GitHub - serverless-dna/tendril

A self-extending agentic sandbox that demonstrates the Agent Capability pattern — where the model discovers, builds, and reuses tools autonomously across sessions.

Built with AWS Strands Agents SDK and Tauri.

What it does

You ask Tendril to do something. It checks its capability registry. If a tool exists, it uses it. If not, it writes one, registers it, and executes it — all without asking. Next time you need the same thing, the tool is already there.

You: "fetch the top stories from Hacker News"

Tendril:
  → searchCapabilities("fetch url hacker news")    # nothing found
  → registerCapability(fetch_url, code)             # builds a tool
  → execute("fetch_url", {url: "https://..."})      # runs it by name
  → "Here are the top stories: ..."

You: "now fetch Lobsters and compare"

Tendril:
  → listCapabilities()                              # found: fetch_url ✓
  → execute("fetch_url", {url: "https://lobste.rs"})# runs it — no rebuild

The registry grows with use. Every session is smarter than the last.

The Agent Loop

The core of Tendril is a Strands agent with three bootstrap tools. That's it — three tools to rule them all.

Where it lives

tendril-agent/src/
├── agent.ts              ← Agent configuration (Strands model + tools)
├── index.ts              ← Orchestrator — wires loop to transport
├── loop/                 ← The agentic loop
│   ├── tools.ts          ← 4 bootstrap tools in cycle order
│   ├── prompt.ts         ← System prompt (autonomous behaviour rules)
│   ├── registry.ts       ← Capability registry (index.json CRUD)
│   └── sandbox.ts        ← Deno subprocess execution with sandboxing
└── transport/            ← Conversation framing + stream observation
    ├── protocol.ts       ← ACP JSON-RPC over stdio
    ├── stream.ts         ← SDK events → loop phases (think/act/observe)
    └── errors.ts         ← Provider error classification

How it works

agent.ts — Creates the Strands agent with a Bedrock model and three tools:

import { Agent } from '@strands-agents/sdk';
import { BedrockModel } from '@strands-agents/sdk/models/bedrock';

const agent = new Agent({
  model: new BedrockModel({ modelId: '...', region: '...' }),
  systemPrompt: TENDRIL_SYSTEM_PROMPT(workspacePath),
  printer: nullPrinter,   // suppress SDK stdout — we own the protocol
  tools: [
    listCapabilities(registry),
    registerCapability(registry),
    executeCode(registry, workspacePath, config),
  ],
});

index.ts — Observes the agentic loop and bridges it to the ACP protocol:

// The agentic loop runs inside agent.stream().
// We observe each phase and forward to the UI.
for await (const event of agent.stream(userText)) {
  const { phase, event: e } = classifyEvent(event);
  switch (phase) {
    case 'think':   emitUpdate(handleThink(e));    break;  // text delta
    case 'act':     emitUpdate(handleAct(e));      break;  // tool call
    case 'observe': emitUpdate(handleObserve(e));  break;  // tool result
  }
}

loop/prompt.ts — The system prompt that makes the agent autonomous:

BEFORE acting on any request:
1. Call searchCapabilities(query) to check if a relevant tool exists
2. If found: call loadTool(name) then execute(code, args)
3. If NOT found: you MUST build the tool yourself.

RULES:
- NEVER ask "would you like me to create a tool?" — just build it.
- If a tool fails, read the error, fix the code, and retry.
- NEVER answer from training data when a tool could get live information.

The "too many tools" solution

Most agent frameworks give the model a big bag of tools and hope it picks the right one. Tendril inverts this — the model always sees exactly three tools. It searches a registry, builds what it needs, and the registry grows over time. The tool surface never changes; the capabilities do.

Architecture

┌─────────────────────────────────────────┐
│ Tauri Shell (Rust)                      │
│                                         │
│  ACP Host ──stdin/stdout──► Agent       │
│  (acp.rs)    NDJSON        (Node.js SEA)│
│     │                         │         │
│  Events  ◄── session/update ──┘         │
│  (events.rs)                            │
│     │                                   │
│  Tauri Events ──►  React Frontend       │
│                    (TailwindCSS v4)     │
└─────────────────────────────────────────┘

Agent internals:
  Strands SDK ── BedrockModel ── Claude
       │
  4 bootstrap tools
       │
  ┌────┴────┐
  │ Registry │ ←→ index.json + tools/*.ts
  └─────────┘
       │
  ┌────┴────┐
  │ Sandbox  │ ←→ Deno subprocess (scoped permissions)
  └─────────┘

Communication: JSON-RPC 2.0 over NDJSON (newline-delimited JSON) on stdin/stdout. The agent is a standalone process — the Tauri host spawns it as a sidecar.

Protocol: Implements the Agent Integrator Specification (ACP) — the same protocol used by Claude Code and similar agent hosts.

Tech Stack

Component	Technology
Desktop shell	Tauri 2.x (Rust)
Frontend	React 18 + TailwindCSS v4
Agent	TypeScript (Node.js SEA binary)
Agent framework	@strands-agents/sdk
Inference	AWS Bedrock (Claude via Strands BedrockModel)
Code sandbox	Deno (bundled, subprocess with permission flags)
Protocol	JSON-RPC 2.0 / NDJSON over stdio

Prerequisites

Node.js 22+ (for building the agent)
Rust toolchain (for Tauri)
AWS credentials configured for Bedrock access (~/.aws/credentials)

Quick Start

git clone https://github.com/serverless-dna/tendril.git
cd tendril
make dev

This will:

Install dependencies (npm install for agent and UI)
Build the agent (esbuild bundle)
Download Deno (bundled as sidecar)
Create sidecar shims with platform triple
Launch Tauri dev mode

On first launch, pick a workspace folder. Configure your AWS profile and model in Settings.

Configuration

All settings live at ~/.tendril/config.json:

{
  "workspace": "/Users/you/tendril-workspace",
  "model": {
    "provider": "bedrock",
    "modelId": "us.anthropic.claude-sonnet-4-5-20250514",
    "region": "us-east-1",
    "profile": "your-aws-profile"
  },
  "sandbox": {
    "denoPath": "deno",
    "timeoutMs": 45000,
    "allowedDomains": []
  },
  "agent": {
    "maxTurns": 100
  }
}

allowedDomains: empty = unrestricted network. Set ["api.example.com"] to restrict.

Capability Registry

Capabilities are stored in the workspace as plain files:

~/tendril-workspace/
  index.json          ← registry (name, triggers, suppression rules)
  tools/
    fetch_url.ts      ← tool implementation (TypeScript, runs in Deno)
    summarize_text.ts
    parse_json.ts

Each capability has:

name: snake_case identifier
capability: one-sentence description
triggers: conversational signals that should invoke it
suppression: conditions that prevent invocation

The model writes these definitions. You can inspect, edit, or delete them — they're just files.

Makefile Targets

make dev       Build agent + sidecars, launch Tauri dev
make build     Build agent (esbuild bundle)
make test      Run agent tests (vitest)
make lint      tsc --noEmit + cargo clippy
make fmt       cargo fmt --check
make check     Full quality gate (fmt + lint + test)
make release   Quality gate + cargo tauri build
make clean     Remove all build artifacts

Project Structure

tendril/
  tendril-agent/        TypeScript Strands sidecar
    src/                Agent source (see "The Agent Loop" above)
    tests/              vitest tests
    package.json
    sea-config.json     Node.js SEA build config

  tendril-ui/           Tauri + React desktop app
    src/                React components + hooks
    src-tauri/          Rust backend (ACP host, event forwarding)
    package.json

  docs/                 Specs and reference implementations
  specs/                Feature specifications and plans
  Makefile

License

MIT

推荐订阅源

Hacker News: Front Page