mcphubEvery MCP server. One hub. The right tool on demand.

Connect each harness once. In lazy mode, agents can match the current task against your live catalog — including unpinned tools — then call the right capability through one gateway while mcphub keeps 15 agent configs in sync.

$ brew install abdul-hamid-achik/tap/mcphub

Get started

What is mcphub?

GitHub

Resolve capabilities by intent

Add natural-language use_when hints once. Search and resolve index them with live tool metadata, so an agent can find the right server even when it is not pinned.

Follow the discovery loop

One gateway, every tool

Agents connect to a single stdio server. mcphub proxies every downstream server behind it and namespaces tools as server__tool — names never collide.

How the gateway works

Sync 15 harnesses

mcphub sync writes the right config into Claude Code, Codex, opencode, Copilot CLI, Cursor, Claude Desktop and more. Dry-run by default, timestamped .bak, hand-added servers survive.

How sync stays safe

Token savings by design

Gateway mode loads one server instead of a dozen tool lists. expose: lazy goes further — eight meta-tools advertised, everything else discovered on demand.

Lazy mode & pinning

Local intelligence

Every proxied call lands in a local SQLite ledger. mcphub stats shows calls, errors, latency and estimated token cost per server and per tool.

Read your stats

Lossless big results

Oversized responses are spooled locally for 24 hours and replaced with a compact receipt — agents page back the exact bytes with mcphub_get_result. Nothing truncated.

Bounded results

Capability routing

Unpinned does not mean unavailable

In lazy mode, mcphub keeps the catalog out of the prompt but not out of reach. Tool metadata and your use_when hints stay searchable, so an agent can resolve the current task, inspect the match, and call it through the gateway.

example catalog routesresolve → describe → call

current task
“Fetch a public HTTP response as Markdown for source-backed research”
matched signals
webpageHTTPMarkdownresearch
recommended capabilityhitspechitspec__fetchBounded HTTP response → inline Markdown
current task
“Investigate evidence across code, CLI output, and databases”
matched signals
evidencecase filetool routingverification
recommended capabilitycortexcortex__investigateBounded investigation with retained evidence
current task
“Inspect the repository plan before developing a feature”
matched signals
repositoryplanconvergenceread-only
recommended capabilitybobbob__planDeterministic plan and digest before edits

See it run

Real output, not a mockup

One YAML file in, twelve synced harnesses out — and a local ledger of what every server actually costs you.

$ mcphub sync
» claude  (claude, gateway) → ~/.claude.json
    update   mcphub
» codex  (codex, gateway) → ~/.codex/config.toml
    update   mcphub
» copilot  (copilot, gateway) → ~/.copilot/mcp-config.json
    update   mcphub
» crush  (crush, gateway) → ~/.config/crush/crush.json
    update   mcphub
» hermes  (hermes, gateway) → ~/.hermes/config.yaml
    update   mcphub
» opencode  (opencode, gateway) → ~/.config/opencode/opencode.json
    update   mcphub

Dry run. Re-run with --write to apply (a .bak is saved first).

How it works

Three commands, and you never hand-edit a config again

1mcphub init
Write mcphub.yaml once — or import the servers your agents already declare with --from-agents.
2mcphub sync --write
Push the right config into every harness. Dry-run by default, a timestamped .bak before any write.
3mcphub mcp serve
Your agent gets one gateway. Lazy mode resolves unpinned tools from use_when hints and live metadata, then records every call locally.

Syncs into 12 harnesses

Claude Code
opencode
Codex
Copilot CLI
Qwen Code
Gemini CLI
Kilo Code
Kimi Code CLI
Crush
Forge
Hermes
local-agent

Give every agent the whole catalog — without loading all of it.

one registry → the right capability → one gateway

Build your hub GitHub