continual-harness-mcp-server

The Continual Harness as an MCP server. It stores durable prompts, subagents, skills, and memories with full CRUD and an evidence-backed refinement pass, scoped local to a session or global across every host you use.

Part of the harness monorepo. It has no model or sandbox dependency, so it stays light and starts fast. State is written in prime-agent's harness_state.json format, so a store written by either side is readable by the other.

Installation

npm install continual-harness-mcp-server

Usage

Run it over stdio, the transport a local host spawns:

npx continual-harness-mcp-server

State is one JSON file per scope, matching prime-agent's layout:

<dir>/harness_state.json     schema, entries bucketed by kind, refinements

local state resolves from CONTINUAL_HARNESS_DIR, then prime-agent's RLM_HARNESS_STATE_DIR / RLM_SESSION_DIR, then .rlm-mcp/harness under the working directory. global state resolves from CONTINUAL_HARNESS_GLOBAL_DIR, then RLM_GLOBAL_HARNESS_STATE_DIR, then the prime-agent home.

The same binary is a CLI, so a person or a shell-driven model can use it directly. The installed command is continual:

continual add-memory "Prefer focused patches" "Small updates validate faster."
continual add-skill "run tests" "invoke the suite" --import agent_skills.test --callable run
continual refine "learned to verify first" --edits '[{"action":"create","kind":"prompt","title":"always verify","content":"Verify every case."}]'
continual get --scope global                    # print the state as JSON

Any subcommand runs the CLI; no subcommand starts the MCP server, so hosts are unaffected.

The package is also a library:

import { Harness, FileStore, loadConfig, systemClock } from 'continual-harness-mcp-server';

const harness = new Harness(new FileStore(loadConfig()), systemClock);
await harness.createMemory('Prefer focused patches', 'Small updates validate faster.');

Tools

| Tool | Purpose | | -------------------------- | ------------------------------------------------------- | | harness_get_state | Read every entry and refinement for a scope | | harness_create_memory | Store a durable memory (component M) | | harness_create_prompt | Store a prompt note added to the system prompt (component p) | | harness_create_skill | Store a skill with a Python call contract (component K) | | harness_create_subagent | Store a subagent spec (component G) | | harness_update | Update any entry by kind and id, bumping its version | | harness_delete | Delete any entry by kind and id | | harness_refine | Apply a multi-component diff and record what changed |

State is also exposed as a resource at harness://state/local and harness://state/global.

Refinement

harness_refine takes a summary and a list of edits, and applies them across any mix of the four kinds in one pass, mirroring the paper's Δ = (Δp, ΔG, ΔK, ΔM). Each edit is a create, update, or delete; the pass records one refinement listing what changed, with its trigger, evidence, and expected outcome.

{
  "summary": "learned to verify before answering",
  "evidence": "two failed validations",
  "edits": [
    { "action": "create", "kind": "prompt", "title": "always verify", "content": "Verify against every case before answering." },
    { "action": "delete", "kind": "skill", "id": "stale_helper" }
  ]
}

The host side reflex

MCP holds the state. Reading it before a turn and calling harness_refine after is the host's job, which the protocol cannot do on its own. This package ships a hook, hooks/inject-harness.mjs, that reads the store and injects every kind including bodies. Wire it on SessionStart to seed each session, and on UserPromptSubmit so a refinement written mid-session re-enters context on the next turn. The hook depends only on the on-disk format, so it runs without building the package. See NOTES.md.

Structure

src/
  harness.ts     create, update, delete, refine, snapshot, overview
  types.ts       HarnessEntry, RefinementEvent, RefinementProposal
  store.ts       HarnessStore port, FileStore (single JSON file per scope)
  config.ts      resolves the local and global directories
  server.ts      builds the MCP server and tools, transport agnostic
  system.ts      systemClock          local.ts   stdio entrypoint
  index.ts       public API
hooks/
  inject-harness.mjs   the read-before-a-turn hook