tokvera for coding agents

tokvera reaches coding agents in three ways:

• An MCP server: tokvera mcp.
• A plugin: for Claude Code and for Codex.
• A skill: tokvera install-skill installs it on its own.

All three send load only to an endpoint the user named, and only after the user agreed to it.

tokvera mcp

A Model Context Protocol server on stdio. Add it to an agent:

claude mcp add tokvera -- tokvera mcp            # Claude Code
codex mcp add tokvera -- tokvera mcp             # Codex

Tool	What it does	Touches
scan_demo	a full graded report from the built-in simulator	nothing (loopback)
scan_plan	what a scan would send: suite, caps, worst-case cost	nothing
scan	measure the endpoint and grade it; needs confirm: true	the target, with load
summarize_report	a report.json’s grade and findings; with baseline, whether it regressed	nothing
verify_report	a signed report’s signature and files (docs/report-signing.md)	nothing
list_suites	the load shapes	nothing
doctor	the installation and the tools tokvera uses	nothing

• scan refuses without confirm: true. Its description tells the agent to show scan_plan to the user first. It is the only tool marked as touching the outside world (openWorldHint) and the only one not read-only.
• Defaults here: scan uses the quick suite unless told otherwise, and writes report.html and report.json to out, or to a new temporary directory whose path it returns.
• Cancellation: a call the client cancels stops the scan’s load.
• Protocol versions:
  • the stateless 2026-07-28 revision (per-request _meta, server/discover);
  • through initialize, 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05.
• Output: stdout carries only protocol messages, and logs go to stderr.

The plugin (Claude Code and Codex)

One plugin, agents/tokvera, carries both tools’ manifests. It brings the tokvera mcp server and the tokvera skill. tokvera itself must be on PATH.

# Claude Code (2.1.275 or later)
/plugin install tokvera --marketplace rajasatyajit/tokvera
# or, from a shell:
claude plugin marketplace add rajasatyajit/tokvera
claude plugin install tokvera@tokvera

# Codex
codex plugin marketplace add rajasatyajit/tokvera
codex plugin add tokvera@tokvera

File	For
.claude-plugin/plugin.json, .mcp.json	Claude Code
plugin.json, mcp.json (the portable agent-plugins format, type: stdio)	Codex
skills/tokvera/SKILL.md	both (the open Agent Skills format)
/.claude-plugin/marketplace.json, /.agents/plugins/marketplace.json	each tool’s marketplace, at the repository root

When tokvera is not installed, the skill tells the agent to ask the user first. Then it runs the official installer, which downloads from GitHub, and verifies the SHA-256 and, when cosign is installed, the release’s signature.

Versioning: the plugin’s version is the release’s. The release preflight refuses a tag the manifests disagree with.

The marketplaces work once the repository is public.

tokvera install-skill

The skill without the plugin:

tokvera install-skill --agents all --dry-run      # the plan and its cost, nothing written
tokvera install-skill --agents claude,codex       # asks, then writes

• Where the skill goes:
  • Claude Code: ~/.claude/skills/tokvera/SKILL.md;
  • Codex: ~/.agents/skills/tokvera/SKILL.md;
  • with --scope project, under .claude/ and .agents/ in the current directory instead.
• The plan comes first, with its cost.
  • The skill’s name and description join each agent’s skill list at every start: about 350 characters, roughly 90 tokens.
  • The whole skill, about 3.3 KB, loads only when the agent uses it.
  • Tokens are estimated at about four characters each.
• Then it asks: a “y” at the terminal, or --yes.
• An existing SKILL.md that differs from tokvera’s is kept unless --force.