Imported from google/capsem (skills/dev-mcp/SKILL.md). Install upstream with npx skills add google/capsem --skill dev-mcp. Copyright stays with the author.

name: dev-mcp description: MCP for Capsem: the host server, the guest relay, and tool routing. Use when working on MCP servers, tool policy, or telemetry, or to debug quickly inside a VM.

MCP in Capsem

Capsem has two MCP components:

1. capsem-mcp (host): MCP server over stdio that lets AI agents (Claude Code, Gemini CLI) control sandboxes -- create/delete VMs, exec commands, read/write files, and read typed telemetry routes. Bridges to capsem-service HTTP API over UDS.
2. Guest MCP relay + MITM MCP endpoint: bridges AI agents running inside a guest VM to external MCP servers on the host via framed MCP records over vsock port 5002.

Using capsem MCP tools for fast debugging

When the capsem MCP server is configured in your AI CLI, you have direct VM control without leaving the conversation. This is the fastest debug loop for any in-VM work.

Available tools

Debug workflow

-- Quick one-shot (no VM management needed):
capsem_run { command: "capsem-doctor -k net" }

-- Iterative debugging (long-lived VM):
1. capsem_create        -- boot a fresh sandbox
2. capsem_exec          -- run the thing you want to test
3. capsem_read_file     -- check config, logs, state
4. capsem_timeline      -- inspect typed telemetry without raw SQL
5. (fix code on host, rebuild with `just build`)
6. capsem_delete        -- tear down
7. repeat from 1

Common debug patterns

Verify a guest command works:

capsem_exec { id: "vm-1", command: "capsem-doctor -k net" }

Check network policy enforcement:

capsem_exec { id: "vm-1", command: "curl -s https://blocked-domain.com" }
capsem_timeline { id: "vm-1", layers: "net", limit: 10 }

Verify telemetry pipeline:

capsem_timeline { id: "vm-1", layers: "tool,fs", limit: 50 }

Read guest runtime state:

capsem_read_file { id: "vm-1", path: "/etc/resolv.conf" }
capsem_read_file { id: "vm-1", path: "/tmp/capsem-init.log" }

Write a test script and run it:

capsem_write_file { id: "vm-1", path: "/tmp/test.sh", content: "#!/bin/bash\necho hello" }
capsem_exec { id: "vm-1", command: "chmod +x /tmp/test.sh && /tmp/test.sh" }

When to use MCP tools vs just recipes

MCP tools are for fast, targeted checks during development. Just recipes are for comprehensive validation before committing.

capsem-mcp (host MCP server)

Architecture

AI Agent (Claude Code) <-> capsem-mcp (stdio, rmcp) <-> HTTP/UDS <-> capsem-service

Uses the rmcp crate with #[tool_router] macro for tool definitions. Stateless -- creates a fresh HTTP connection to ~/.capsem/run/service.sock per request.

Parameter conventions

MCP tools use camelCase on the wire (ramMb, cpuCount) because that is the MCP/JSON convention. The capsem-service HTTP API uses snake_case (ram_mb, cpus). The conversion happens inside each tool method -- the #[serde(rename)] attributes on param structs handle deserialization, and the tool builds a new JSON body with the service's field names.

Key source files

Configuration

Registered in AI CLI settings:

{ "mcpServers": { "capsem": { "command": "cache/target/cargo/debug/capsem-mcp" } } }

Environment variables

MCP subprocess architecture

The guest MCP path is not a single process. capsem-process (the per-VM host process) owns the MITM MCP endpoint and spawns two privilege-isolated subprocesses that together handle MCP traffic from the guest:

Rationale: isolating external-server connections in a low-privilege subprocess means a compromised third-party MCP server cannot reach the host filesystem or the session DB. The built-in tool server runs in its own process for the same reason.

Wire protocol between capsem-process and the aggregator: length-prefixed msgpack frames on stdio ([4-byte big-endian length][msgpack payload]). Between the aggregator and the built-in server: stdio MCP (standard JSON-RPC per line). Between the in-guest AI agent and capsem-process: /run/capsem-mcp-server relays stdio JSON-RPC as bounded framed MCP records over vsock port 5002. Tool activity writes the canonical tool_calls ledger. Visible MCP protocol facts are represented as typed security events and security_rule_events.

Binaries land in ~/.capsem/bin/ at install time: capsem-mcp-aggregator, capsem-mcp-builtin.

Guest MCP Endpoint

The guest MCP relay bridges AI agents in the guest VM to the host MITM MCP endpoint. It runs over vsock port 5002 using bounded length-prefixed MCP frames that carry JSON-RPC payloads and per-frame process attribution.

Framed guest MCP over vsock:5002 must be tested as the default transport, not as an opt-in benchmark mode. The minimum hardening matrix for that path is:

• parser/interpreter: bounded frames, invalid JSON, malformed flags, stream-id reuse, notification/request-id mismatch
• dispatch: initialize, tools/list, builtin tools/call, configured external stdio tools/call, resources/list, prompts/list, and method error mapping
• policy: live policy mutation, per-tool block, resource URI rule, argument-name rule, argument-value rule, return-value rule, deny-over-allow precedence, and proof that blocked requests/responses do not leak original data
• telemetry: session.db rows for success, denial, timeout, process attribution, request/response previews, policy fields, and terminal errors
• boundary: aggregator remains DB-free; MITM/process owns MCP audit writes
• VM E2E: boot a real VM, run /run/capsem-mcp-server with no transport override, then query session.db

Architecture

Guest (Claude/Gemini) -> capsem-mcp-server (stdin/stdout relay)
  -> framed vsock:5002 -> MITM MCP endpoint (capsem-core)
  -> SecurityEvent rule check -> Route to: builtin tools | external MCP servers (via rmcp)
  -> Telemetry -> session.db tool_calls table plus security_rule_events protocol evidence

Wire format

Length-prefixed MCP frames over vsock. Each frame contains a bounded JSON-RPC payload plus a stream id, flags, and sanitized process name.

Handshake

Guest sends NUL-prefixed metadata line first:

\0CAPSEM_META:claude\n

Then JSON-RPC messages:

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}
{"jsonrpc":"2.0","id":2,"method":"tools/list"}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"github__search","arguments":{"q":"rust"}}}

Tool namespacing

Tools are namespaced with __ (double underscore) to prevent collisions:

• github + search_repos -> github__search_repos
• builtin + http_get -> builtin__http_get

The endpoint parses the namespace to route to the correct server.

Supported methods

Security evaluation

1. Parse MCP frame into typed `SecurityEvent` MCP fields.
2. Apply the shared security engine plugin/rule rail.
3. Dispatch only if the effective action allows it.
4. Log canonical tool row, optional MCP protocol row, and matched security rule rows.

Config hierarchy: corp config constrains profile config. Profile config owns MCP servers, tools, resources, default rules, and plugin policy. There is no MCP-specific decision provider or user.toml override rail.

Decisions use the shared security action enum: allow, ask, block, rewrite, preprocess, and postprocess. ask waits for an approval or denial before dispatch; block returns a policy JSON-RPC error without calling the tool.

Built-in tools

Snapshot tools (VirtioFS mode only)

snapshots_list, snapshots_changes, snapshots_create, snapshots_delete, snapshots_revert, snapshots_history, snapshots_compact

HTTP tools (always available)

http_get, http_post, http_put, http_patch, http_delete, http_head

All use namespace prefix builtin (e.g., builtin__http_get).

Endpoint key source files

Telemetry (tool_calls and security events)

tool_calls is the canonical user/security ledger for all tool origins: model-native, built-in/local, and MCP-origin. UI counts, CEL rules, and forensic tool activity must use this table.

MCP protocol facts such as initialize/list/resource frames are security events. A visible MCP tools/call without a matching tool_calls row is a bug.

Every request/response logged with: timestamp, server_name, method, tool_name, compact request/response display fields, decision, duration_ms, error_message, process_name, bytes sent/received. Full MCP request/response payloads share the same event_body_blobs ledger table as HTTP and model traffic; do not add a second MCP-only body rail.

Read references/mcp-wire.md for the full wire format details.

Testing

Unit tests

cargo test -p capsem-mcp -- param serde roundtrips, UDS path resolution, tool router registration, schema constants.

cargo test -p capsem-core mcp -- gateway, policy, server manager, type serialization.

Integration tests (Python)

The MCP integration tests (tests/capsem-mcp/) are black-box tests that boot a real service + VM and exercise the full MCP protocol over stdio.

Run with: uv run --project build_system --frozen pytest tests/capsem-mcp/ -m mcp -v

Test files:

Fixture architecture:

• capsem_service (session scope) -- spawns capsem-service on isolated temp socket, codesigns binaries on macOS
• mcp_session (per-test) -- fresh capsem-mcp subprocess with JSON-RPC handshake, returns McpSession helper
• shared_vm (session scope) -- one long-lived VM for non-destructive tests, avoids repeated boot overhead
• fresh_vm (per-test factory) -- creates uniquely named VMs with auto-cleanup for destructive tests

McpSession helper (tests/capsem-mcp/conftest.py): wraps capsem-mcp subprocess with JSON-RPC 2.0 protocol. Key methods:

• request(method, params) -- send NDJSON, read response
• call_tool(name, args) -- call tool, assert success, parse JSON content
• call_tool_raw(name, args) -- raw response (no assertions)

In-VM diagnostics

just exec "capsem-doctor -k mcp" -- tests tool routing and domain blocking inside the guest.

Manual validation

Boot interactively, run a workload, then inspect telemetry:

just shell
# (in another terminal)
python3 build_system/scripts/doctor/check_session.py <vm_id> "SELECT * FROM tool_calls WHERE origin = 'mcp'"

Or use MCP tools directly (see "Fast debugging" section above) for the same workflow without leaving Claude Code.

Lessons learned

1. Never prepend headers to JSON output. MCP tool responses with format=json must return raw, parseable JSON. Do not wrap JSON in pagination headers, content-length prefixes, or any other text. The snapshots_changes tool broke because paginated_response() prepended "Content length: ...\nShowing: ...\n" to the JSON array, making json.loads() fail. Rule: if a tool offers both text and JSON formats, branch early and return JSON directly without passing through text-oriented helpers like paginated_response().