Spikes

/ Spikes

The feedback loop for AI-assisted building

AI can build a prototype in an hour.

Turning feedback into action is still the slow part.

Quick Start · CLI Reference · Widget Docs · Hosted Dashboard · Self-Hosting

What Is This?

Spikes is a feedback tool for AI-assisted development. It lets reviewers leave targeted feedback directly on web pages — no screenshots, no "that button over there", no lost context.

Click any element. Rate it. Comment. Spikes captures the exact CSS selector, bounding box, and page context. Your AI agent gets structured JSON it can act on immediately.

No accounts required. No build step. Works on `file://`, localhost, anywhere.

Quick Start

1. Install the CLI

curl -fsSL https://spikes.sh/install.sh | sh

Or: cargo install spikes

2. Initialize your project

spikes init                     # Defaults to hosted spikes.sh

spikes init --self-host # Opt out — scaffold a self-hosted config instead

spikes init creates .spikes/config.toml with a [remote] section pointing at https://spikes.sh by default. Pass --self-host (or answer s at the prompt) to skip the hosted defaults.

3. Add the widget to your HTML

spikes inject ./mockups/        # Injects widget script tag (uses configured endpoint)

spikes serve # http://localhost:3847

4. Collect and use feedback

spikes list                     # See all feedback

spikes list --json # Feed to your agent spikes list --rating no # Find problems spikes hotspots # Elements with most feedback spikes resolve # Mark items done

CLI Commands

All commands support --json for scripting. See full CLI reference.

Viewing spikes for a hosted project:

The hosted dashboard at [https://spikes.sh/dashboard](https://spikes.sh/dashboard) lists every project you own and lets you drill into individual spikes (filter by page, rating, resolved, toggle resolved inline). Sign in directly in your browser: click Sign in, open the verification page, and confirm via the magic link sent to your email — no CLI required.

You can also hit the JSON API directly with your bearer token ($SPIKES_TOKEN from spikes login):

# GET /me/projects

List all projects you own (with spike_count + last_activity)

curl -H "Authorization: Bearer $SPIKES_TOKEN" \ https://spikes.sh/me/projects

GET /me/projects/:key/spikes

List spikes for one of your projects (paginated, filterable: page, per_page,

filter_page, filter_rating, filter_resolved)

curl -H "Authorization: Bearer $SPIKES_TOKEN" \ "https://spikes.sh/me/projects/my-project/spikes?page=1&per_page=50"

PATCH /me/projects/:key/spikes/:id

Toggle the resolved flag on one spike

curl -X PATCH -H "Authorization: Bearer $SPIKES_TOKEN" \ -H "Content-Type: application/json" \ -d '{"resolved": true}' \ https://spikes.sh/me/projects/my-project/spikes/

POST /projects

Create a new project (user bearer only)

curl -X POST -H "Authorization: Bearer $SPIKES_TOKEN" \ -H "Content-Type: application/json" \ -d '{"key": "my-project", "allowed_origins": ["https://example.com"]}' \ https://spikes.sh/projects

Single-line variant (copy-paste friendly)

curl -H "Authorization: Bearer $SPIKES_TOKEN" https://spikes.sh/me/projects

All /me/* endpoints are user-scoped: you only ever see projects and spikes you own. Cross-tenant requests return 404 PROJECT_NOT_FOUND (enumeration-proof).

Security note: API keys (sk_spikes_*) and the admin SPIKES_TOKEN are not accepted on /me/* endpoints or POST /projects — those require a user bearer token from spikes login. The POST /spikes collect endpoint enforces a per-project origin allowlist and per-project rate limits; rejected widget requests surface as a small red #spikes-error-dot indicator on the widget button (404 / 403 / 429).

AI Agent Integration

Spikes speaks agent natively. Your AI coding assistant can read, write, and manage feedback without ever leaving its workflow.

npx spikes-mcp            # Zero-install MCP server — just works

or: spikes mcp serve # If you have the CLI installed

MCP Server — 16 Tools

spikes mcp serve starts a Model Context Protocol server that exposes 16 tools:

Supports stdio and HTTP transports, local and remote data modes:

spikes mcp serve                              # stdio, local JSONL

spikes mcp serve --remote # stdio, hosted API spikes mcp serve --transport http --port 3848 # HTTP for sandboxed agents

API Keys

Agents get their own identity. No email, no magic link, no human step:

spikes auth create-key --name "my-agent"   # → sk_spikes_...

spikes auth list-keys # See all keys spikes auth revoke-key # Revoke

Keys support read/write/full scopes and optional budget caps.

Agent-Tier Billing

Consumption-based pricing for agent-scale usage. Pay per spike, not per seat:

spikes usage    # See current spend and limits

Budget enforcement returns 429 BUDGET_EXCEEDED when caps are hit — agents can check before they burn.

Context Exports

Export structured markdown optimized for agent consumption:

spikes export --format cursor-context > cursor-feedback.md

spikes export --format claude-context > claude-feedback.md

Discovery

• [llms.txt](https://spikes.sh/llms.txt) — All 9 MCP tools, parameters, agent quickstart
• [agents.md](https://spikes.sh/agents.md) — Machine-readable landing page for agents
• [Smithery](https://smithery.ai) — Listed in the MCP server registry
• spikes mcp install — Generates config for Claude Desktop / Cursor

Full details: llms.txt · agents.md

GitHub Action

Gate CI on feedback quality. The spikes-action fails builds when unresolved negative feedback exceeds your threshold.

name: CI

on: [push, pull_request]

jobs: feedback-gate: runs-on: ubuntu-latest steps:

• uses: actions/checkout@v4
• uses: bierlingm/spikes/action@v0.3.1 with: threshold: 0 # Fail if any blocking spikes ignore-paths: "" # Optional: pages to ignore require-resolution: false

See action/README.md for full documentation.

Widget Attributes

Simple setup — just add your project key and you're ready to collect feedback:

<script src="https://spikes.sh/spikes.js" data-project="my-app"></script>

When you set data-project, feedback syncs automatically to https://spikes.sh/spikes. No need to configure an endpoint unless you're self-hosting.

Full configuration — all available attributes:

<script src="https://spikes.sh/spikes.js"

data-project="my-app" data-position="bottom-right" data-color="#e74c3c" data-theme="dark" data-reviewer="Pat" data-collect-email="true" data-admin="true">

Self-hosting — point to your own backend:

<script src="/spikes.js"

data-project="my-app" data-endpoint="https://my-worker.workers.dev/spikes">

See Widget Attributes Reference for complete documentation.

Architecture

Spikes has three components that work together or standalone:

┌─────────────┐     ┌─────────────┐     ┌─────────────┐

│ CLI │────▶│ Widget │◄────│ Worker │ │ (Rust) │ │ (Vanilla │ │ (Cloudflare │ │ │ │ JS) │ │ + D1) │ └─────────────┘ └─────────────┘ └─────────────┘ │ │ │ spikes.sh (hosted) │ └────────────────────────────────────────────┘

CLI — Rust binary for local development, spike management, and deployment. Stores spikes in ~/.local/share/spikes/.

Widget — 14KB gzipped vanilla JS. Captures element selectors, bounding boxes, ratings, and comments. Works offline via localStorage.

Worker — Optional Cloudflare Worker + D1 backend for multi-reviewer sync, sharing, and hosted deployments. Lives in spikes-hosted/.

Development

CLI

cd cli

cargo build --release cargo test # 160+ tests cargo run -- --help

Widget

cd widget

Edit src/spikes.js

Test by running: spikes serve from the project root

Widget Regression Tests

CI suite that catches the silent-data-loss bug class (e.g., widget loaded on a non-allowe