Connect with MCP

Connect an Agent tool client to the same data. MCP is not the full HTTP API.

MCP is the tool set an Agent calls explicitly. It talks to the same Server and the same data. It is not the full HTTP API.

Automatic recall, capture, and flush still belong to a host hook or framework adapter. MCP only exposes operations you call yourself.

Field names and error codes follow the latest published official docs. This page only helps you connect /mcp and see what it can do.

Default URL

http://127.0.0.1:8000/mcp

This is Streamable HTTP. The default Server listens on loopback with Bearer auth off. If the Server is down, the client should fail to connect. It should not show an empty tool set that looks healthy.

Minimum client config

Merge this into the Agent MCP config. Do not write a machine-local filesystem path, and do not put a token in the URL.

{
  "mcpServers": {
    "powercontext": {
      "type": "http",
      "url": "http://127.0.0.1:8000/mcp"
    }
  }
}

Some Agents spell the type as streamable-http. Point them at the same /mcp. Codex, Claude Code, and WorkBuddy use different field names and header helpers. Those details stay on their pages.

What it can and cannot do

MCP deliberately exposes a smaller surface than HTTP. Use this table to decide whether the next step can stay in the Agent, or whether you should switch to the Python Client or HTTP.

Can	Cannot
remember_memory, search_memory, exact get / list / revise / retire	prepare_context. A hook or adapter normally prepares the final PreparedContext
capture_content_source	flush_memory
Work Contract, Handoff, Task Outcome	Skill generate / export
Candidate list / get / approve / reject / revise	Experience operations
Read and write the same data as HTTP and the Python Client	Reproduce each integration’s automatic recall / capture / flush

The default MCP tool names in current tests are those operation ids. Read-only Handoff report tools appear only when that feature is on. Do not treat them as the default surface.

readOnlyHint and destructiveHint on MCP tools are confirmation hints for the Agent. They are not an authorization boundary. A proposing Agent can still receive approve tools unless the host withholds them.

handoff_current_work writes a boundary Source. Discussing a design, or saying preview and make no writes, must not call it.

After you enable Bearer

Loopback can leave auth off. Once POWERCONTEXT_SERVER_AUTH_ENABLED=true:

• /mcp itself needs Authorization: Bearer <token> for discovery and for tool calls;
• /openapi.json needs the same header;
• /docs can still open without a token. That does not mean MCP or the API is authorized.

Do not put the token in the URL. Use that Agent’s header helper or environment variable, such as POWERCONTEXT_CODEX_AUTHORIZATION on Codex.

401/403 is not empty. A tool failure must stay visible. Do not pretend a write succeeded.

How you know it worked

• The client connects to http://127.0.0.1:8000/mcp and lists remember_memory and search_memory.
• A Memory written through MCP is visible to the Python Client on the API tour.
• The tool list does not include prepare_context.
• remember_memory fails visibly after you stop the Server.
• With Bearer enabled, /mcp without a header is not a connected client.

If PowerContext fails while automatically fetching context, keep working. If you explicitly write a Memory, commit a Handoff, approve a proposal, or export a Skill, you must see the failure. Read fail-open for the full rule.

Next step

Check the counts in Dashboard

Use the same Scope as the Client and check that the counts match.

Open the Dashboard

How coding Agents register MCP

Codex and Claude Code use different field names. The portable path is Agent Plugin.

Pick a coding Agent