supernovae-st/nika

Nika

Intent as Code. The workflow language for AI: one file, 4 verbs, one binary.

Useful AI work shouldn't disappear into chats. Nika turns repeatable AI work into files you can run, review, diff and share. If you do the same AI task twice, make it a workflow.

The pipeline is a file. A graph of model calls, tools and processes: a shape you declare, not glue you program. Nika audits that file before a token is spent (cost ceiling, permissions, secret flows, types), runs it on whichever LLM you choose, local first, no cloud required, and leaves every run a verifiable receipt. The language is an open Apache-2.0 spec; this repo is the reference engine, a single Rust binary (AGPL-3.0). The way SQL pairs with PostgreSQL, or the Dockerfile with Docker.

Does it run today?

Yes.

brew install supernovae-st/tap/nika    # or: curl -LsSf https://nika.sh/install.sh | sh
nika examples run 01-hello --model mock/echo            # zero setup: no key, no model server
nika examples run 01-hello --model ollama/llama3.2:3b   # got Ollama? the same run, real + local
# (first run loads the model into memory; later runs are much faster)

Nika audits a workflow before a single token is spent (plan, cost ceiling, secret flows, types, tool args), then runs it:

$ nika check brief.nika.yaml
 ✔ PLAN     2 wave(s) · 2 task(s) · max parallelism 1
      wave 1 fetch_notes (exec · sh -c)
      wave 2 brief (infer · ollama/llama3.2:3b)
 ✔ SECRETS  no information-flow escapes
 ✔ TYPES    every deep output reference fits its declared shape
 ✔ TOOLS    every nika: tool names a canonical builtin
 ✔ ARGS     every invoke arg key is declared + every required arg is present
 ✔ SCHEMA   every authored schema: is satisfiable
 ✔ audited · 2 task(s) · 2 wave(s) · permits none · est ≥$0.0000

$ nika run brief.nika.yaml
  🦋 nika · daily-brief · 2 tasks
  ✔  fetch_notes  exec · cat
  ✔  brief        infer · ollama/llama3.2:3b
  ── 2/2 done · $0.000 · elapsed 16.2s ───────────────────────────

The loop: check → fix → run → receipt

An agent (or you, at 2am) writes a workflow. nika check audits the file statically and names every fix; the run streams live; the trace is a hash-chained receipt nika trace verify re-proves. The whole loop, captured against the real binary:

What the audit catches before a single token is spent. Each finding carries its NIKA-XXXX code, the exact source span, and the fix:

After the run, the receipt: every run journals to .nika/traces/ as an append-only, hash-chained record. The run prints its chain head; nika trace verify recomputes it. Tamper-evident, local, zero services.

The capture above is honesty-gated like every asset here: its fixtures + tape are committed, and scripts/media/validate-media.sh keeps the broken half failing nika check and the fixed half clean.

What a workflow looks like

# review.nika.yaml: read a PR diff, judge its risk, comment only when it's high.
nika: v1
workflow:
  id: pr-risk-review
model: ollama/qwen3.5:9b             # local by default. swap to any provider

tasks:
  diff:                               # exec: a read-only shell command
    exec:
      command: ["git", "diff", "origin/main...HEAD"]
      capture: structured

  assess:                             # infer: structured LLM judgment
    with:                             # the binding IS the edge — no separate wiring
      patch: ${{ tasks.diff.output.stdout }}
    infer:
      prompt: "Risk-assess this diff (secrets, breaking changes, missing tests). Be terse.\n${{ with.patch }}"
      max_tokens: 300
      schema:
        type: object
        required: [risk]
        properties:
          risk: { type: string, enum: [low, medium, high] }

  comment:                            # invoke: the only write, gated on the verdict
    with:
      risk: ${{ tasks.assess.output.risk }}
      verdict: ${{ tasks.assess.output }}
    when: ${{ with.risk == 'high' }}
    invoke:
      tool: "mcp:github/pr-comment"
      args:
        body: ${{ with.verdict }}

Check before it runs

nika check is a static audit. It catches broken references, missing dependencies, schema and permission problems before any model is called, and when something is off, it points at the exact fix:

The two fixtures behind this capture live in scripts/media/fixtures/, gated in both directions by scripts/media/validate-media.sh: the broken one must keep failing nika check, the fixed one must stay clean.

The same audit holds the workflow's declared blast radius. A permits: block makes the file itself the security boundary: hosts, paths, programs, tools, all default-deny once declared. A task that reaches beyond it is caught statically, with the machine-applicable fix, before anything runs:

And failure handling is part of the file, not an ops runbook. When a task dies, on_error: recover: degrades to a declared fallback: the run completes, and the output says what it is:

Images are workflow citizens

nika:image_generate renders through the same discipline as everything else: a declared permits: boundary gates every save, the run ledger meters real spend, and provenance is structural, not a promise:

• Local-first: provider: local speaks the OpenAI-images wire any self-hosted server exposes (LocalAI · Ollama · stable-diffusion.cpp · SGLang · vLLM-Omni). The base URL is engine config, never workflow data. Clouds when you choose: openai · gemini · xai, and mock renders real, decodable PNG files offline for CI.
• Provenance survives cp: every saved PNG carries a nika tEXt chunk (tool · engine · provider · model · prompt · seed), the practice ComfyUI and InvokeAI standardized and no other workflow engine ships. The sidecar manifest adds sha256, resolved request and timing.
• Honesty is enforced: magic bytes beat declared MIME types, lossy provider mappings warn loudly, a provider returning fewer images than asked is a visible count_shortfall:, result URLs are never fetched, and base64 never rides workflow outputs: assets, not blobs.
• Real spend in the ledger: a render's exact cost (xAI bills in ticks) lands on the task line and the run total, the same honest-spend channel infer: uses.

Daily commands

nika welcome                         # the mirror · what Nika is + what it sees here · offline
nika welcome --deep --json           # the whole workspace, audited, one JSON (agents start here)
nika inspect flow.nika.yaml          # anatomy · tasks · waves · cost floor
nika check flow.nika.yaml            # the audit · exit 0 clean · 2 findings
nika explain flow.nika.yaml          # the story · waves · cost BEFORE a token · what it touches
nika explain NIKA-VAR-001            # any code · cause · category · fix-form
nika run flow.nika.yaml --var topic=rust   # launch inputs · repeatable
nika test flow.nika.yaml --update    # pin the golden · then `nika test` = offline CI
nika run flow.nika.yaml --task hero    # regenerate ONE task + its upstream cone
nika run flow.nika.yaml --resume .nika/traces/<run>.ndjson   # skip journaled successes
nika run flow.nika.yaml --resume <trace> --answer approve=true  # re-arm a paused gate
nika trace show .nika/traces/<run>.ndjson   # re-render any past run
nika catalog                         # the embedded provider/model catalog · capabilities · env vars
nika catalog --tools                 # the nika:* builtin catalog · what invoke reaches without MCP
nika model pull Qwen/Qwen3-0.6B-GGUF # a GGUF + its tokenizer land in ~/.nika/models
nika model serve --model Qwen/Qwen3-0.6B-GGUF  # your own loopback OpenAI endpoint · no cloud, no daemon
nika doctor --ping                   # are the local servers actually listening?

Every run writes its own journal to .nika/traces/ (opt out per run with --no-trace-file, globally with NIKA_NO_TRACE_FILE); --resume and nika trace read it directly. A paused run exits 4 (a blocking nika:prompt journals its question); cache hits on resume are always visible: nothing is skipped silently.

Pick a workflow

The binary embeds a versioned pack of runnable examples. Browse with nika examples list, read one with nika examples show <slug>, preview any of them with --model ollama/llama3.2:3b (or offline with --model mock/echo):

The full gallery, every workflow sha256-pinned and proven in CI, lives in examples/: foundation patterns, business showcases, and the skeletons nika new --from <template> instantiates (nika examples and nika new --from '?' list the live shelves).

Shared workflows live on nika-registry: every entry pinned to a full commit + sha256 and re-proven by CI (the conformance oracle + this engine's static certificate: exec · tools · cost, visible before anything runs). nika add is on the roadmap; today the registry's get.py does resolve → verify → audit in one command.

The model

Four verbs, and nothing else. A small core that composes into arbitrary real-world workflows. The Unix and SQL discipline of "small surface, large composition."

Everything sits under one frozen, versioned envelope, nika: v1, that won't break. Three properties hold across every workflow:

• Provider-agnostic, local-first. Local Ollama or LM Studio, any API — or no server at all: every release binary serves GGUFs itself (nika model pull → nika model serve, loopback, 0.101+). Your workflow doesn't change when the model does.
• Safe by construction. A read-XOR-write capability model. A step that reads cannot silently write; every effect is explicit and gated.
• Reproducible. The file and its execution trace are an auditable, re-runnable record.

flowchart LR
    F["workflow.nika.yaml<br/><i>portable · readable · verifiable</i>"] --> E["<b>nika</b><br/>single Rust binary"]
    E -->|infer| L["LLMs<br/>Ollama · LM Studio · any API"]
    E -->|exec| S["shell"]
    E -->|invoke| T["tools · MCP"]
    E -->|agent| A["autonomous loop"]

Dependencies make every workflow a graph: independent tasks run in parallel, an agent step fans out, joins wait for every branch, and the whole plan is known, costed and audited before execution starts (the capture at the top of this page shows exactly that).

Why Nika

The closest analogues aren't products. They're standards. SQL. The Dockerfile. A portable specification with a reference engine. The language is the contribution, not a product to sell.

As AI agents start acting on the real world, the interface where they act can't be free text (too vague) or raw code (too risky). It has to be a verifiable action language: one an AI writes, a human reviews and approves, and a machine runs deterministically. Kept open and sovereign, not locked inside one vendor's cloud.

What no existing workflow tool offers together: a single Rust binary · portable declarative YAML · local-first · read-XOR-write capability security · AGPL · no cloud required · bring-your-own-LLM.

Coming from Airflow, Dagster, LangGraph, Temporal or plain GitHub Actions? The docs keep an honest how-Nika-compares page, including when not to use Nika.

Status

Nika is built in the open.

The language (the nika: v1 envelope and its four verbs) is stable and won't break. The engine is a strict, modular Rust workspace. The latest tagged public release is whatever the badge at the top of this page says: always the releases page, never a number typed here (numbers in prose rot). main moves immediately to the next -dev version after each release so local contributor binaries cannot be confused with Homebrew assets. The 1.0.0 launch remains gated by the release checklist, not by a date. The code, the spec, and the example workflows are all readable, and development happens on main in the open.

The nika: v1 language envelope is frozen forever. It is a separate axis from the engine version. Every release is complete for its declared scope; no half-features parked behind a future version.

Get started

Install (macOS · Linux):

# Homebrew (macOS · Linux): on your PATH immediately
brew install supernovae-st/tap/nika
nika welcome   # thirty seconds: what Nika is + what it sees on this machine · offline

# …or, without Homebrew: the install script. It downloads the verified release
# binary into ~/.nika/bin and prints the single PATH line to add to your shell
# profile (then reopen the terminal, or `source` it, and `nika --version` works).
curl -LsSf https://nika.sh/install.sh | sh

Prefer a guided page? Every install path, step by step: nika.sh/install.

Fully manual / air-gapped? Download the platform tarball + SHA256SUMS from the latest release, verify with sha256sum -c SHA256SUMS --ignore-missing, then move nika onto your PATH.

Already carrying a toolchain?

# cargo: fetches the PREBUILT release tarball (no compile); the binary lands
# as `nika-cli` until the crates.io publish; symlink the public name once:
#   ln -sf ~/.cargo/bin/nika-cli ~/.cargo/bin/nika
cargo binstall --git https://github.com/supernovae-st/nika nika-cli

# nix: builds the exact release source via the flake (first run compiles,
# the store caches it); `nix profile install` for a durable install
nix run github:supernovae-st/nika

Deploying to containers (Dokploy · Coolify · Railway · any PaaS)?

# docker: the official multi-arch image (ghcr · linux/amd64 + linux/arm64),
# built on the release train from the same verified release binaries;
# mount your workdir to check/run workflows
docker run --rm -v "$PWD:/work" -w /work ghcr.io/supernovae-st/nika check hello.nika.yaml

# the MCP lane (stdio JSON-RPC): wire it into any MCP client config
docker run -i --rm ghcr.io/supernovae-st/nika mcp

Your first workflow runs with zero setup: no model, no API key:

cat > hello.nika.yaml <<'YAML'
nika: v1
workflow:
  id: hello
tasks:
  greet:
    exec:
      command: "echo hello from nika"
YAML

nika check hello.nika.yaml   # static audit, before a single token is spent
nika run hello.nika.yaml     # execute locally

Adding an AI step? Point it at a local model and nothing leaves your machine:

model: ollama/llama3.2:3b    # local · or mistral/..., anthropic/..., any provider
tasks:
  greet:
    infer:
      prompt: "Say hello in one sentence."

(No model handy? The built-in mock/echo previews any workflow offline.)

For real inference, run a local model (Ollama / LM Studio) or set a provider key, then see what's wired:

nika doctor                  # provider keys + local servers, with the exact fix
nika init                    # schema wiring + AGENTS.md for this repo
nika wire cursor             # explicit MCP wiring · also: vscode · windsurf · claude · codex · zed · all
nika examples list           # browse the embedded examples
nika examples run 01-hello --model ollama/llama3.2:3b   # a real local run

From source (contributors): git clone https://github.com/supernovae-st/nika.git && cd nika && cargo test --workspace --lib. End-user docs: docs.nika.sh.

Work with your agents

Nika is built to be written by agents and reviewed by you. nika init drops the schema wiring, AGENTS.md, the Cursor rule and a repo-level agent skill into your repo so Claude Code, Cursor, Codex and friends author valid workflows on the first try. nika wire <cursor|vscode|windsurf|claude|codex|all> points each client's MCP config at the engine: idempotent, and it preserves your other servers. nika mcp exposes a read-only oracle any MCP client can call: 8 tools, nika_check and nika_explain through nika_catalog and nika_tools. nika lsp speaks LSP to every editor.

The full map of every door (install paths, IDEs, agents, skills, MCP, CI, SDKs) is one page: docs.nika.sh/integrations/everywhere.

Wondering which model on your machine can actually drive a workflow? The model bench is a workflow that benches your models: same tasks, every local/cloud model you point it at, an honest scorecard out (unpriced never reads as free).

Or install everything as a plugin: this repo hosts one plugin (the authoring skill · the nika-author subagent · three commands · a check-on-edit hook · the read-only MCP oracle) for three ecosystems:

codex plugin marketplace add supernovae-st/nika-agents && codex plugin add nika@nika
claude plugin marketplace add supernovae-st/nika-agents && claude plugin install nika@nika
# Cursor: search "nika" in Settings → Plugins · one Add installs the bundle

Agents discover a working prompt chain once; Nika keeps it as a file your team can check, run and replay.

Send us a workflow

Do you repeat an AI task every week, in ChatGPT, Claude, Cursor, Codex, or scripts? That loop is exactly what a workflow keeps:

Describe yours at nika.sh/convert or open a "convert my workflow" issue. We convert the best ones into runnable .nika.yaml examples, credited to you.

Editor support

This repo is the engine. It ships the language server (nika lsp, over stdio). The VS Code / Cursor / Windsurf / VSCodium extension lives in its own repo and is published as supernovae.nika-lang (and on Open VSX for Cursor / Windsurf / VSCodium):

• Install it from your editor's marketplace. It auto-downloads the matching nika release binary on first use (or reuses the nika already on your PATH).
• Source + issues: supernovae-st/nika-vscode.
• Any other editor: nika lsp speaks LSP over stdio. Wire it into any LSP client.

The constellation

This repo is the engine. Everything around it lives in its own repo, each with one job:

Examples live right here: examples/, the embedded gallery (nika examples list shows the same shelf from the binary).

Building Nika? The engine is crafted under a strict workspace discipline: context-window-sized crates, a per-crate admission checklist, zero .unwrap() in src/ (CI-enforced), downward-only layering. The design lives in docs/architecture/ and the decisions in docs/adr/; the roadmap is in ROADMAP.md.

License

The engine is AGPL-3.0-or-later (see LICENSE): modify it and run it as a hosted service, and users of that service get the source. The spec is Apache-2.0, maximally permissive for a standard.

A commercial license (Grafana model) is available for organizations that can't accept AGPL's network clause. Contact contact@supernovae.studio. Security reports: security@supernovae.studio.

© 2024–2026 SuperNovae Studio · 🦋 Nika, the butterfly on the SuperNovae flag. Prompt once. Run forever.