent manifest

  Public snapshots carry a read-only projection with content hashes and line spans.

• 01

  Agent long-term memory

  Claude Code, Codex, Hermes and OpenClaw recall scoped, cited facts from one store and retrieve documents on demand.

• 02

  Past agent sessions

  Find the turn where you picked a library, next to the proposal the agent made before it, across Codex and Claude Code.

• 03

  Due diligence

  Mixed-format client corpus pinned local_only; each finding backed by a re-verifiable Capsule.

• 04

  Second brain

  Obsidian vault, journals, saved research. Semantic search over what you actually wrote.

• 05

  Engineering knowledge

  ADRs, RFCs, runbooks across repos. Structural chunking for TS, JS, Python, Go, Rust.

• 06

  Meetings and mail

  Transcript and mailbox exports indexed as individually citable records.

slim-tuned (default)

Tuned local query expansion

slim

Untuned slim expansion

balanced

Qwen2.5 3B expansion and answers

quality

Qwen3 4B expansion and answers

Roles are separable: embed, rerank, expand and gen can each point at a local GGUF or a remote HTTP endpoint, and a single collection can override the global embedding model when code and prose want different things. The default expansion model is a published fine-tune, trained on a benchmark fixture; your documents are never part of training.

Chunk size and overlap are set per index. Leave them out and you keep about 800 tokens with 15% overlap; change them for a retrieval experiment and gno status reports the policy actually applied.