🌳 Grove

Production-grade AI agents. OTP-style supervision, compile-time optimization, live time-travel inspection, behaviour diff. Built for engineers who need agents that don't fall over in prod.

bun add @vyntral/grove-core @vyntral/grove-runtime

---

Why Grove

96% of organizations are running AI agents. Only 1 in 9 has them in production at scale.

The problem isn't the models. It's that today's agent frameworks were built as scripts: brittle, opaque, expensive, and impossible to debug once they're live. Most failures originate at the boundaries between processes — exactly where current tooling has nothing to say.

Grove takes the lessons from 30 years of Erlang/OTP fault tolerance, adds a workflow compiler, and wraps both in a developer experience that feels like Next.js. It is the runtime layer agents have been missing.

The three superpowers

1. Supervised processes, not scripts

Define your agent topology. The runtime spawns each agent as a supervised process. When something crashes — and it will — the supervisor restarts it under your declared strategy (one_for_one, one_for_all, or rest_for_one), with crash-loop guards. Let it crash; let the system heal.

import { agent, supervise, tool } from '@vyntral/grove-core'
import { start } from '@vyntral/grove-runtime'

const research = agent({
  name: 'research',
  model: 'anthropic/claude-opus-4-7',
  tools: [search, summarize],
})

const tree = supervise({
  strategy: 'rest_for_one',
  children: [research, writer],
  restart: { intensity: 5, period: 60_000 },
})

const { handle } = await start(tree)
await handle.run({ topic: 'sparse attention' })

2. Compile-to-determinism + runtime cache

Most agent workflows have stable, deterministic paths — formatting, parsing, lookups, cached responses. Grove's compiler analyzes your topology, identifies which steps don't actually need an LLM, and projects the cost reduction. The runtime cache then makes the projection real: deterministic tool calls hit a SQLite-backed cache before invoking the model.

$ grove compile examples/research.ts
  total cost projection: $0.1440 → $0.0144 (10.0× cheaper)

$ bun packages/examples/src/persist.ts        # first run
run: 599ms
cache: { entries: 1, hits: 0, misses: 1, hitRate: 0 }

$ bun packages/examples/src/persist.ts        # second run (cached)
run: 99ms
cache: { entries: 1, hits: 1, misses: 0, hitRate: 1 }

The cache survives across processes — restart your dev server, redeploy your worker, the warm cache stays warm.

3. The Bench — live time-travel inspector

Every run is recorded automatically. Open the Bench (grove bench) and watch your agents execute in real time, scrub through any past session, click any step to see its inputs/outputs/cost, and see cache hits / saved $ live.

$ grove bench
✓ bench listening on http://localhost:4773

4. Hot reload, only what changed

Edit agent.ts, save, and only the children whose definitions actually changed restart. Siblings keep running, sessions keep their state.

$ grove run --watch agent.ts
✓ started agent.ts
  watching for changes — Ctrl+C to stop
→ hot-reloaded: research

5. Behaviour diff — catch regressions before merge

Declare eval cases, save a profile, diff across runs. CI exits non-zero on regression.

import { suite, evalCase, contains } from '@vyntral/grove-eval'

export const suite = suite('word-counter', [
  evalCase({
    id: 'short',
    input: 'How many words: "hello world"?',
    assertions: [contains('2')],
  }),
])

$ grove eval examples/eval-suite.ts
word-counter: 4 passed, 0 failed in 7569ms
  ✓ short                1772ms
  ✓ medium               1644ms
  ...

$ grove diff word-counter
  drift: 1
  regressed: 1
    regressed   short  — output does not contain "2"

6. Anthropic prompt caching, auto-on

Long system prompts are cached at the model layer for anthropic/* models — Grove writes the cache_control: ephemeral breakpoint for you. 90% input-token discount on cache reads, surfaced as prompt_cache events.

7. MCP-native — any tool from any MCP server, plug-and-play

import { mcpServer } from '@vyntral/grove-mcp'

const fs = await mcpServer({
  command: 'npx',
  args: ['-y', '@modelcontextprotocol/server-filesystem', process.cwd()],
})

const a = agent({
  name: 'librarian',
  model: 'anthropic/claude-haiku-4-5',
  tools: fs.tools, // MCP tools surface as Grove tools, with prefix
})

Spawned as real stdio child processes. Tool calls flow through Grove's recorder + cache transparently.

---

Install

# Bun (recommended)
curl -fsSL https://bun.sh/install | bash

# Then in your project:
bun add @vyntral/grove-core @vyntral/grove-runtime @vyntral/grove-compiler @vyntral/grove-cli

Grove targets Bun 1.3+ for native SQLite + native TypeScript. Node support tracked in #node-support.

Quickstart

$ grove init agent.ts
✓ created agent.ts
  run it with: grove run agent.ts

$ grove run agent.ts
→ running agent.ts
hello, world 🌳
✓ done in 102ms
  session: s_1778016019099_k8lnf5
  inspect: grove inspect s_1778016019099_k8lnf5

How it differs

	LangChain / CrewAI	DSPy / TextGrad	Replay / AgentOps	Grove
Supervised processes	–	–	–	✅
Restart strategies	–	–	–	✅
Compile-time optimization	–	prompt-only	–	workflow-level
Runtime cache	–	–	–	✅ persistent
Hot reload (per-child)	–	–	–	✅
Time-travel debug	–	–	✅	✅ + fork
Recorder DB built-in	–	–	–	✅
Zero-config local UI	–	–	–	✅ (Bench)
MCP-native	partial	–	–	✅ stdio adapter
One install	partial	–	–	✅

Backends

Grove ships with a mock backend so you can run all examples (and demo the runtime) without API keys. To talk to a real frontier model, plug in the AI SDK backend:

# Option 1 — Vercel AI Gateway (recommended; one key, any provider)
export AI_GATEWAY_API_KEY=sk_gateway_xxx
bun packages/examples/src/smoke-llm.ts

# Option 2 — direct provider (Anthropic only at the moment)
export ANTHROPIC_API_KEY=sk-ant-xxx
export GROVE_DIRECT_PROVIDER=1
bun packages/examples/src/smoke-llm.ts

The AI SDK backend addresses models with the canonical provider/model syntax and works with any provider on the Gateway:

• anthropic/claude-opus-4-7
• openai/gpt-5.5
• google/gemini-3.1-pro
• subq/subq-1m-preview
• … and any other AI Gateway-compatible model.

Tests

Two suites:

# Unit + integration against MockBackend — fast, no API key needed.
bun run test                    # 37 tests, ~1.5s

# Real LLM integration against Anthropic — ~15s, costs a few cents.
ANTHROPIC_API_KEY=sk-... GROVE_DIRECT_PROVIDER=1 bun run test:live   # 6 tests

The live suite covers AISDK roundtrip, typed tool calling, deterministic cache hits, Anthropic prompt caching (cache_read > 0 on the second call), streaming chunk emission, and MCP stdio end-to-end. Skipped automatically when the env var is missing, so contributors don't need a key. CI runs the unit suite on every push and the live suite daily + on manual dispatch.

Architecture

See docs/architecture.md for the runtime model, docs/spec.md for the API spec, and docs/manifesto.md for the design philosophy.

License

MIT © Grove contributors