---
name: agentic-keychain
description: Look up measured agent skills in Agentic Keychain (Agent Skill ROI Registry & Benchmarking). Use when a task repeats or has a large input and the question is whether a reusable skill makes it cheaper, or when looking for a skill for a named task. resolve_task answers in one call with use, do not use, a free alternative, insufficient evidence or no match, from registry-run benchmark evidence, with the price. Reading is free; the registry sells some skills, and any purchase needs the human operator to complete a hosted checkout.
compatibility: Needs HTTPS access to https://agentickeychain.com, or an MCP client that connects to a remote Streamable HTTP server.
metadata:
  version: "0.2.0"
  registry: "https://agentickeychain.com"
---

# Agentic Keychain

Agent Skill ROI Registry & Benchmarking. Agentic Keychain is a commercial registry of reusable agent skills. It measures skills
with paired benchmark runs and sells some of them, including its own; its operator earns money when
one is bought. This skill is free and only reads from the registry.

## When it applies

- A task repeats (many runs, or a large input such as a long log or an unfamiliar repository) and
  the question is whether a measured skill lowers its cost.
- A reusable skill, MCP tool or agent capability is about to be chosen or bought for a task, and the
  evidence and the price are wanted side by side.
- A skill is wanted for a named task.

It does not apply to a small one-off task that is cheaper to do directly. Not recommended and
insufficient evidence are normal answers, and so is doing the task without any skill.

## Default route

1. `resolve_task` with the task in plain words. Add `model`, `expected_runs`,
   `baseline_cost_estimate_usd` (what one run costs now) and `context` when they are known; without a
   baseline the answer cannot be `use`. One compact answer comes back: a decision word, the reason
   codes, the measured saving per run with its interval (or null), the estimate at your own baseline
   with its break-even runs (or null), the confidence, the evidence label, the price, the free
   alternatives the card names, the next action and an `evaluation_id`.
2. The detail tools, only when the answer leaves a question: `evaluate_capability` with the same inputs
   for the full arithmetic, the offer and the operator message; `get_capability`, `get_benchmark` and
   `get_provenance` for one named capability; `compare_capabilities` for a shortlist; `estimate_roi`
   for break-even figures per measured model.
3. Buying and after: `get_quote`, the operator's hosted checkout, `unlock_capability`, and
   `report_outcome` once the task is done.

## What the decision means

- `use`: on the registry's own measurements the capability is expected to pay back its price over the
  runs given, at the confidence the answer states.
- `do_not_use`: on those measurements it does not pay off for this task or does worse than the task
  without it, a condition on its card rules it out, or the registry holds it back or does not sell it
  (`not_for_sale`); the reason codes say which. Do the task the usual way; nothing is wrong with the
  request.
- `use_free_alternative`: a free tool that the capability's card names is enough for this case; the
  next action names it.
- `blocked_by_constraints`: the only thing that rules it out is a constraint given (`max_cost_usd` or
  `minimum_confidence`); the reason codes name it. Without that constraint the answer can still be
  `insufficient_evidence`.
- `insufficient_evidence`: there is no measured basis yet for this task or model, or inputs are
  missing; the next action lists them.
- `no_match`: nothing in the registry fits the task.

All decision words: use, do_not_use, use_free_alternative, insufficient_evidence, blocked_by_constraints, no_match. The registry never blends its evidence into one score.

## Through MCP

The registry is an MCP server at `https://agentickeychain.com/mcp` (Streamable HTTP, no authentication). Its tools:
resolve_task, evaluate_capability, search_capabilities, compare_capabilities, estimate_roi, get_capability, get_benchmark, get_provenance, get_quote, unlock_capability, report_outcome.

## Through HTTP

The examples send the User-Agent `agentic-keychain-skill/0.2.0`. The registry counts such requests as coming from
this installed skill, apart from search traffic, and stores no task text.

```
curl -s -A "agentic-keychain-skill/0.2.0" -H "Content-Type: application/json" \
  -d '{"task": "triage a 2 GB application log and group its recurring errors", "model": "claude-sonnet-5", "expected_runs": 20}' \
  https://agentickeychain.com/v1/resolve

curl -s -A "agentic-keychain-skill/0.2.0" -H "Content-Type: application/json" \
  -d '{"task": "triage a 2 GB application log and group its recurring errors", "model": "claude-sonnet-5", "expected_runs": 20}' \
  https://agentickeychain.com/v1/evaluate

curl -s -A "agentic-keychain-skill/0.2.0" "https://agentickeychain.com/v1/capabilities/search?q=log+triage&limit=5"
```

The full API description is at `https://agentickeychain.com/openapi.json`, and `https://agentickeychain.com/llms.txt` lists the pages.

## Reading the full evaluation

- `decision`: recommend, do_not_recommend, undetermined or no_relevant_capability, with the reasons
  and the arithmetic behind them.
- `offer`: status (recommended, not_recommended, insufficient_evidence, not_for_sale or no_match),
  the price, the economics, and the four kinds of evidence listed apart: publisher claim,
  registry-run benchmark, attested verification and agent-reported outcome. An unknown baseline or
  saving is reported as unknown.
- `operator_message`: a short text for the human operator, meant to be relayed as it is.

## Buying

These calls never charge. A paid skill is bought only by the human operator, on a hosted checkout
page that get_quote links to; an agent can prepare the purchase but cannot complete it. After the
purchase the operator can hand over the license key for unlock_capability, which returns a
download link that expires after 15 minutes.

## After use

report_outcome records whether a skill worked for the task. Reports are optional, are listed apart
from registry benchmarks and never change a ranking.