MCP Context
Access Rate
The fraction of agent-authored PRs that were informed by rkito design context before code generation began.
This page describes the intended design. It requires MCP server access-log tracking, which doesn't exist anywhere in the product yet. It doesn't appear on your dashboard yet, and no org currently has a live value for it.
MCP tool called before
first commit
÷ total agent-authored PRs)
× 100
What it signals
MCPCAR is the MCP integration adoption metric. It measures the fraction of agent-authored PRs that were informed by rkito design context before code generation began. Low MCPCAR means agents are writing code without Ledger context, Steering context, or CHI scope — producing output that cannot be expected to conform to a design system the agent never consulted.
High MCPCAR is the prerequisite for positive AHCR. Agents cannot outperform humans on design conformance without first loading the governed design context that defines what conformance means. An agent that codes without calling rkito MCP tools is operating on implicit or stale context — exactly the condition that produces high ADRIFT scores and negative AHCR.
MCPCAR is the earliest leading indicator in the agentic health stack. Before AHCR can turn positive, MCPCAR must be high. Before LSCR can improve, the Lens context must actually be loaded. MCPCAR is the signal that tells you whether the investment in connecting agents to rkito is being realised operationally.
How rkito produces it
MCP server access logs are correlated with branch creation timestamps. A session that called get_ledger_state, get_applicable_lenses, get_steering_context, or get_approved_chi before the first commit on the associated branch increments MCPCAR. Detected automatically from MCP server access logs.
Who this metric is for
MCPCAR is the adoption metric for the MCP integration. It shows whether the investment in connecting agents to rkito is being realised operationally. If MCPCAR is high and AHCR is still negative, the context is being loaded but the agents are not using it effectively — a different problem requiring agent configuration review, not deployment work.
MCPCAR tells you whether your .mcp.json and CLAUDE.md bootstrap instructions have been deployed across all agent-enabled repositories. A low MCPCAR is not an agent quality problem — it is a deployment gap. The agents that are loading context consistently outperform those that are not. Fix the deployment and the metric follows.
Watch
MCPCAR measures that an MCP tool was called, not that the returned context was meaningfully used. An agent that calls get_ledger_state() and then ignores the response will show MCPCAR=100 and still produce high ADRIFT. The call is logged. The usage is not.
Pair MCPCAR with AHCR and LSCR to confirm that high MCP access is translating into conformant output. MCPCAR rising without AHCR improving is the signal that agents are loading context but not acting on it — an agent configuration or prompt engineering problem, not a deployment problem.
What to do at each threshold
Step-by-step response playbook for each signal state.
MCP integration is broadly adopted. Agents are loading design context before coding begins.
Monitor AHCR and LSCR to confirm context loading is producing conformant output. High MCPCAR with negative AHCR means agents are loading context but not acting on it — shift focus from deployment to agent configuration review.
Track MCPCAR by repository to identify any new repos that onboarded agents without the MCP bootstrap instructions. High org-level MCPCAR can mask individual repo gaps as the agent footprint expands.
Partial adoption. Some agent configurations or workflows are not loading MCP context before coding.
Identify which agent configurations or workflows are not loading MCP context. The most common cause: agents configured before MCP integration was deployed. These sessions have no bootstrap instruction and no default MCP call in their startup sequence.
Update agent startup instructions or system prompts to include rkito MCP bootstrap calls. A single line in CLAUDE.md — “call get_ledger_state before writing any code” — is sufficient to convert an MCPCAR miss into a hit.
Most agents are coding without design context. MCP integration is not deployed broadly enough.
Prioritise deployment of .mcp.json and CLAUDE.md bootstrap instructions across all agent-enabled repositories. This is a deployment task, not an agent tuning task — the agents will perform better the moment context is available.
Do not interpret low MCPCAR as an indictment of agent capability. Agents without context are expected to produce drift — that is the AHCR baseline condition. The fix is not to restrict agents; it is to give them the design context they need before the first line is written.
MCPCAR is the leading indicator for AHCR. High MCPCAR is the prerequisite for positive AHCR — agents cannot outperform humans on design conformance without first loading the governed context that defines what conformance means.
LSCR confirms that high MCPCAR is translating into Lens-level conformance. If MCPCAR is high but specific Lens scores remain low, agents are loading context but not applying the criteria it contains.
The companion governance metric for agent sessions. AIFI measures whether agents file Change Intentions before coding. MCPCAR measures whether they load design context. Both must be high for agent governance to function.
What is your org's MCPCAR?
rkito computes MCPCAR automatically by correlating MCP server access logs with branch creation timestamps. Connect your repo and see your agent context adoption rate within minutes.