Connecting to agent environments

The agent environment of the RAGRAF author (zwork 2.0.2, Electron): the product was designed on it and the published results were measured on it. Connection runs over two independent channels — the CLI without configuration and MCP through the product interface; --agent-kind autoclaw is a real code path (telemetry attribution, extended journal mode). Everything in the guide is integration-session measurements and facts from the application code.

1. 1Start with the CLI channel — it works without configurationThe first step of a task is python <repo>\tools\ragraf.py ensure <directory>, then index operations instead of reading files. The CLI depends neither on app settings nor on restarts and cannot "silently fall off". If there are no ragraf_* tools in the agent list, that is not a breakage: the index, search, schema and graphs remain fully operational through the CLI (session measurement: tool_search → 0).
2. 2MCP is registered through the product interfaceMeasured: there is no file-based registration — no mcp.json or mcpServers in the profile, no server table in runtime.sqlite (all 106 tables checked). The server is entered in the project/application settings: the .venv interpreter, ragraf-mcp.py --agent AutoClaw --agent-kind autoclaw, working directory <repo>. --project is not needed: the corpus is resolved from the working directory, switching is ragraf_use_project.
3. 3Skill into both accounts, rules into ~/.pi/agent/AGENTS.mdragraf-core is deployed into the skill tree of each account (source — docs/skills of the project; copy freshness is controlled by SHA-256). Section 8 of the user rules prescribes starting tasks with ensure and stays executable in both configurations — with MCP and without.
4. 4Verify by measurement, not assumptiontool_search for ragraf: 0 before registration and 59 after the restart — that difference is the proof. Through the CLI: catalog (registry), ensure (status: actual), check (matching checksums). usage --agent AutoClaw shows ensure in the breakdown — the first-step rule is being followed.

команда:            <repo>\.venv\Scripts\python.exe
аргументы:          <repo>\tools\ragraf-mcp.py
                    --agent AutoClaw --agent-kind autoclaw
рабочий каталог:    <repo>

An open platform for a team of AI employees (Streamlit, MIT, v1.13.1): assistants, autonomous orchestrators, the built-in DevAgent developer agent, knowledge bases and a skills library. The platform code has no MCP client — verified by a repository search — so RAGRAF connects through the platform native mechanisms: the ragraf-core skill and CLI calls via the run_code tool.

1. 1Build the corpus once — shared by every employeeThe index is read-only for agents: several orchestrators and dialogs work with one corpus at the same time. The first step is ensure <project directory>.
2. 2Install the ragraf-core skill into the libraryInstall from a GitHub URL, ZIP or folder (skills/skills.json registry; limits: 200 MB, 5000 files, 50 MB per file). A third-party skill does not reach the system prompt until DevAgent (Skill Developer) completes adaptation — mark_skill_adapted.
3. 3Put the regimen into orchestrator instructionsThe five rules of the universal fragment are the minimal sufficient set; add them to the employee instructions (persisted across restarts) or to the project AGENT.md. Without a regimen agents fall back to search_in_files — immediately visible in the call statistics.
4. 4CLI calls — via run_coderun_code executes commands in an isolated subprocess (300 s timeout; dangerous patterns only with UI confirmation): catalog, ensure, query symbol, check, usage --agent saga-worker. The agent reads the response text itself — no schema interpretation needed.

{
  "mcpServers": {
    "ragraf": {
      "command": "python",
      "args": [
        "<repo>/tools/ragraf-mcp.py",
        "--agent", "saga-worker",
        "--agent-kind", "external"
      ]
    }
  }
}

The DeepSeek Harness 0.2.0-rc.2 Electron app with a built-in MCP client (the @deepseek-ai/dsh-mcp-client plugin). RAGRAF connects to it as an MCP server over stdio; its tools surface as mcp__ragraf__ragraf_*. The pairing was debugged end to end and confirmed by live-session measurements: everything below is reproducible commands, paths and measured outputs.

Full Harness setup guide

Your environment is not listed? If the client can spawn a stdio process and speak JSON-RPC — that is enough. Three steps: server config, regimen into the system prompt, capabilities check as the first call.

# Регламент работы с RAGRAF (фрагмент AGENTS.md)
1. Навигация — только через ragraf_symbol / ragraf_search.
   Полный файл — только ragraf_live_read и только по факту из индекса.
2. Проверяй конверт: stale_warnings → перечитай затронутое (live_read);
   capability=degraded → не планируй задачи со slice.
3. Перед любой записью в БД — ragraf_write_check (всегда).
4. Перед правкой символа — ragraf_impact --depth 3.
5. «Не найдено» — валидный ответ: сообщай limits из конверта, не додумывай.

# сервер отвечает на инициализацию MCP и раскрывает режим корпуса
python tools/ragraf-mcp.py --project corpus.yaml --agent probe --agent-kind external
# → JSON-RPC: 59 инструментов ragraf_*; list_capabilities:
#   {"capability":"full","operations":[...22...]}   (или degraded + reason)

# телеметрия после пробных вызовов
ragraf -p corpus.yaml query usage_stats
# → usage_call содержит записи agent=probe, kind=external