pidp10-mcp
pidp10-mcp
An MCP server for driving an ITS session on a PiDP-10 (simh KA10) emulator over its raw TCP terminal line.
It exists because generic telnet MCP servers do not work against this target: they cannot transmit raw control bytes, they tie the TCP connection's lifetime to tool-call cadence, and they reconnect dead sessions in the background — which, on a port that maps to a single terminal line, produces zombie connections fighting each other for it.
What it does differently
Raw control bytes get through. A
~-escape syntax insendputs exact bytes on the wire:~zfor the Ctrl-Z that calls ITS,~efor the ESC that DDT prints as$,~xNNfor anything else.The connection belongs to the server process, not to tool calls. A background reader drains the socket continuously into a 256 KB scrollback. Nothing cares how long the client spends thinking between calls; a three-minute gap is invisible to the session.
It never reconnects on its own. If the socket dies, the session is marked dead and the next tool response says so. Reopening is an explicit decision.
Closes hard.
close()setsSO_LINGERto zero so the socket is reset rather than left in a half-closed state that keeps the line marked busy, and the same teardown runs fromatexitplus SIGTERM/SIGHUP handlers.Terse responses. Only output produced since the last call, VT52 noise stripped, followed by one trailer line. No banners, no echoed inputs, no re-dumping the session log.
Install
Requires Python 3.11+ and the official mcp SDK 2.x.
uv sync # or: pip install -e .
uv run pidp10-mcp # stdio transport (default)Register it with an MCP client — for Claude Code:
claude mcp add pidp10 -- uv --directory /path/to/pidp10/mcp run pidp10-mcpor by hand, in an mcpServers config block:
{
"mcpServers": {
"pidp10": {
"command": "uv",
"args": ["--directory", "/path/to/pidp10/mcp", "run", "pidp10-mcp"],
"env": { "PIDP10_HOST": "pidp10.local", "PIDP10_PORT": "10018" }
}
}
}Streamable HTTP
uv run pidp10-mcp --http --http-host 127.0.0.1 --http-port 8010Configuration
Env var | CLI flag | Default | Meaning |
|
|
| Emulator host |
|
|
| Emulator TCP port (one line) |
|
|
| Bind address for |
|
|
| Bind port for |
open(host, port) can override the host and port per call.
Escape syntax
Escapes are expanded in send's input. An unknown escape is an error rather
than being passed through as text — silently sending ~q to DDT is worse than
a rejection.
Escape | Byte | Meaning |
|
| Ctrl-Z — calls ITS; a fresh line ignores all other input |
|
| ESC / altmode — DDT's |
|
| Ctrl-C |
|
| Ctrl-G |
|
| Rubout |
|
| Ctrl-S |
|
| Ctrl-O |
|
| CR, when an explicit one is wanted mid-line |
|
| LF — a DDT command (examine next location), not a newline |
|
| Tab |
|
| Any byte, two hex digits |
|
| Literal tilde |
| — | At the very end: do not append the automatic CR |
Line endings
send appends a CR (0x0D) automatically, because that is what the line
editor wants, and normalises any literal LF or CRLF in the input to CR. A bare
LF is not a newline on this system — it is a DDT command. If you genuinely
want to send one, ~n is exempt from normalisation.
raw=true sends the expanded bytes verbatim: no CR appended, no normalisation.
Tools
Tool | Purpose |
| Connect. Idempotent — reports status if already open. Returns any greeting bytes. |
| Send input, return the output it produced. |
| Collect more output without sending anything. |
| Re-show recent scrollback without moving the read cursor. |
| Connected?, host:port, uptime, bytes, pending output, death reason. |
| Hard-close the socket so the emulator frees the line. |
How send and read decide to return
They return on whichever comes first:
expect(a regex) matches the new output → reasonmatchedthe line has been silent for
quiet_ms→ reasonquiettimeout_mselapses → reasontimeout
The quiet timer only starts after the first byte arrives, so a program that
takes five seconds to say anything is not cut off at 700 ms. A call that sees
no output at all runs to timeout_ms and returns timeout.
With auto_more (default on), a trailing --More-- (Space=yes, Rubout=no)
prompt is answered with a space and collection continues, up to 20 pages; the
answered prompts are removed from the returned text. Hitting the cap returns
reason more_limit, and read continues from there. To flush a pager instead
of paging through it, send ~d (rubout).
send and read return only output produced since the last call. peek
does not move that cursor.
Output filtering
Raw output carries VT52 escape sequences, NUL padding and occasional telnet
IAC bytes. The filter drops NULs, ESC+letter sequences, ESC Y <row> <col>
cursor addressing, IAC negotiation (without implementing any telnet stack), and
other nonprinting bytes; it keeps text, tabs and newlines. CR, LF and CRLF all
become a single \n. Trailing whitespace and runs of blank lines are collapsed
in returned text to save tokens; the scrollback keeps the unabridged text.
Typical session
open()
send("~z", expect="Happy hacking|ITS") # ^Z calls ITS -> DDT banner
send(":login rms")
send(":listf") # pages collected automatically
send("foo~ej") # f-o-o ESC j CR (DDT's foo$j)
close()A detached but logged-in ITS session gets auto-logged-out after about five minutes; ITS itself never times out an idle line, so a held-open session is stable indefinitely.
Tests
uv run pytest # offline: filter, escapes, session, tools
PIDP10_LIVE=1 uv run pytest -m live # acceptance tests on a real emulator
PIDP10_LIVE=1 PIDP10_LIVE_SLOW=1 uv run pytest -m live # ...including the 3-minute idle testThe offline tests run the whole stack — including the MCP tool layer, via an in-process client — against a fake TCP line, so no emulator is needed.
Live tests are skipped unless PIDP10_LIVE=1; they take the single terminal
line for their duration. They honour PIDP10_HOST / PIDP10_PORT, plus
PIDP10_USER (default guest) and PIDP10_LISTF_DIR (default sys;, which
needs to be a directory big enough to make the pager appear).
The test_three_minute_gap_does_not_drop_the_session case sits idle for 190
seconds on purpose — it is the regression test for the failure that motivated
this server — so it needs PIDP10_LIVE_SLOW=1 as well.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/masto/pidp10-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server