Agent Skills and MCP

Callra ships two machine-facing surfaces for AI agents:

1. Standard Agent Skills in the repository skills/ directory.
2. MCP tools served locally by the CLI or remotely from your Callra app at /api/mcp.

Use Agent Skills when you want portable instructions and workflows that any standards-compatible agent can load on demand. Use MCP when you want the agent to execute live queries against your Callra data.

What ships in the repository

Callra includes the following standard Agent Skills:

• skills/callra-site-resolution
• skills/callra-web-analytics
• skills/callra-speed-regression
• skills/callra-release-impact

Each skill follows the Agent Skills specification:

skill-name/
├── SKILL.md
├── references/
└── assets/ or scripts/ (optional)

The SKILL.md file contains YAML frontmatter plus Markdown instructions. Compatible agents can discover the skill from its metadata, then load the full instructions only when the task needs it.

Installation example: use the repository skills/ directory directly

If your agent supports the Agent Skills standard and lets you point at a local skills directory, use the top-level skills/ directory from this repository.

/your-checkout/callra/skills

Typical setup flow:

1. Clone or open the Callra repository.
2. Find the absolute path to the skills/ directory.
3. Add that directory to your agent's configured skills locations.
4. Restart or refresh the agent so it re-indexes the available skills.

Because the format is standardized, the important part is the directory structure, not a vendor-specific packaging format.

Installation example: copy one skill into a shared skills workspace

If your team keeps a central skills workspace, you can copy a single Callra skill directory into it.

mkdir -p ~/agent-skills
cp -R /path/to/callra/skills/callra-release-impact ~/agent-skills/

That gives the agent one focused release-analysis skill without requiring the whole repository.

Installation example: validate a skill before distributing it

The Agent Skills ecosystem provides a reference validator.

skills-ref validate ./skills/callra-web-analytics

Use this when you edit or extend Callra's bundled skills.

Pair skills with live MCP access

Skills explain how an agent should investigate a problem. MCP gives the agent the live data tools it needs to do the work.

For release regressions, the recommended higher-level flow is now:

1. callra.release_snapshot
2. callra.release_breakdown
3. callra.release_compare

That keeps most investigations within 2-3 tool calls instead of forcing the model to guess whether it should drill into raw web or speed tools first.

Callra supports both a local and a remote MCP flow.

Local MCP via CLI

Start the local stdio MCP server:

callra mcp serve

This is useful when your agent or editor can launch local MCP servers directly.

Remote MCP via /api/mcp

Callra also exposes a remote MCP endpoint:

https://your-callra-host.example/api/mcp

To use it:

1. Open Dashboard -> Site settings -> Agent access.
2. Mint a site-scoped agent token.
3. Configure your MCP-compatible agent to connect to the remote endpoint.
4. Send the token as a Bearer token.
5. Use the endpoint URL shown in the dashboard.

Callra's remote MCP endpoint supports:

• MCP initialization
• site-scoped tool calls
• resources
• prompts
• session headers for streamable HTTP flows

Recommended setup pattern

For most teams, the best experience is:

1. Load the standard Callra skills from skills/.
2. Connect the same agent to Callra's MCP endpoint.
3. Let the skills decide when to call:
   • callra.resolve_site_context
   • callra.web_snapshot
   • callra.speed_snapshot
   • callra.release_snapshot

This gives you both:

• reusable investigation workflows
• live analytics and release data
• a short, deterministic release-regression loop built around release_snapshot, release_breakdown, and release_compare

Editor and agent examples

Codex / CLI-style agent

Use the remote MCP endpoint from Site settings -> Agent access and ask:

Why did the latest production release get worse?

The intended path is:

1. callra.resolve_site_context
2. callra.release_snapshot
3. callra.release_breakdown
4. callra.release_compare only if the first answer still looks noisy

Claude Code / Cursor

If your tool can launch local MCP servers, run:

callra mcp serve

Then prompt with:

Investigate the latest release regression for my default site and summarize the top metric, affected pages, and likely commits.

Remote MCP over HTTP

If your MCP client supports remote HTTP transports, point it at:

https://your-callra-host.example/api/mcp

Use a site-scoped bearer token minted by the site owner. The same token governance surface also shows audit rows with actor type, transport, tool, and timestamp.

Which surface should you use?

Use Agent Skills when you need

• reusable workflows
• standard task instructions
• portable, repo-versioned agent expertise
• compatibility across many agent products

Use MCP when you need

• live data
• authenticated access to a specific Callra site
• real Web Analytics, Speed Insights, and Release Analytics results
• machine-readable tool outputs

Use both when you need

• the best overall agent experience
• consistent reasoning plus live telemetry

Next step

After your agent can load the Callra skills, configure live MCP access and then try one of these prompts:

• "Analyze traffic changes for the last 7 days."
• "Did the latest release hurt LCP?"
• "Which release looks most suspicious this week?"