Vaner
Status: alpha (pre-1.0). Interfaces may evolve quickly while we stabilize core behavior.
What is Vaner?
Vaner is a local-first context engine for coding agents that turns idle compute into useful future context.
Instead of waiting for a prompt and then starting retrieval from cold, Vaner continuously prepares likely
next-context packages, scores them, and serves the best fit quickly when the real question arrives.
Vaner predicts what the user will likely ask next and prepares evidence-backed context packages in advance.
Instead of reacting from a cold start, it continuously turns idle compute into prepared context that can be
served quickly when the real prompt arrives.
It is built around evidence-backed scenario memory, not chat-log accumulation:
- context is compiled from repo evidence and stored as reusable scenarios
- memory has explicit lifecycle states (
candidate, trusted, stale, demoted)
- fingerprints tie memory to concrete sources so drift is detected automatically
- conflict handling can abstain instead of blending contradictory evidence
Install
One-line installer (Linux/macOS)
curl -fsSL --proto '=https' --tlsv1.2 https://vaner.ai/install.sh | bash
Common installer flags:
--yes (non-interactive), --dry-run, --verify--backend uv|pipx--version <tag>--with-ollama--minimal (legacy minimal extras; pairs with --no-mcp when needed)--backend-preset ollama|lmstudio|vllm|openai|anthropic|openrouter|skip--backend-url <url>, --backend-model <model>, --backend-api-key-env <env>--compute-preset background|balanced|dedicated--max-session-minutes <n>--no-mcp (skip MCP extras)
Manual install (advanced)
pipx install 'vaner[all]'
# or
uv tool install 'vaner[all]'
Upgrade
vaner upgrade
vaner version
What Vaner stores and when it expires
Vaner stores local runtime state in your repository under .vaner/ (for example store.db, scenarios.db,
metrics.db, and telemetry.db). Retention is controlled by limits.max_age_seconds in .vaner/config.toml.
Old signal/replay/query and stale cache entries are purged during engine/precompute activity. There is not yet a
separate prune or compact command.
Useful commands:
vaner config set limits.max_age_seconds 3600 --path .vaner forget --path . (remove local Vaner state for this repo)vaner uninstall --path . --keep-state (remove client wiring while keeping state)
Configure
Run the setup wizard per repo:
vaner init --path .
Key non-interactive flags:
--backend-preset--backend-model and --backend-api-key-env--compute-preset background|balanced|dedicated--max-session-minutes <n>--force--interactive/--no-interactive--no-mcp
What vaner init does by default:
- writes repo-local config to
.vaner/config.toml
- writes repo-local Cursor MCP wiring to
.cursor/mcp.json
- writes managed feedback skills to
.cursor/skills/vaner/vaner-feedback/SKILL.md and ~/.claude/skills/vaner/vaner-feedback/SKILL.md
- installs per-client usage primers (short guidance about when and how to use Vaner) into each detected client's rules surface:
.claude/CLAUDE.md (Claude Code), .cursor/rules/vaner.mdc (Cursor), .github/copilot-instructions.md (Copilot), AGENTS.md (Codex CLI), .clinerules (Cline), .continue/rules/vaner.md (Continue). Primers are non-destructive: existing files get a delimited <!-- vaner-primer:start v=… -->…<!-- vaner-primer:end --> block that re-runs replace in place without touching content outside it.
Use vaner init --no-mcp --path . if you want config only and do not want Vaner to wire MCP clients or managed
skills. Use --no-primer to skip the client-level guidance. Use --user-primer to also install the Claude Code
primer at ~/.claude/CLAUDE.md (always-on across every session) in addition to the repo-level .claude/CLAUDE.md.
If Cursor is already open, reload the window after vaner init so the new MCP server is picked up.
Start / verify runtime:
vaner up --path .
vaner status
vaner logs --daemon
vaner down
Supported MCP clients
