@kynver-app/mcp-agent-os

Kynver Agentic OS MCP server. Exposes the agent_os_* family that proxies to /api/agent-os/{slug}/* on a Kynver deployment. Each tool accepts an optional slug argument; if omitted, it falls back to KYNVER_AGENT_OS_SLUG env, then the account's primary AgentOS slug, then to "ghost".

This is the external surface only. The in-app Kynver agent does not dispatch these tools — they exist for MCP clients (Claude Code, Cursor, OpenClaw, etc.) that connect to a running Kynver deployment.

Tools

The full family:

• agent_os_get_context
• agent_os_open_session
• agent_os_close_session
• agent_os_log_session_event
• agent_os_log_session
• agent_os_list_goals
• agent_os_update_goal
• agent_os_get_projects
• agent_os_create_project
• agent_os_update_project
• agent_os_search_memory
• agent_os_write_memory
• agent_os_update_memory
• agent_os_list_skills
• agent_os_get_skill
• agent_os_create_skill
• agent_os_update_skill
• agent_os_import_skills
• agent_os_bind_skill
• agent_os_write_skill_memory
• agent_os_get_contacts
• agent_os_create_contact
• agent_os_update_contact
• agent_os_consolidate_memory

AgentTask board (Phase 2 — batch 1 CRUD + batch 1.5 steering/lease):

• agent_os_task_create
• agent_os_task_get
• agent_os_task_list
• agent_os_task_update
• agent_os_task_log_event
• agent_os_task_close
• agent_os_task_steer
• agent_os_task_abort
• agent_os_task_claim
• agent_os_task_renew_lease
• agent_os_task_operator_priority

AgentPlan first-class plans (Phase 5A — versioned operational artifacts separate from goals/projects/tasks):

• agent_os_plan_create
• agent_os_plan_list
• agent_os_plan_get
• agent_os_plan_update
• agent_os_plan_add_version
• agent_os_plan_mark_current
• agent_os_plan_add_link
• agent_os_plan_list_links
• agent_os_plan_list_versions

Schemas are mirrored in lib/mcp/tool-manifests/kynver-mcp-agent-os.json in the Kynver repo (used as the static fallback when stdio spawn isn't available — e.g. on Vercel).

Memory write/update tools are the memory-writer surface. They accept sourceRefs for structured provenance plus memoryType, confidence, reviewStatus, projectId, goalId, contactId, and skillId. Use those fields to attach repo paths, commits, PRs, map files, sessions, tools, URLs, and review state to important memories instead of burying those pointers in freeform text. Set reviewStatus: "needs_review" for important or uncertain memories that a human should be able to find and correct in the dashboard.

Active-state and correction tools expose the "what is true right now" layer: agent_os_record_state / agent_os_get_state manage one current active_state row per live PR, branch, or deployment, while agent_os_correct_memory writes an auditable correction that supersedes stale or wrong remembered claims instead of silently editing history.

Skill tools expose the AgentOS skill registry and skill-tagged memory contract. Use agent_os_get_skill to fetch instructions on demand, agent_os_create_skill /agent_os_update_skill to manage user-authored skills, agent_os_import_skills to import external skill markdown/JSON as draft unbound user skills, agent_os_bind_skill to enable or update a workspace binding, and agent_os_write_skill_memory for lessons learned while applying a skill. Built-in skill memories use the built-in slug as skillId; user-authored skill memories use AgentSkill.id.

agent_os_log_session_event appends structured session events. Close-session summaries can then be synthesized from actual topics, actions, decisions, and follow-ups instead of generic lifecycle text.

AgentPlan tools manage first-class plans — versioned operational artifacts distinct from goals/projects/tasks. Each AgentPlan owns a chain of immutable AgentPlanVersion snapshots (one stamped current at a time) plus typed AgentPlanLink rows pointing at goals/projects/tasks/memories/sessions/plans or remote refs (repo/branch/commit/PR/URL/markdown). The plan body itself is immutable — to record a new snapshot call agent_os_plan_add_version, which appends a new version and (by default) advances currentVersionId atomically. Use planSlug for the plan's own stable slug; the workspace slug argument is reserved for the AgentOS workspace. Workspace-owned link targets are re-validated server-side against the same AgentOS workspace; remote refs carry metadata only.

AgentTask tools expose the durable task board. agent_os_task_create is idempotent within an AgentOS — pass idempotencyKey (or requestId) from the foreground turn so a retry or restart returns the existing task instead of spawning a duplicate. agent_os_task_log_event accepts a stable eventKey so a board-down executor can replay its local event spool without creating duplicates; artifact events take an artifactVisibility tier (internal | telegram_safe | public_pr_safe, default internal). agent_os_task_claim and agent_os_task_renew_lease implement the lease model — an executor must hold a live lease to own a task, and a claim is an atomic conditional update that increments attempt.

Env

| Var | Purpose | |---|---| | KYNVER_API_URL | Kynver origin, e.g. https://kynver.com | | KYNVER_API_KEY | Bearer token for external authentication | | KYNVER_AGENT_OS_SLUG | Default agent slug when the tool's slug arg is omitted (defaults to "ghost") | | KYNVER_SERVICE_SECRET | Required when running under SSE and forwarding internal userId headers | | MCP_SSE_HOST | SSE bind address (defaults to 127.0.0.1) | | PORT | SSE port (defaults to 3101) |

OpenClaw / Claude Code config

Once this package is published, the Ghost workspace's MCP config needs two separate Kynver MCP server entries — one for the analyst surface and one for this Agentic-OS surface:

{
  "mcpServers": {
    "kynver-analyst": {
      "command": "npx",
      "args": ["-y", "@kynver-app/mcp-analyst"],
      "env": {
        "KYNVER_API_URL": "https://kynver.com",
        "KYNVER_API_KEY": "${KYNVER_API_KEY}"
      }
    },
    "kynver-agent-os": {
      "command": "npx",
      "args": ["-y", "@kynver-app/mcp-agent-os"],
      "env": {
        "KYNVER_API_URL": "https://kynver.com",
        "KYNVER_API_KEY": "${KYNVER_API_KEY}",
        "KYNVER_AGENT_OS_SLUG": "ghost"
      }
    }
  }
}

The split mirrors the server-side change of 2026-05-14: the agent_os_* family was extracted from @kynver-app/mcp-analyst into this dedicated package so each AgentOS deployment can mount it independently from the analyst tools.