Bricolage
An agentic system that turns free-tier cloud resources into working outcomes. It orchestrates six specialist agents to plan, acquire, and assemble free services into deployable applications.
README
Bricolage
An agentic system that turns free-tier resources into working outcomes. Deployable on OpenClaw, ZeroClaw, Claude Code, or any MCP-compatible runtime. Starts with cloud compute; extends into any domain you bring it.
Free-tier services from GitHub, Vercel, Supabase, OpenAI, Cloudflare, and hundreds of others give you enough capacity to build, deploy, and run applications without paying for infrastructure. Bricolage is six specialist agents that know which resources exist, how to claim them, and how to combine them into something that runs.
The name is from Lévi-Strauss: bricolage is the craft of making do with what's at hand. Scan the available parts, pick the ones that fit the goal, assemble. That is what these agents do.
Bricolage is a domain-agnostic protocol for assembling free resources into outcomes. It ships a compute domain with 10 fully verified resources (detailed guides, token validation regex, verified 2026-03-23) plus 401 bulk-catalog entries surveyed the same day for discoverability. Other domains — education, health, finance, research — can be scaffolded on the fly by the discovery agent when a user brings a goal the existing domains don't cover.
Naming note. This project used to be called Universal Basic Compute (UBC). The name implied a kind of commons, but most free tiers are customer-acquisition spend, not public goods. Bricolage is the more honest name. MCP tool names still carry the
ubc_prefix for backward compatibility — the prefix is a stable token, not a claim.
Quick Start
Bricolage exposes a standards-compliant MCP server over stdio. Any MCP-compatible runtime works. See INTEGRATE.md for the full wire-up.
Via Claude Code (one of several supported runtimes):
/plugin install chiKeka/bricolage
Via any other MCP runtime (OpenClaw, ZeroClaw, Claude Desktop, Cursor, Codex, custom):
git clone https://github.com/chiKeka/bricolage.git
cd bricolage
npm install
# then point your runtime at `npx tsx tools/src/index.ts` — see INTEGRATE.md §2
Then in your runtime:
/setup
The master agent takes over from there.
What's Inside
6 Agents
| Agent | What it does |
|---|---|
| master | Your entry point. Takes the user's goal, figures out which domain it belongs to, delegates to the right specialist. |
| planner | Reads a domain's catalog and picks the minimal set of resources to achieve the goal. Outputs a structured plan. |
| guide | Walks users through acquiring access one step at a time. Creating accounts, getting API keys, enrolling in programmes, storing tokens. |
| assembler | Takes the plan plus acquired access and builds the outcome. For compute: scaffolds the project, wires the services, deploys. |
| discovery | Researches new domains and resources. When a goal doesn't fit any existing domain, discovery scaffolds one, populates its catalog, and creates starter patterns. |
| infra | Helps deploy Bricolage's own agent runtime (OpenClaw, MiniClaw, PicoClaw) onto user-provisioned free compute. Compute-domain only. |
411 Free Resources (compute domain)
Two-tier catalog:
- 10 detailed guides with full signup instructions, credential capture walkthroughs, and validation regex. GitHub, Vercel, Supabase, OpenAI, Cloudflare, Netlify, Render, Neon, Resend, Upstash.
- 408 bulk entries in
domains/compute/resources/catalog.yamlcovering AI, cloud infra, storage, automation, APIs, and developer tools. (7 of these share a name with a detailed guide; the loader deduplicates so the merged total is 411.)
Every resource is classified by free-tier durability so the planner can prefer tiers that will outlast the project:
| Classification | Meaning | Example |
|---|---|---|
cross_subsidy |
Subsidised by a paid business elsewhere. Durable. | Cloudflare, GitHub |
public_good |
Funded by a non-commercial institution. Gated but durable. | GitHub Education, .edu datasets |
cac_funded |
Customer acquisition. Decays when unit economics change. | Most SaaS free tiers |
community |
Volunteer or donation-funded. Availability tracks donations. | Many OSS hosting projects |
trial |
Time-limited. Flag, don't rely. | OpenAI's $5 starter credit |
unknown |
Not yet classified. Treat as cac_funded. |
Most bulk-catalog entries |
Every resource also carries a verified_at date. Resources aged past 90 days get a warn tag; past 180 days, stale. Pattern planning can exclude stale entries with exclude_stale=true. Free tiers are decaying assets - Heroku-in-2022 is the canonical case - so the catalog treats them that way.
The compute catalog was last verified 2026-03-23.
5 Assembly Patterns (compute domain)
A pattern is a known-good combination of resources that produces an outcome. Previously called "recipes."
| Pattern | Resources Used | What You Get |
|---|---|---|
| blog-ai | GitHub + Vercel + Supabase + OpenAI | Next.js blog with AI summarisation and smart search |
| portfolio | GitHub + Vercel + Supabase | Personal site with contact form and analytics |
| saas-starter | GitHub + Vercel + Supabase + Cloudflare | Full-stack app with auth, database, CDN |
| ai-chatbot | GitHub + Vercel + Supabase + OpenAI | GPT chat with conversation memory |
| api-backend | GitHub + Supabase + Cloudflare | REST API on Cloudflare Workers with Postgres |
10 MCP Tools
| Tool | Description |
|---|---|
ubc_domains |
List all registered domains with trust level |
ubc_create_domain |
Scaffold a new domain (defaults to trust_level: user_scaffolded) |
ubc_catalog |
Browse resources by domain, category, or search. Returns staleness tag on every entry. |
ubc_resource_guide |
Full setup guide for one resource |
ubc_patterns |
List all assembly patterns for a domain |
ubc_pattern_detail |
Full details for one pattern |
ubc_status |
Check what's acquired, what's pending, which pattern is active |
ubc_update_status |
Update a resource's provisioning status |
ubc_store_access |
Store an access token. Validation regex enforced before encryption. |
ubc_get_access |
Retrieve stored tokens. Values masked unless reveal=true. Every reveal is audited. |
Six legacy aliases (ubc_service_guide, ubc_recipes, ubc_recipe_detail, ubc_store_credential, ubc_get_credentials) still work for backward compatibility but log a deprecation note in their description.
4 Slash Commands
| Command | What it does |
|---|---|
/setup |
Interactive onboarding. Checks state, asks about your goal, walks you through everything. |
/explore |
Browse the resource catalog. |
/build |
Assemble the outcome from acquired resources. |
/status |
See what's configured and what's missing. |
Protocol
The protocol has three schema files in /protocol/:
domain.schema.yaml- how to register a new domainresource.schema.yaml- what a resource looks likepattern.schema.yaml- what an assembly pattern looks like
Executable counterparts live in tools/src/schemas.ts (Zod). Every domain, resource, and pattern is validated at load time. Malformed content is rejected with a structured error on stderr rather than silently accepted. This is what keeps the protocol load-bearing instead of decorative.
Trust Level
Every domain declares a trust_level:
blessed— ships in the main repo; reviewed by maintainersuser_scaffolded- created locally by the discovery agent; not reviewedexternal- added from an outside source (future: signed registry)
User-scaffolded domains still work, but agents should warn users before treating them as authoritative. This is the minimum governance the protocol needs to avoid becoming a free-for-all.
How It Works
You: "I want to build a blog with AI features"
master:
"Let me plan that out for you."
→ delegates to planner
planner:
Reads domains/compute/resources/ and patterns/.
Returns: "Use the blog-ai pattern. GitHub + Vercel + Supabase + OpenAI."
master:
"Let's set up each service. Starting with GitHub..."
→ delegates to guide
guide:
"Do you have a GitHub account?"
"No? No problem. Go to github.com/signup..."
"Now let's get your API token..."
Walks through each resource one at a time. Stores each token.
master:
"All resources are ready. Let me build your blog."
→ delegates to assembler
assembler:
Scaffolds the project, wires the services, deploys.
"Your blog is live at https://your-blog.vercel.app"
Works With Other Agents Too
The MCP server is platform-agnostic. Connect it to OpenClaw, Cursor, Codex, or any MCP client:
{
"mcpServers": {
"bricolage": {
"command": "npx",
"args": ["tsx", "tools/src/index.ts"],
"cwd": "/path/to/bricolage"
}
}
}
Extending
Add a resource - Create a YAML in domains/compute/resources/ (detailed) or append to domains/compute/resources/catalog.yaml (bulk). Include free_tier_type and verified_at.
Add a pattern - Create a YAML in domains/compute/patterns/ following protocol/pattern.schema.yaml.
Add a domain - Tell the master agent what you want to accomplish. If no existing domain fits, the discovery agent will scaffold one for you. Or call ubc_create_domain directly.
Add an agent - Create a .md file in agents/ with frontmatter (name, description, model, tools).
Update the catalog - Tell the discovery agent: "Search for new free-tier AI services and update the catalog."
Repository Structure
.claude-plugin/plugin.json Claude Code plugin manifest (one of several runtimes)
.mcp.json MCP server config for Claude Code / Desktop / Cursor / Codex
INTEGRATE.md Wire-up guide for any MCP-compatible runtime
commands/ 4 slash commands
agents/ 6 agent definitions
protocol/ 3 schema reference files
domains/compute/ compute domain (411 resources, 5 patterns)
domains/<other>/ additional domains scaffolded on the fly
tools/ MCP server (Zod validation, audit log, AES-256-GCM encryption)
tools/test/ npm test — schema, catalog, access, MCP handshake, race contract
skills/ Runtime skill manifests (e.g. OpenClaw)
CLAUDE.md Context file for Claude-family runtimes
SECURITY.md Trust model, custody, audit log
Security
Tokens are encrypted with AES-256-GCM using a machine-local key in .ubc/.key (mode 0600).
Important: encryption defends against offline file-copy. It does not defend against an agent that is already running with this MCP server. Any such agent can call ubc_get_access with reveal=true and see every token in plaintext. That is inherent to MCP. Reveal events are logged to .ubc/audit.log so you can see what looked at what. See SECURITY.md for the full threat model.
Contributing
Pull requests welcome. Good places to start:
- Add resources - Find a free tier we're missing? Add it to
domains/compute/resources/catalog.yamlwithfree_tier_typeandverified_at. - Add detailed guides - Write a step-by-step YAML for a resource in
domains/compute/resources/. - Add patterns - Create a new assembly blueprint in
domains/compute/patterns/. - Classify durability - Help backfill
free_tier_typeacross the 408 bulk entries (currently most areunknown). - Refresh verifications - Walk through a resource, confirm the free tier still matches our entry, bump
verified_at.
License
MIT
Free compute is everywhere. It just needs to be assembled.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.