# 20 — MCP server (risc-v-gsp-rm-debugger agent interface)

> Status: implemented, documented and tested (audit: docs/27-TOOL-AUDIT.md; value demos: docs/25-VALUE-VALIDATION.md).

## 1. What it is

`plugins/mcp_server.py` is the risc-v-gsp-rm-debugger **MCP (Model Context Protocol)
server** — roadmap [`R3`](08-ROADMAP.md#r3-mcp-server-plugin) and
[`H7`](../RISC-V-GSP-RM-DEBUGGER-ROADMAP.md). It exposes the *active* `Session` as MCP
tools so an AI agent (Claude Code, opencode, any MCP client) can inspect
symbols, set breakpoints/watchpoints, run the target, read memory, snapshot,
analyse/compare recordings, inject services, and drive the harnesses.

* **Stdlib-only**: JSON-RPC 2.0 over newline-delimited stdin/stdout, no MCP
  library, no network listener unless explicitly asked.
* **Plugin first**: `register(ge)` adds an `mcp-serve` REPL/CLI command that
  serves the session already loaded by risc-v-gsp-rm-debugger.
* **Standalone too**: the same file runs as a script that launches its own
  session (`--kernel`, `--backend`, `--record`) and then serves MCP.
* **Read-only by default**: memory writes, remote calls and injection require
  explicit startup flags.
* **Defensive capabilities**: tools owned by other components (mutation and
  portability harnesses, native trace, injection) return a structured
  `not_available` tool error when the module/file is absent instead of
  crashing. The tool list always advertises every capability.

The core risc-v-gsp-rm-debugger package is untouched; everything lives in the plugin.

## 2. Running it

All commands run from `open-iov/debug/risc-v-gsp-rm-vm/`.

### 2.1 Standalone (the usual MCP client mode)

```bash
# bundled demo, QEMU, recorded session
python3 plugins/mcp_server.py --stdio --record

# native simulator, recorded
python3 plugins/mcp_server.py --stdio --backend native --kernel firmware/uproc_demo/build/uproc_demo.elf --record

# enable writes / injection (never on by default)
python3 plugins/mcp_server.py --stdio --allow-write --allow-inject

# local loopback TCP instead of stdio (for clients that cannot spawn a process)
python3 plugins/mcp_server.py --tcp 127.0.0.1:8765 --record
```

Flags: `--stdio` (default), `--tcp HOST:PORT` (loopback only),
`--kernel ELF`, `--backend qemu|native`, `--record [PATH]`, `--name NAME`,
`--plugins PATH` (repeatable), `--allow-write`, `--allow-inject`.
Logs go to stderr; **stdout carries protocol frames only**.

### 2.2 Plugin mode (serve a live session)

```bash
./risc-v-gsp-rm-debugger.sh attach --kernel firmware/uproc_demo/build/uproc_demo.elf \
    --plugin plugins/mcp_server.py
risc-v-gsp-rm-debugger> mcp-serve                      # read/observe
risc-v-gsp-rm-debugger> mcp-serve --allow-write        # + memory writes / remote calls
risc-v-gsp-rm-debugger> mcp-serve --allow-inject       # + service injection
```

`mcp-serve` blocks until the MCP client disconnects or calls the `close`
tool, then returns to the REPL. In plugin mode `close` stops the server but
**does not** close the session (the REPL still owns it); in standalone mode
`close` also closes the launched session.

Both `--stdio` and `--tcp` behave identically. The bundled plugin directory is
loaded by default, so `mcp_server.py` is auto-discovered; the command only
serves when invoked.

## 3. Client configuration

The MCP server must be started **before** the session it should serve, so
client configs use standalone mode. Use absolute paths.

### 3.1 Claude Code / `.mcp.json`

```json
{
  "mcpServers": {
    "risc-v-gsp-rm-debugger": {
      "command": "python3",
      "args": [
        "/home/rmcore/Git/integdev_gpu_drv/open-iov/debug/risc-v-gsp-rm-vm/plugins/mcp_server.py",
        "--stdio",
        "--backend", "native",
        "--kernel", "/home/rmcore/Git/integdev_gpu_drv/open-iov/debug/risc-v-gsp-rm-vm/firmware/uproc_demo/build/uproc_demo.elf",
        "--record"
      ]
    }
  }
}
```

Or, one line: `claude mcp add risc-v-gsp-rm-debugger -- python3
<tool-root>/plugins/mcp_server.py --stdio --record`.

### 3.2 opencode (`opencode.json`)

```json
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "risc-v-gsp-rm-debugger": {
      "type": "local",
      "command": [
        "python3",
        "/home/rmcore/Git/integdev_gpu_drv/open-iov/debug/risc-v-gsp-rm-vm/plugins/mcp_server.py",
        "--stdio", "--backend", "native", "--record"
      ],
      "enabled": true
    }
  }
}
```

For an agent that must write memory or inject, add `--allow-write` /
`--allow-inject` to `args` and keep the session recorded (the recording is
the audit trail).

## 4. Protocol

| Request | Behaviour |
|---|---|
| `initialize` | returns `protocolVersion` (echoes `2024-11-05`, `2025-03-26` or `2025-06-18`; otherwise the server default), `serverInfo`, `capabilities.tools` |
| `notifications/initialized` | accepted silently (notifications never get a response) |
| `ping` | `{}` |
| `tools/list` | the 24-tool catalogue with JSON input schemas |
| `tools/call` | `{"name": ..., "arguments": {...}}` → `{"content":[{"type":"text","text":"<json>"}]}`, or `isError: true` for tool errors |

JSON-RPC errors: `-32700` parse error, `-32600` invalid request, `-32601`
unknown method, `-32602` unknown tool / bad params. Tool-level failures
(bad address, symbol not found, gating, capability absent) are returned as
`isError: true` results whose text is a JSON object with `error` and
`message` — the server never dies on a bad call.

Tool result payloads are always a single JSON text block so agents can parse
them deterministically.

## 5. Tool catalogue

| Tool | Arguments | Class | Notes |
|---|---|---|---|
| `session_info` | — | read | backend, PC, symbols, recording, breakpoints/watches, enabled capabilities |
| `symbols_resolve` | `target` | read | symbol, `symbol+0xNN` or hex address |
| `bp_set` | `target`, `kind=sw\|hw` | observe | hits recorded as `bp_hit` |
| `bp_clear` | `target` | observe | removes every breakpoint at that address |
| `watch_set` | `target`, `size=8`, `kind=w\|r\|a` | observe | hits recorded as `watch_hit` |
| `run` | `duration=1.0`, `max_stops`, `timeout` | observe | bounded; services hooks/watchpoints |
| `run_until` | `target`, `timeout=5.0` | observe | temporary breakpoint |
| `step` | `count=1` (max 10000) | observe | dispatches hooks/watchpoints |
| `read_mem` | `addr`, `size=64` (max 1 MiB) | read | lowercase hex |
| `write_mem` | `addr`, `hex` | **write** | requires `--allow-write`; recorded as `mem_write` |
| `hexdump` | `addr`, `size=64` (max 1 MiB) | read | symbol-annotated text |
| `snapshot` | `addr`, `size` (max 16 MiB), `label=mcp` | read | zlib + sha256 into the session |
| `call_function` | `target`, `args=[]`, `timeout=5.0` | **write** | requires `--allow-write`; Frida-like remote call |
| `inject_service` | `service`, `slot=6`, `load_addr`, `cmd=1`, `arg0`, `arg1`, `path=remote\|firmware\|both`, `settle=0.25` | **inject** | requires `--allow-inject`; service ELF built with `make sdk` |
| `trace_start` | `exec=true`, `start`, `end`, `max_events` | trace | native backend + recording session only |
| `trace_stop` | — | trace | flushes `blobs/trace.jsonl`, returns counts |
| `analyze_current` | `max_rows=40` | read | `render_summary` for the active recording |
| `compare_sessions` | `a`, `b`, `symbols=true`, `max_rows=40` | read | A/B recorded sessions |
| `export_hits` | `session`, `out` | read | Ghidra `greeneye_hits.tsv` |
| `probe_firmware` | `elf`, `instructions=2000`, `duration=1.0`, `no_relocate`, `symbols`, `timeout=300` | harness | QEMU-only probe subprocess; first divergence |
| `mutation_run` | `kernel`, `out`, `iterations=10`, `seed=1`, `timeout=0.25`, `settle=0.2`, `memory_region`, `allow_foreign_elf` | harness | `harness/mutation_harness.py`, QEMU-only |
| `portability_matrix_run` | `targets[]`, `out`, `probe_instructions=2000`, `scenario_duration=0.4`, `symbols`, `max_rows` | harness | `harness/portability_matrix.py`, QEMU-only |
| `plugins_list` | — | read | bundled plugins, `GREENEYE_PLUGINS`, REPL commands |
| `close` | — | session | stop serving; standalone also closes the session |

Capability classification follows the manifest in
[15-AGENTIC-GSP-IDE](15-AGENTIC-GSP-IDE.md) §5: read / observe are
unattended-safe; write and inject require the explicit startup flags and a
recorded session. `bp_set`, `watch_set`, `run` and `step` mutate target
*observation state* only and stay in the observe class.

Harness and probe tools shell out to the existing CLIs with bounded budgets
and return their stdout/stderr; they never touch hardware, a guest or a
kernel module.

## 6. Example agent sessions

### 6.1 Monitor an automated firmware-release run

```json
{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"session_info","arguments":{}}}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"symbols_resolve","arguments":{"target":"rpc_ping"}}}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"bp_set","arguments":{"target":"rpc_ping"}}}
{"jsonrpc":"2.0","id":4,"method":"tools/call","params":{"name":"run","arguments":{"duration":0.5,"max_stops":100}}}
{"jsonrpc":"2.0","id":5,"method":"tools/call","params":{"name":"analyze_current","arguments":{"max_rows":20}}}
```

The agent watches `bp_hit`/`watch_hit` counts in the report, snapshots the
control block, runs the release stimulus, then diffs the new recording
against the golden one:

```json
{"jsonrpc":"2.0","id":6,"method":"tools/call","params":{"name":"snapshot","arguments":{"addr":"gsp_ctrl","size":392,"label":"release"}}}
{"jsonrpc":"2.0","id":7,"method":"tools/call","params":{"name":"compare_sessions","arguments":{"a":"dumps/golden","b":"dumps/release","symbols":true}}}
```

### 6.2 Dump memory

```json
{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"hexdump","arguments":{"addr":"0x80000840","size":128}}}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"read_mem","arguments":{"addr":"0x80000840","size":16}}}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"snapshot","arguments":{"addr":"0x80000000","size":4096,"label":"ram-window"}}}
```

### 6.3 Inject a service

Start the server with `--allow-write --allow-inject --record`, then:

```json
{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"inject_service","arguments":{"service":"sdk/build/injected_demo.elf","slot":6,"cmd":1,"arg0":65,"arg1":66,"path":"both"}}}
```

The result contains the load address/entry and both the direct (`remote`) and
firmware-mailbox (`firmware`) return values; every segment write and call is
recorded, and the session can be compared like any other run.

## 7. Security rules

1. **Local only.** Default transport is stdin/stdout. `--tcp` binds loopback
   addresses only (`127.0.0.1`, `localhost`, `::1`); any other host is
   refused at argument parse time. There is no auth, so never expose it.
2. **No guest, no hardware.** The server drives the simulator/QEMU session
   only. It does not touch `/dev`, a kernel module, an mdev, a guest or a
   physical GPU; the agent runs against the simulator by default
   ([H7](../RISC-V-GSP-RM-DEBUGGER-ROADMAP.md), [15-AGENTIC-GSP-IDE](15-AGENTIC-GSP-IDE.md) §5).
3. **Read-only by default.** `--allow-write` enables `write_mem` and
   `call_function`; `--allow-inject` enables `inject_service`. The agent's
   client config decides which manifest class it gets; a read/observe agent is
   configured without those flags.
4. **Sessions are the audit trail.** Writes, calls, injections and snapshots
   land in `events.jsonl`; run with `--record` (the trace tools require it).
   A finding without a session is not a finding.
5. **Budgets everywhere.** `run` defaults to 1 s, `step` is capped, reads and
   snapshots have size caps, and harness subprocesses have timeouts. A wedged
   target is left stopped and restarted clean, never poked.
6. **Protocol hygiene.** stdout is reserved for JSON-RPC; diagnostics go to
   stderr. One bad tool call returns `isError` and the server keeps serving.
7. **No product changes.** MCP can propose evidence, not merge emulation
   changes; human approval still gates `open-iov-vfio`.

## 8. Absent capabilities

`probe_firmware`, `mutation_run`, `portability_matrix_run`, `trace_*`,
`inject_service`, `analysis` and the Ghidra export are checked at call time.
If a module/file is missing the tool returns:

```json
{"content":[{"type":"text","text":"{\"error\": \"not_available\", \"message\": \"mutation harness not present: ...\"}"}],"isError": true}
```

so an agent can plan around a capability instead of seeing the server crash.
The tool list is stable regardless of what is present.

## 9. Testing

```bash
python3 -m unittest tests.test_mcp -v     # 34 tests, no external deps
```

The suite drives `initialize`/`tools/list`/`tools/call` over in-memory
streams (including unknown method/tool, parse errors, write/inject gating,
no-stdout-pollution and structured `not_available`), plus a standalone
subprocess test on the native backend that skips cleanly when the native
simulator is absent.

## 10. Roadmap

* **R3 (done)** — MCP plugin exposing sessions as agent tools; local stdio,
  no network by default; capability gating.
* **H7 remaining** — named harness entry points and budgets, a session index
  for agents, enforced capability manifest files, and the physical-escalation
  gate. Those build on this server, not around it.