Show HN: Heku – Dynamic MCP Tooling via JSON Configs

One server. Any API. Any LLM.

Your agent's tool manifest breaks around ten MCP servers. Every server you add fattens the manifest until the context fills up and the model starts forgetting which tools exist. heku is one MCP server that removes that ceiling: you describe each tool as a JSON config, and heku serves them lazily — the manifest stays a few hundred tokens whether you've loaded ten configs or two hundred, and the model pulls in only the tools it needs, when it needs them.

One server, any number of APIs, no context bloat. Agents can even write their own configs live from API docs.

npx @rapidthoughtlabs/heku start

** Website** ·

·

Launch post →·

Console

heku hub** console.rapidthoughtlabs.space** — hosted console you can point at any running heku instance. Connect, browse configs, chat with your tools, and inspect the system prompt — no local build needed.

** app.rapidthoughtlabs.space** —

heku hub, the online registry for browsing, installing, and publishing heku configs. Find community-built connectors for GitHub, Slack, Linear, and more — or publish your own.

Install (requires Node.js ≥ 20):

npx @rapidthoughtlabs/heku start

Create mcp-configs/mcp.github.json

:

{
  "id": "github-http",
  "name": "GitHub API",
  "description": "Manage GitHub repos, issues, and pull requests",
  "connector": {
    "type": "http",
    "base_url": "https://api.github.com",
    "auth": { "type": "bearer", "token_env": "GITHUB_TOKEN" }
  },
  "tools": [
    {
      "name": "list_repos",
      "description": "List repositories for the authenticated user",
      "method": "GET",
      "path": "/user/repos",
      "params": [
        { "name": "per_page", "type": "number", "required": false, "location": "query", "description": "Results per page" }
      ]
    }
  ]
}
heku auth setup github-http   # writes GITHUB_TOKEN to .env
heku start

Your LLM now has a github-http.list_repos

tool — and the manifest grew by only that one entry.

heku sits between your LLM and every API you've configured. Three ideas make it different from running a pile of separate MCP servers:

Lazy discovery. The manifest the model sees stays small no matter how many configs you load. Tools are surfaced on demand instead of dumped up front, so you never hit the ~ten-server context wall.Configs, not code. A tool is a JSON file — an endpoint, its params, and how to authenticate. No per-integration server to build, ship, or maintain.Self-managing. Point heku at API docs and it can author and edit its own configs through internal tools, then hot-reload them without a restart.

Tool names follow the pattern config_id.tool_name

— e.g. github-http.list_repos

, linear-graphql.create_issue

.

8 connector types— 4 standard (HTTP, GraphQL, gRPC, child-MCP) + 4 experimental (CLI, File, SQL, MongoDB)** Lazy tool discovery**— manifest stays a few hundred tokens regardless of how many configs are loaded** Hot-reload**— add or edit a config, tools update live without restart** Auto-discovery**— gRPC reflection, GraphQL introspection, and child MCP tool listing fill in tools automatically** Response stripping**— base64 blobs, null fields, oversized strings, and long arrays are trimmed before the model sees them, keeping context lean without losing dataBuilt-in console UI— React dashboard for chat, config editing, and registry browsing** heku hub**— publish and install community configs fromapp.rapidthoughtlabs.space** Auth handled**— bearer, basic, API key, and OAuth2 with.env

-based credential managementSelf-managing— the server can create and edit its own configs via internal tools

Each connector type wraps a different kind of backend as MCP tools. Full config schemas, field references, and examples live in ** connectors.md**.

Connector	Status	What it wraps	Tool discovery
http

graphql

grpc

.proto

)mcp

cli

file

sql

mongodb

Experimental connectors are functional, but their config schema and behaviour may change in future releases.

All credentials read from environment variables — heku auth setup

writes them to .env

, and nothing ever travels over the MCP protocol.

Type	Header
bearer
Authorization: Bearer {token}
basic
Authorization: Basic base64(user:token)
api_key
Custom header, e.g. X-API-Key
oauth2_static
Pre-acquired OAuth2 access token

{ "type": "bearer",       "token_env": "GITHUB_TOKEN" }
{ "type": "api_key",      "key_env": "MY_KEY", "header_name": "X-Api-Key" }
{ "type": "basic",        "username_env": "MY_USER", "token_env": "MY_PASS" }
{ "type": "oauth2_static","token_env": "MY_OAUTH_TOKEN" }
heku start [config-dir]      Start the MCP server (stdio by default)
                             Flags: --http, --port <n>, --debug

heku list [service]          List loaded configs + auth status
heku auth                    Check or set up credentials interactively
heku auth status             Show per-service auth health
heku auth setup [service]    Walk through env-var setup, write to .env

heku login                   Authenticate with the registry
heku logout                  Clear stored registry credentials

heku install <target>        Install a config from the registry
                             Target: @ns/slug or @ns/slug@version
heku uninstall <target>      Remove an installed registry config
heku publish [file]          Publish a local config to the registry
heku fork <namespace/slug>   Fork a published config into your namespace

heku discover                Scan Claude Desktop / Cursor for MCP servers
heku update                  Update heku to the latest version
heku help                    Show usage

Start with the console UI:

heku start --http --port 3456

Then open ** console.rapidthoughtlabs.space** and connect to

http://localhost:3456

.The dashboard is a React + Vite app — available hosted at ** console.rapidthoughtlabs.space** or embedded when you run

heku start --http

.Chat— test tools through a model of your choice (OpenAI, Together AI)** Configs**— visual editor for connector and tool definitions** Prompts**— inspect the system prompt layers and token counts** Registry**— browse, install, and publish configs** Auth**— credential status across all configs at a glance

** app.rapidthoughtlabs.space** is the default registry for sharing configs — browse community connectors, install with one command, and publish your own.

heku install @rtl/github
heku install @rtl/slack@1.2.0
heku publish mcp-configs/mcp.stripe-http.json

Use --registry <url>

to point at a self-hosted registry.

Drop heku.config.json

in your config directory:

{
  "log_level": "info",
  "rate_limits": {
    "github-http": { "requests_per_minute": 60 }
  },
  "self_config": true
}
git clone https://github.com/RapidThoughtLabs/heku
cd heku
npm install

npm run dev          # client (5173) + console server (3456) in parallel
npm run dev:mcp      # MCP stdio server only — for testing with Claude Desktop
npm run build        # bundle CLI to dist/cli.js via tsup
npm run typecheck    # tsc --noEmit
npm test             # vitest
src/             MCP server core (CLI, connectors, auth, , executor)
server/          Express backend for the console UI
client/          React + Vite dashboard
protos/          Example .proto files for gRPC connector testing
scripts/         Registry seed scripts
mcp-configs/     Local config files (gitignored)

TypeScript · Node.js (ESM) · @modelcontextprotocol/sdk

· Express · React 19 · Vite · Zustand · @grpc/grpc-js

· GraphQL · tsup · Vitest

Response stripping— every tool response is structurally trimmed before it reaches the model. Base64 blobs become size markers, null/empty fields are dropped, strings over 8 000 chars are head-truncated, and arrays over 100 items are capped. Structure-only — never decides which field matters.

• Renamed meta-tool namespace from mcp.one.*

toheku.*

across all connectors, prompts, and client code - Fixed deployed console manifest style switcher (settings API calls now use the correct bridge base URL)

• Fixed prompt page config catalog not refreshing when heku connects after page load
• Markdown rendering in the demo chat — assistant responses now render formatted text
• Config catalog descriptions now show in composed prompt preview; falls back to display name when description is absent
• Dual manifest preview in Prompts page — flat and namespaced styles with separate token counts
• heku server version now reads from package.json

at runtime in dev mode instead of showing0.0.0-dev

• Registry versioning overhaul — semver-based publish/install flow
• CLI registry commands: install

,uninstall

,fork

,publish

• Console registry browser tab

• SQL and MongoDB connector types (experimental)

• Config write lock — block LLM agents from mutating configs

• Hot-reload watcher improvements

@ruchitnannavare— creator & maintainer@SayanSwaroopROy

Apache License 2.0 — see LICENSE.

source & further reading

github.com — original article