snapdec

Snap decisions for coding agents. One command, any machine, any agent: fast, calibrated typed decisions (classify / check / score / rank) from a local model or hosted API — exposed to every coding agent through one shared MCP server plus a portable Agent Skill.

• One prompt setup — snapdec shows your PC spec and a menu of every backend your hardware can actually run, with honest accuracy stats. You choose.
• Local = fully automatic — pick a local model and snapdec creates a runtime venv, downloads the model, launches the server (127.0.0.1 only), and wires it into every coding agent it finds. No other steps.
• Hosted = paste a key — provider auto-detected from the key format, model auto-selected (free variant first), validated by a live canary call.
• Honest by design — every answer carries calibrated probabilities and an auto | review decision. Uncertain → the agent decides itself. Advisory only; never used to approve destructive actions.

Install

uv tool install snapdec     # or: pipx install snapdec / pip install snapdec
snapdec init

Dev (from source): uv venv && uv pip install -e ".[dev]" && uv run pytest

The first-run menu (real example, 16 GB laptop, no dGPU)

Recommended for this machine (13th Gen i5 · 16 GB RAM · Intel UHD · Windows 11)

  LOCAL — free · private · offline
    [1] Laya EN (421M)  —  DI ~0 zero-shot (specialize-first base)
        5–15 ms GPU/Apple · 50–450 ms CPU · setup: ~2 GB        (recommended)
    [2] Laya multilingual (322M)  —  DI ~0 zero-shot · 100+ languages
    [3] Kev 0.8B  —  DI 13.26 · OOD acc 0.652
        ~40–80 ms CUDA · CPU: seconds/question · setup: ~3 GB   [slow on CPU]
  HOSTED — API key · best accuracy
    [4] OpenRouter  —  free keys (openrouter.ai/keys) · hosts Jev + Kev · Jev DI 51.67 (best known)
    [5] TypeSafe Jev  —  native /v1/systemone · Jev DI 51.67 (best known) · paid per call
    [6] Other /v1/systemone URL

Choice [1]:

Only what fits your hardware is listed — a machine without a 12 GB+ GPU never sees Kev-4B (DI 31.31); every listed option is genuinely runnable. Stats are zero-shot Decision Index numbers (chance-corrected, 40 benchmarks; Jev 51.67 is the best known reference) from our research.

Picking [3] on the machine above works — you get the honest [slow on CPU] flag first. Freedom within your spec.

After you choose, snapdec automatically:

1. provisions the backend (local: venv + download + server; hosted: key validation + model canary),
2. registers its MCP server in every coding agent it detects (Claude Code, Codex, Cursor, OpenCode, Antigravity, Windsurf, VS Code, Cline — CLI-first, backup-first, exactly reversible),
3. installs the Agent Skill globally so every repo gets it,
4. starts the shared daemon and prints a live health check.

What agents get

Multi-item tools fan out one request per item (small-context backends can't answer N items against one blob) and run the calls in parallel.

Example (CLI mirror of the MCP tool, real output from Laya on a CPU laptop):

$ echo '{"items":[
    {"id":"t1","text":"Tests failed: OSError network unreachable on runner"},
    {"id":"t2","text":"AssertionError: expected status 200, got 500"},
    {"id":"t3","text":"ModuleNotFoundError: No module named requests"}],
  "classes":{"infra":"network/runner problem","bug":"real code bug",
             "deps":"missing dependency"}}' | snapdec classify --input -

t1: bug    p=0.73  decision=review
t2: bug    p=0.94  decision=auto
t3: deps   p=0.61  decision=review

auto results can be acted on in bulk; review items go back to the host agent — that abstention is the product.

CLI

snapdec init                 # the wizard (above)
snapdec init --yes --api-key sk-or-…      # non-interactive, provider auto-detected
snapdec init --yes --backend local        # non-interactive managed local setup
snapdec doctor --live        # verify install, daemon, every agent registration
snapdec daemon start|stop|status|logs
snapdec agents list|add|remove|print      # integrations, all reversible
snapdec classify|check|score|rank --input -   # CLI mirrors of the tools
snapdec project-facts .      # tier-0 facts straight from the terminal
snapdec uninstall            # reverses every integration exactly

Architecture (30 seconds)

agents (Claude Code, Codex, Cursor, …)
   │ stdio MCP (one thin shim per agent, <1 s start, no ML imports)
   ▼
snapdec daemon (one shared process, loopback+token IPC / UDS)
   ▼
backend: managed Laya (8901) · managed Kev (8902, pinned git SHA)
         · hosted /v1/systemone (OpenRouter, TypeSafe, any URL) · mock

Keys live only in snapdec's own state (env-var reference or 0600 file) — never in agent configs. Local servers bind 127.0.0.1 only. No telemetry.

Full design doc with research citations: SYSONE_ARCHITECTURE.md · decisions: docs/adr/

Status

Phase 1/2 complete for: tier-0, managed Laya + Kev provisioning, hosted backends, 9 agent integrations, MCP v2 shim + shared daemon. Windows is verified on real hardware; macOS/Linux covered by CI and follow the same paths. Honest gaps: Kev on CPU is slow (flagged in the menu, not hidden); Laya zero-shot is weak (it's a specialize-first base — the stats say so in the menu); calibration refit and micro-batching are Phase 4.

License

Apache-2.0. Not affiliated with TypeSafe/Jev, Kev, or Laya — snapdec routes to them and credits them. Model licenses: Apache-2.0 (Kev, Laya).