Public knowledge MCP endpoint

Hive can publish a small, anonymous MCP endpoint for operational knowledge that other agents can read without joining the hive. Use it when you want a troubleshooting assistant, local coding agent, or support script to reason from the hive’s accumulated fixes, regressions, and integration notes instead of from model training data.

The endpoint is POST /mcp/knowledge on the spoke. It is available in Hive v5.134.0 and later, shipped for hivecommons/hive#10615 by hivecommons/hive#10621.

Note: this is currently controlled by environment variables. There is no dashboard toggle yet; a separate follow-up is tracking that operator UI.

Turn it on or off

Variable	Default	Effect
`HIVE_PUBLIC_KNOWLEDGE`	disabled	Set to `1`, `true`, `yes`, or `on` to enable `POST /mcp/knowledge`. Any other value makes the path return `404`, even to the owner.
`HIVE_PUBLIC_KNOWLEDGE_TAGS`	all public-type facts	Optional comma-separated tag allow-list. When set, facts carrying at least listed tag are served.

Hive reads both variables on every request, so an owner can close the surface instantly by unsetting or changing HIVE_PUBLIC_KNOWLEDGE; a spoke restart is not required.

Example Kubernetes patch:

kubectl -n hive set env deploy/hive HIVE_PUBLIC_KNOWLEDGE=1
kubectl -n hive set env deploy/hive HIVE_PUBLIC_KNOWLEDGE_TAGS=troubleshooting,linux

To close it:

kubectl -n hive set env deploy/hive HIVE_PUBLIC_KNOWLEDGE-

What public means operational fact types are eligible:

pattern
gotcha
regression
test_scaffold
integration
coverage_rule
general

Ideation and governance facts are never served: idea, vision, constitution, requirement, constraint, stakeholder, and decision. Requests for a non-public slug return “not found” rather than revealing that a private fact exists.

Public facts include slug, title, type, body, tags, related, and scored confidence. Hive strips source links, authors, usage counters, confidence reasoning, and lifecycle phase. A fact’s related list is also filtered to slugs that are public under the same rules.

MCP protocol and limits

The endpoint speaks MCP protocol 2025-06-18 over Streamable HTTP with plain JSON responses.

Limit	Value
HTTP method	`POST`
Request body	64 KiB maximum
JSON-RPC batching	Not supported
Search results	`knowledge_search` returns at most 50 facts
Writes	No write tools or mutation path

The exposed tools are marked with readOnlyHint: true.

Tool	Arguments	Returns
`knowledge_search`	`query` required; optional `type`, `limit`	JSON `{query, count, results[]}`
`knowledge_get`	`slug` required	public fact, or an MCP tool error saying the fact was not found
`knowledge_export`	none	Markdown export of the public base, grouped by type, with `_meta.etag` and `_meta.facts`

Connect agents

Replace https://hive.example.org with the public URL for your spoke.

Goose

Add a Streamable HTTP extension to ~/.config/goose/config.yaml:

extensions:
  project-hive:
    type: streamable_http
    uri: https://hive.example.org/mcp/knowledge
    enabled: true

Claude Desktop and Claude Code

Claude Code can register the HTTP MCP server directly:

claude mcp add --transport http project-hive https://hive.example.org/mcp/knowledge

For Claude Desktop, add the same URL as an HTTP MCP server in the app’s MCP server configuration and restart the app so it reloads the configuration.

Copilot CLI

Add the server to ~/.copilot/mcp-config.json:

{
  "mcpServers": {
    "project-hive": {
      "type": "http",
      "url": "https://hive.example.org/mcp/knowledge"
    }
  }
}

Raw curl

Initialize the MCP session:

curl -s https://hive.example.org/mcp/knowledge \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1.0.0"}}}'

List the tools:

curl -s https://hive.example.org/mcp/knowledge \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'

Search public knowledge:

curl -s https://hive.example.org/mcp/knowledge \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"knowledge_search","arguments":{"query":"bluetooth regression","limit":5}}}'

Verify it is working

With HIVE_PUBLIC_KNOWLEDGE unset, POST /mcp/knowledge should return 404.
Set HIVE_PUBLIC_KNOWLEDGE=1.
Run the tools/list curl command above and confirm the three knowledge tools are present.
Run a knowledge_search for a tag or phrase you know exists.
If HIVE_PUBLIC_KNOWLEDGE_TAGS is set, search for an unlisted tag and confirm those facts are absent.
Unset HIVE_PUBLIC_KNOWLEDGE and confirm the endpoint returns 404 again without restarting the spoke.

Security posture

This surface is anonymous but read-only. It exposes no create, update, delete, import, promotion, vault, or GitHub write path, and it does not expose source attribution or author data. Hosted hub front doors can pass /mcp/knowledge through without a browser session, but the spoke still re-checks HIVE_PUBLIC_KNOWLEDGE and returns 404 when the owner switch is off.

Treat the response body as public information. Before enabling the endpoint, consider tag scoping with HIVE_PUBLIC_KNOWLEDGE_TAGS, review which operational facts should be public, and use a reverse-proxy allow-list if specific agent networks should reach the endpoint.