# Feature Ideas — serial-mcp
> Trimmed from the original roadmap. Items already shipped (device identity,
> hot reconfigure, `get_status`, profiles, statistics, framing/packet decoder,
> AT/JSON/shell protocol parsers, auto-reconnect, richer matching modes
> regex/glob, event log, native_sim diagnostics, `--version`, the 7 protocol
> presets) were removed. See [CHANGELOG.md](../../CHANGELOG.md) and the README
> tool list for what landed.
>
> **Protocol roadmap** (HDLC/PPP framing, Modbus RTU, MIDI, Firmata — all
> deferred) is tracked per-protocol in
> [protocol-matrix.md](protocol-matrix.md); this file tracks everything else.
>
> Priorities: **Near-term** · **Later** · **Wish** · **Needs architecture review**.
> Non-feature work (refactors, CI, release mechanics) lives in
> § Infrastructure / tech debt at the bottom.
## Near-term
### Local-only usage statistics for development
- collect local metadata on tool-call frequency, stop-reason
distribution, option-usage frequency (which `from` variants agents
actually pick, how often `match`/`framing` options get used, average
`max_buffered_bytes` actually configured, etc.) to
drive evidence-based decisions on which options to keep, trim, or default
differently
- strictly local: write to a file on the host (e.g. under
`~/.local/share/serial-mcp/` or a configured path), never transmit
over the network, no telemetry, no remote endpoint, opt-out by
default with an explicit enable
- the goal is development insight (which tools/options are dead weight,
which defaults are wrong), not user tracking — design the schema
around questions we actually want to answer ("is `from: {"type":"now"}` used
enough to justify keeping it?")
- pairs with the shipped `configure` tool + connection-default trim (the
stats inform which fields to cut, and cutting fields makes the remaining
stats cleaner)
### Declarative checksums on generic framing
- a `checksum: { algorithm, ... }` option on `Delimiter` / `LengthPrefixed` /
`StartEnd` rx_framing/tx_framing — generalizes what the NMEA/Modbus presets
hardcode
- covers the long tail of proprietary vendor protocols without building the
full plugin API (see "External decoder/plugin API" below — this is the
lighter first step)
- natural follow-on to the checksum-helper refactor that landed in 0.7.3
### TX pacing / throttling
- inter-chunk or inter-line delay on `write` (per-call field + connection
default)
- bootloaders, GRBL controllers, and cheap AT modems drop bytes without flow
control; today an agent must hand-loop small writes with sleeps it cannot
actually perform
### Modbus ASCII TX auto-LRC
- TX-side counterpart to the shipped RX LRC validation: hex-encode a binary
PDU and append the LRC on write (`:` + hex + LRC + `\r\n`)
- deliberately split out of the NMEA TX auto-checksum work that landed
in 0.7.3 — needs hex-encoding of a binary payload, not just a
checksum append
- **refactor trigger (one-consumer rule):** when this lands, extract a shared
TX checksum-append layer instead of growing `TxFramingMode` variant-by-
variant. The `TxFramingMode::Nmea` encode arm (shipped) is the first
checksum-appending TX mode; Modbus ASCII TX would be the second and should
share the "compute checksum → validate-or-append → wrap with markers"
skeleton so a third (CRC-16 / FCS-16) is a one-line diff.
- **refactor trigger (one-consumer rule):** when CRC-16 (Modbus RTU) lands,
make `emit_frame`'s per-frame validation policy pluggable. Today
`emit_frame` hard-codes the `ChecksumMismatch` → drop-and-count behavior
for all parsers; a multi-checksum world (XOR/LRC now, CRC-16/FCS-16 future)
wants the validation outcome (drop+count vs. stream-fatal vs. emit-with-
flag) driven by the parser's declared checksum width, not a match on the
error variant.
### Config import/export
- likely pairs with profiles (already shipped)
- needs sharpening before it earns implementation: the profiles TOML file
already covers copy-between-machines; the added value would be exporting a
*running* server's full state (open connections + their framing/parser
defaults) as importable profiles
### External decoder/plugin API
- useful after the in-process decoders (AT/JSON/shell) shipped
- allow plugging in custom frame decoders / parsers
- prefer shipping "Declarative checksums on generic framing" (above) first —
it covers much of the demand at a fraction of the API surface
### Decoder integration / export hooks
- export capture or frames to external decoder tools if in-process support stays small
## Later
### MCP Tasks extension for long-running operations
- SEP-2663 / `io.modelcontextprotocol/tasks`: let a tool return a task handle
immediately, then expose `tasks/get`, `tasks/update`, and `tasks/cancel` with
cooperative cancellation and bounded result retention/TTL
- candidates: long `read`, `transact`, `capture_boot`, `send_break`, and any
future firmware/file-transfer operation; keep normal synchronous execution
for clients that do not declare Tasks support
- not a replacement for open-ended RX availability notifications: current
rmcp clients must poll task state and task-status delivery through
`subscriptions/listen` is not available yet
- needs ownership/lifecycle design for connection close, client disconnect,
server restart, task expiry, profile learning, and partial serial results
### Positive MCP cache TTL policy
- future feature only: the 2026-07-28 cache fields already ship, but as the
non-cacheable baseline (`ttlMs=0`, `cacheScope: private`) on every cacheable
family — enabling positive TTL is NOT shipped and is what this item proposes
- possible policy if pursued: long/public for static tool, prompt, and
resource-template catalogs; short/private for live port lists; zero/private
for open connections, logs, status, and RX data
- only enable positive TTL after resource/list notification invalidation,
authorization partitioning, pagination keys, and rmcp's stale-on-error client
behavior have public-boundary tests
### Standard HTTP parameter headers
- MCP 2026-07-28 + rmcp 3 automatically provide `Mcp-Method` and `Mcp-Name`;
selected primitive tool inputs can later opt into `Mcp-Param-*` through
top-level `x-mcp-header` schema annotations
- possible low-risk first field: `connection_id`; assess `port` and `profile`
separately because proxies commonly log headers
- never promote commands, serial payloads, match data, selectors, credentials,
or capture filenames into infrastructure-visible headers
### Flow-control-aware ring backpressure (pause-on-full)
- follow-up to the 0.8.0 RX ring redesign: the always-on pump
drains the kernel buffer continuously, so with RTS/CTS enabled the kernel
never deasserts RTS and the device is never throttled — sustained unread
traffic wraps the ring (oldest bytes lost, observably via `bytes_lost`)
- opt-in per-connection mode (`on_full: "wrap"` (default) | `"pause"`): when
the ring is full and RTS/CTS is active, pause the pump instead of
overwriting — the kernel buffer fills, RTS deasserts, the device pauses,
and hardware backpressure semantics are restored end-to-end
- emit a log event + MCP notification on entering/leaving the paused state
so agents know the device is being held off rather than silent
- only meaningful with hardware flow control; with none enabled, pausing
just moves the loss from our ring to the kernel buffer/UART FIFO
### Persistent per-connection framing decoder
- deferred from the 0.8.0 RX ring redesign: framing is per-call
and applied to the drained ring window, so a frame torn at a call boundary
(partial drain, ring wrap) decodes as garbage or a SLIP error
- carrying decoder state across `read` calls fixes that, but requires
binding framing config to the connection rather than the call and
rethinking its place in the 4-layer precedence model — needs design
### Per-client RX cursors
- deferred from the 0.8.0 RX ring redesign: multiple HTTP
clients share the one consuming read cursor, so concurrent consuming reads
interleave their drains
- offsets in every result already make this diagnosable; the fix, if shared
multi-agent access becomes real, is named cursor groups (one cursor per
client) — overlaps with "Socket sharing / tee" below
### Baud-rate auto-detection *(deferred)*
- classic serial-tool feature and a strong agent diagnosis step for
"unknown device on /dev/ttyUSB0": try candidate rates, score the RX per
rate, return ranked guesses
- deferred: generic host-side detection over an ordinary USB-serial adapter
is heuristic, not waveform measurement — the adapter's UART already
re-clocked the bits at the configured rate, so the host sees decoded
garbage that can score deceptively well at the wrong rate; a built-in
tool should return **inconclusive** rather than guess
- existing solution/reference worth studying before building anything:
EXPLIoT's `uart.generic.baudscan`
([repo](https://gitlab.com/expliot_framework/expliot),
[baudscan.py](https://gitlab.com/expliot_framework/expliot/-/blob/master/expliot/plugins/serial/baudscan.py))
cycles candidate rates, reads a bounded sample per rate, and ranks each
by printable-ASCII percentage (default rates
1200/2400/4800/9600/19200/38400/57600/115200; accepts 100% or the best
rate above 90%) — useful for continuously emitting ASCII devices, weak
for silent or binary devices
- reference only: no dependency, integration, or adoption implied
### Modem input lines + UART error counters in `get_status`
- read CTS/DSR/CD/RI (the serialport crate exposes them) — today the server
can *set* DTR/RTS but cannot *read* any input line
- parity/framing/overrun error counters where the platform provides them
- cheap additive fields on `ConnectionStatus`
### Per-frame timestamps
- `Frame` carries an index but no time; correlating frames against the event
log or across two connections is currently impossible
- small additive field, but it changes the wire format — decide before 1.0
### GRBL / G-code preset
- line-based `ok`/`error` protocol; popular real-world target (CNC, laser,
3D-printer controllers)
- becomes nearly free once TX pacing lands — the `transact` tool already
ships; implement after pacing
### Safety policies for dangerous commands
- optional dangerous-command confirmation patterns
- includes the profile-level safety policy intent carried over from the
removed `ProfileDefaults.safety_policy` field
### Capture bookmarks / annotations
- useful if logs/captures grow further
### Expect/script automation *(needs architecture review)*
- interesting, but discuss at a higher architecture level before adding
- can become huge if designed poorly; can conflict with simpler read/write model
- conservative first design if pursued: JSON transaction steps only, bounded
step types, no shell access, deterministic transcript output
- the shipped `transact` tool is the minimal kernel of this — revisit
whether scripting is still needed
### Filtering/search across captures
- unclear value — maybe an LLM can use grep/glob instead
- worth it only if searches include direction, timestamps, event types, parsed fields
### Recording + replay
- useful but niche: reproducible bugs, test fixtures from real hardware,
decoder/parser regression tests
- **The safe persistent capture foundation shipped** (disabled-by-default
`--capture-dir` store, portable filename-only `export_log`, quotas,
no-overwrite atomic commits, advisory locks) — a prerequisite, not the
feature. Continuous raw capture lifecycle is specified (NOT implemented)
in [safe-continuous-capture-design.md](safe-continuous-capture-design.md);
recommendation: do not implement until concrete task evidence.
### RS-485 options
- half-duplex bus semantics, direction control timing, RTS-based send control
- needs a new testing strategy/firmware
### RFC2217 backend support
- server opens a remote serial device over network with control signals
- backend transport feature, not MCP transport replacement
### Bridge mode
- proxy observation, reverse engineering, test harnessing
- very complex
### MCP Bundle (MCPB) distribution
- package existing native release binaries as an MCP Bundle for one-click local
stdio installation in supporting desktop clients
- promising fit: `server.type = "binary"`, no language runtime, platform-specific
command overrides, and optional user configuration
- separate release/distribution project, not protocol work; decide one
cross-platform bundle versus per-platform bundles, manifest version, signing,
update flow, and clean-machine Claude Desktop tests before implementation
- validate/pack with a pinned `@anthropic-ai/mcpb` release when scheduled
## Wish
### Earlier MCP protocol revisions (pre-2025-11-25)
- potential future feature only — NOT current support and not a near-term
commitment: the supported set remains exactly `2026-07-28` (preferred) and
permanent `2025-11-25`;
- possible older candidates are `2025-06-18`, `2025-03-26`, and `2024-11-05`;
- implement only with concrete user/client demand — never merely because
rmcp lists the version in `KNOWN_VERSIONS`;
- each version would require: an explicit product policy row,
lifecycle/capability/cache review, raw-wire tests, official conformance
support where available, and a real historical client fixture.
### User-facing loopback / virtual port backend
- expose a virtual echo/scripted device as an openable backend (the
native_sim harness exists but is test-only)
- lets agents demo and develop flows with no hardware attached
### Socket sharing / tee / shared live access
- not the HTTP MCP transport — this exposes a live serial stream/session to
another consumer
- seems complicated; keep as future wish
### File transfer protocols
- do not turn project into a full DFU/flashing suite
- only consider generic serial-native transfer helpers if ever added
### Non-intrusive sniffing / proxy observation
- most realistic path: proxy/bridge observation, not universal passive sniff
### Human + agent shared session / tee mode
- overlaps with socket sharing
## Explicit skip for now
- **MRTR product flows** — rmcp 3 supports SEP-2322 multi-round tool/resource/
prompt requests (`InputRequiredResult`) for client elicitation and retries,
but current schemas, defaults, destructive hints, and cancellation cover the
serial workflows we have. Revisit only for a concrete need such as physical
power-cycle guidance or destructive reset confirmation; any echoed
`requestState` must be integrity-protected.
- **Remote monitor** — skip, keep off active roadmap
- **SECURITY.md / vulnerability disclosure policy** — not important at the
current project size; revisit if outside contributors arrive
## Infrastructure / tech debt
Non-feature work, roughly in suggested order. From the 2026-07-05 repo review.
### `UInt` newtype to kill schemars `uint_schema` boilerplate
- per-field `#[schemars(schema_with = "crate::schema_helpers::uint_schema")]`
annotations are sprinkled across the tree, with a documented
regression history (b12b09fd, bc37a0b0, PortInfo miss) — missing one
is a known bug vector
- schemars 1.x emits non-standard `"format": "uintN"` for unsigned
integer fields; the per-field attribute is the workaround
- `#[serde(skip_serializing_if = "Option::is_none")]` fields must also carry
`#[serde(default)]`: schemars 1.2.2 does not see through `schema_with` to
the `Option` type, so without it the field lands in `required` while
serialization omits it (the PortInfo vid/pid/interface miss)
- a `UInt` (and `OptionUInt`) newtype that derives `JsonSchema` natively
without the format keyword, OR a schemars visitor that strips the
format globally, would collapse the entire class of bug and remove
the per-struct `check_schema!` maintenance burden
- coordinate with any schemars 2.x migration if one is on the roadmap;
the upstream fix may make the newtype redundant