agentskm-toolkit 0.4.4

Local multi-agent knowledge capture toolkit with CLI and MCP server.

AgentsKM Toolkit

Reusable local toolchain for conversation-driven knowledge capture across multiple Agent hosts. Live knowledge data stays in a user-controlled Vault and is never bundled in the toolkit package or Codex plugin.

The product boundary is:

Python package provides CLI/MCP capabilities
Codex plugin provides host experience and Skill behavior
%USERPROFILE%\.agentskm\config.json connects Profiles to the user Vault
Vault stores the user's knowledge assets

Components

• src/agentskm_toolkit: productized Python package entrypoints
• tools/km-cli: deterministic capture, review, promotion, merge, locking, transactions, and audit log source implementation
• adapters/mcp: role-aware stdio MCP server source implementation
• plugins/agentskm-toolkit: Codex plugin, capture Skill, and MCP template
• integrations: host-specific notes for Claude, Cursor, Hermes, and others
• scripts: plugin packaging and MCP configuration generation
• tests: temporary-Vault end-to-end acceptance coverage

Package Entry

For local development:

pip install -e .
agentskm --version
agentskm doctor --profile codex
agentskm mcp --profile hermes-agent --host hermes

For install-on-use hosts after publishing:

uvx --from agentskm-toolkit agentskm mcp --profile hermes-agent --host hermes

The first MCP start creates %USERPROFILE%\.agentskm\config.json, the selected Profile, and the default empty Vault at %USERPROFILE%\Documents\AgentsKM\vault when no configuration exists. Custom Vault paths remain controlled by the user config file.

Codex Plugin

The repository is also a Codex Git marketplace source through .agents/plugins/marketplace.json. The plugin supplies the agentskm-capture Skill and an MCP configuration template. Before the Python package is published, the plugin template remains self-contained and starts its bundled MCP source. After publishing, productized MCP startup should use:

{
  "command": "uvx",
  "args": ["--from", "agentskm-toolkit", "agentskm", "mcp", "--profile", "codex", "--host", "codex"]
}

During development, the plugin still carries synchronized CLI/MCP runtime copies so existing local plugin installs remain self-contained. Use scripts/build_plugin.py --check to verify those copies are current.

Other Agents

Generate a role-limited MCP configuration:

python scripts/render_mcp_config.py `
  --agent hermes `
  --profile hermes-agent `
  --role contributor `
  --output hermes.mcp.json

Use --runtime source only for local source-tree testing. Published/user-facing configs should use the default uvx runtime or --runtime agentskm after a local package install.

Roles

The role boundary is invariant across startup modes:

contributor: search / propose / respond
reviewer:    + review / dashboard
compiler:    + promote / merge

A contributor can record user intent with respond, but cannot approve, promote, merge, or write formal Wiki pages.

Update Semantics

Source checkout and Git marketplace installs may use agentskm update to check or refresh local code. In uvx/PyPI-style execution, update should be treated as a version check plus reconnect instruction; the running package should not mutate itself in place.