Files
HermenixandClaude Opus 5 10176356b6 docs(mcp)!: mcp-v4, a second transport, and what its token is not
The contract said "stdio, and only stdio" and now describes two
transports, so it gets a version -- **v4**, because v3 (the live stream)
was published to main while this branch was in review.  Nothing here
breaks: no tool renamed, no argument reshaped, no resource moved, so a
client written against v2 over stdio still works unchanged.  v2 is the
only bump that took something away, and it took away five tools; v3 and
v4 are additive.

§1 is rewritten around the asymmetry rather than around the feature,
because the asymmetry is the part that is easy to get wrong later: stdio
is unauthenticated **on purpose** and that is now written down with its
argument, so the next reader does not "fix" it.  A stdio client spawns
the server, so it has already passed a filesystem permission check and
holds the broker's 0660 control socket either way -- unix permissions are
the gate, exactly as `mcp.enable`'s own option text has always said.  A
port is the first time the caller passed neither, which is the entire
reason the token exists.

§2 gains `HHSTREAM_MCP_STATE_DIR` (and why it is NOT `sessions.stateDir`:
that directory is 0700 owned by the broker's user and holds the audit
trail), the `serve --http` flags, and the `token` verbs with the three
decisions behind them -- digests not tokens, wall-clock expiry where
every broker deadline is BOOTTIME, and a flock because two processes
write the table.  §3 records that a token's toolsets intersect the
server's and cannot widen them.

§10 is new and is the part worth having: **what this is not.**  Not OAuth
2.1, not RFC 9728, no metadata discovery -- so a client that can only
*discover* an authorization server will not connect, and one that lets
you set a header will.  And a token narrows tools but not resources, so a
`frames` token can still read `hhstream://sessions`.  Both are now TODO
files with their own evidence and recommended routes, beside the two
limits the SDK imposes:

* `mcp-http-auth-is-not-oauth.md` -- medium, and the severity is about
  interoperability rather than exposure.  Nothing is weaker than it
  claims; what is missing is the part that lets a client somebody else
  configured connect without a human pasting a header.  It records that
  `ssh host hhstream-mcp` is genuinely the better answer for one operator
  with one machine, because that is what this competes with.
* `mcp-tokens-do-not-scope-resources.md` -- medium, and nothing at all
  until a second token is minted for someone else, at which point the
  restriction is narrower than it reads.

README gains the HTTP section with the three things to know before
enabling it (the unit's refusal and self-heal, no TLS, not OAuth);
README.dev gains the token commands in the dev-shell walkthrough and the
new check's row; ARCHITECTURE gains the two transports in its diagram and
the reason the shim has a state directory while owning no state about the
desktop.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KchcvnbfhhZhUHy83YL3Ya
2026-10-12 18:00:05 +00:00

39 KiB

CONTRACT — the MCP server (mcp-v4)

This file is the interface of hhstream-mcp: the resource URIs, the tool names, the arguments each takes, what comes back, and which of the CLI's exit codes arrives as which kind of MCP error. It is the sibling of docs/CONTRACT.md (the frame socket's wire format) and docs/CONTRACT-sessions.md (the broker, the containers and the input lease).

Immutable without a version bump. An agent that has learned these names and these argument shapes is a consumer like any other, and renaming a tool breaks it exactly as renaming a wire verb would.

Version: mcp-v4 (2026-10-12). v1 was the first freeze and lasted one day; v2 split state out into resources and removed five tools; v3 adds the live stream; v4 adds a second transport. v3 and v4 are additive — every v2 name still means what it meant.

v4 changes: a second transport, and the credential it exists for

§ Change Why
§1 A second transport: Streamable HTTP, opt-in, behind a bearer token. stdio stays the default and stays unauthenticated. A token means nothing on stdio — the client spawns the server, so it has already passed a filesystem permission check and holds the broker's 0660 socket either way. A port is the first time the caller is someone who did neither, and the transport and the credential are therefore one feature.
§2 serve --http and its flags; token mint|list|revoke|prune beside it. The bare hhstream-mcp and hhstream-mcp --toolsets=… forms are unchanged. A client's command entry is a configuration someone else wrote — an HTTP endpoint is a configuration someone else hosts.
§3 A token may carry its own toolsets, and the intersection with the server's --toolsets applies. --toolsets is a launch-time flag, and one HTTP process serves many callers: a per-process tool budget stops being a per-caller one the moment the transport is shared.
§10 What the HTTP transport's token is not: not OAuth 2.1, not RFC 9728, no metadata discovery. A limit worth stating in the same register as the two the SDK imposes.

v3 changes, all of them the per-session live stream (docs/CONTRACT-sessions.md §4, v4):

§ Change Why
§3 Tools session_stream_enable and session_stream_disable. 21 tools became 23. Switching a stream on is an act, and a consequential one: it publishes a desktop on the LAN behind a phrase. Nothing about it is a read.
§3 session_start gains stream, stream_phrase and stream_fps, and its reply carries the URL. A stream needs the container output that starting creates, so sequencing start-then-enable in the caller means every caller repeats the ordering and the half-done case. One call, one answer, and the phrase comes back with the session.
§4 Resources hhstream://streams and hhstream://sessions/{id}/stream. 8 resources became 10. What is being published, and who is watching, is state — so it is read, not called.
§6 A connected viewer is the one exception to the idle-clock rule. About once a minute the broker stamps touch on any session somebody is watching. It is not a verb and cannot be invoked; it is the broker noticing that a person is looking.

v2 is a breaking change, and the only one: the verbs that merely read are resources now, not tools.

§ Change Why
§3, §4 status, session_list, session_status, session_logs and pointer_state are removed from the tool list and are hhstream://… resources instead. 26 tools became 21. The protocol's own split. A resource is application-controlled context — the host decides what to put in front of the model — and "what is on this machine right now" is the host's question. A tool is model-controlled: an act the model chooses. Nothing in those five acts.
§4 Eight resources: five fixed, three templates, with query parameters for the arguments the verbs took. A resource read takes nothing but a URI, so --lines becomes ?lines=. Only declared parameters are accepted.
§5 A refused resource read is a protocol error carrying the same remedy a tool error would have. CallToolResult has isError; ReadResourceResult has nothing. Returning the broker's failure object as the body would make "no such session" read as a successful read of a session.
§3 --toolsets selects which families are listed as tools. A tool listing is a budget. MCP has no grouping primitive — the Tool object has nowhere to put one — so the only lever is which tools the server lists, chosen at launch.

What did not move, and the reasons are the interesting part:

  • shot and record are still tools, though neither changes the desktop. Capturing is an act at a moment: an agent in a loop has to be able to decide to look now, and in many clients a resource is something a person attaches rather than something a model can fetch on its own initiative. The same frames are also readable as hhstream://screen and hhstream://outputs/{name}, for a host that wants to put one in context itself.
  • session_container is still a tool, for a harder reason: with an id the CLI runs container.reconcile, which re-reads the compositor and writes the record back. It is not a read at all. Only the no-id form is a resource (hhstream://containers). A resource read must never write.

0. What this is, and what it is not

It is a shim. Each tool and each resource builds the argv list the hhstream CLI's own parser takes, hands it to the CLI's own handler in this process, and maps what came back. Nothing in it decides anything:

Rule Whose it is
the idle clock, and that only touch resets it the broker
the pointer lease: who holds it, its TTL, what it lifts the broker
the allowlist, and which arguments an app may be given the broker's policy
the exit-code table docs/CONTRACT-sessions.md §4
the refusal messages (reason, hint) the CLI and the broker
coordinate normalisation, the scroll bound, the type cap the broker and the driver

What this layer owns is the argv list, the routing of a URI to one of them, the mapping of a result to MCP content, and one sentence of remedy per exit code — the "what to do next" that the CLI says to a human in the words of a command line and an agent needs in the words of tools and resources.

In process, not as a subprocess: hhstream.cli is importable Python whose parser already takes an argv list and whose handlers already return a code from the frozen table. Two of its paths cannot be called here, and both are documented in hhstream_mcp.invoke: cli.main's sg re-exec, and cmd_run's execvpe. Consequences of the first are in §5.

1. Transport

Two, and stdio is the default.

transport how authentication
stdio hhstream-mcp, or hhstream-mcp serve. One process per client, spawned by the client; no port and no daemon. None, deliberately. See below.
Streamable HTTP hhstream-mcp serve --http, or the mcp.http.enable unit. One process, many clients, one port. A bearer token on every request, or 401.

Why stdio is unauthenticated, and why that is not an oversight

The stdio server is spawned by its client. To do that the client already had to pass a filesystem permission check, and whoever can run hhstream-mcp can run hhstream — which talks to the same 0660 control socket owned by sessions.user, in the same group. Unix permissions are the gate, and mcp.enable's own option text has always said so: it "grants no access the hhstream CLI does not already grant to whoever runs it". A token checked by a process the caller started is a lock whose key is in the same pocket.

Which is also the whole argument for the HTTP transport existing. A port is reachable by a caller who passed no filesystem check, and at that moment a credential is the only gate there is. So the two arrived together: v1 and v2 said "stdio, and only stdio" because, for what the server then was, that was the honest answer.

stdio: the protocol is not on file descriptor 1

Once the server is running: the tools run the CLI in process, the CLI prints, and record spawns an ffmpeg whose output the CLI does not capture. The server takes a private duplicate of stdin and stdout at start-up, then points fd 1 at its log and fd 0 at /dev/null. A stray write is therefore a log line and never a corrupted stream.

This is also why token mint prints a secret on stdout and is safe: the token verbs return before anything claims the transport, which tests/unit/test_mcp_authorizations.py asserts with a claim_transport that fails the test if it is reached.

HTTP: what is on the wire

  • Authorization: Bearer <token> on every request. The scheme is compared case-insensitively (RFC 9110). Anything else — absent, malformed, wrong scheme, unknown token, revoked token, expired token — is one answer: 401, an empty body, WWW-Authenticate: Bearer realm="hhstream", Cache-Control: no-store. The token is never echoed, in the body or in a header, and the response does not distinguish which of those it was — telling them apart would tell someone who guessed a real id that they had.
  • Failed attempts are rate-limited per client, reusing hhstream_live.server's RateLimiter; past the limit the refusal is 429.
  • The token is checked before the Host. An unauthenticated caller learns nothing about the rebinding allowlist, which is itself a hint about how the server is deployed.
  • Any path but the mounted one is a 401 without a token and a 404 with one. An unauthenticated caller does not learn what is mounted, or that anything is.
  • DNS-rebinding protection is on. The Host allowlist is derived from the bind address when none is given, because the SDK's check is allowlist-only with no implicit localhost — protection on with an empty list 421s every request. A wildcard bind derives nothing and is refused at launch, which makes binding every interface the one case where the hostnames must be named.
  • No TLS. The token crosses the wire in clear text. Loopback is the default; anything else wants a reverse proxy and says so in a module warning.
  • Tool calls are serialised across every caller, not per session. The Runner holds a process-wide lock because the CLI it calls prints to a stdout this process shares — invisible while each client had a process of its own, visible now that one process serves many. Two agents on two tokens queue behind each other, and a slow call delays everyone's next one rather than only its own caller's. Latency, never a hang: the event loop answers a ping throughout. See docs/TODO/mcp-in-process-has-no-call-deadline.md.

2. Configuration

There is no --socket flag. The sockets are the CLI's own environment, read by the CLI:

Variable What it names Default
HHSTREAM_SOCKET the broker's control socket /run/hhstream/control.sock
HHSTREAM_FRAME_SOCKET the published frame socket /run/hypr-stream/stream.sock

Set them in the MCP client's env block. One place to configure them is one place for them to be wrong.

The HTTP transport reads one more, and hhstream-mcp token reads the same:

Variable What it names Default
HHSTREAM_MCP_STATE_DIR where the authorization table lives /var/lib/hhstream-mcp

Not sessions.stateDir. That directory is 0700 owned by the broker's user and holds events.jsonl, the audit trail; the MCP server runs as whichever user's client spawned it, and parking a token file there would mean granting this shim write access to the audit trail's directory for no reason. The two also have different lifetimes: revoking every token must not be able to touch a session record.

The server's own command line. hhstream-mcp with no subcommand, and hhstream-mcp serve, are the same thing — the bare form is not deprecated and will not be, because an MCP client's command entry is a configuration somebody else wrote:

flag where what it does
--toolsets NAMES either comma-separated families to offer as tools: frames, sessions, pointer, keyboard. Omitted, every tool is offered. An unknown name is argparse's exit 2, never a silently empty server. Resources are not filtered by it — they are the host's context, not the model's tool budget, and making hhstream://sessions conditional on the session tools would couple two decisions that are not the same decision.
--list-tools either print the table and exit. Honours --toolsets, so it is also how to see what a selection would offer.
--list-resources either print the resource URIs and templates and exit.
--http serve Streamable HTTP instead of stdio.
--address HOST serve default 127.0.0.1. Loopback, because reaching the LAN is a second decision.
--port N serve default 9989, one past the live stream's 9988.
--path P serve default /mcp.
--state-dir DIR serve, token overrides $HHSTREAM_MCP_STATE_DIR.
--allowed-host HOST serve repeatable. Omitted, derived from --address; a trailing :* is a wildcard port.
--allowed-origin ORIGIN serve repeatable. Nothing needs this unless a browser is the client.
--allow-no-tokens serve start although the table is empty. Without it that is a usage error, because a listener that 401s every request is almost always a forgotten token mint.
--stateless serve no session to pin, for a deployment behind a load balancer.
--json-response serve one JSON body instead of an SSE stream.

hhstream-mcp token

verb what it does
token mint [--label L] [--expires-in D] [--toolsets NAMES] print a new token once, on stdout, with the prose for a human on stderr beside it. D is 30s/5m/3h/90d or bare seconds.
token list [--json] the table. The tokens are not in it and cannot be.
token revoke <id> | --all by id, so ending a grant never means pasting a live secret onto a command line that lands in a shell history.
token prune drop expired grants. Housekeeping only — an expired token is refused whether or not it is still on disk.

The table holds SHA-256 digests, not tokens, so authorizations.json is not itself a credential; a token that was not written down can only be revoked and replaced. No salt and no KDF: a token is secrets.token_urlsafe(32) behind an hhsmcp_ prefix — 256 bits, so there is no dictionary to attack and a work factor would only slow the request path. The prefix is there so a leaked token is identifiable on sight, which is the convention secret scanners are built around.

Expiry is wall clock, and it is the one place in this project where that is right. Every deadline in the broker is CLOCK_BOOTTIME (hhstreamd.clock) because an idle window has to survive an NTP step and count time spent suspended. "Valid for thirty days" is a calendar statement that has to outlive a reboot, and BOOTTIME resets to zero there.

Every read-modify-write takes an exclusive flock on a sidecar lock file. Two processes write this table — the server stamping last_used, and an operator running token mint — and the atomic replace alone prevents a torn read, not a lost update.

These verbs sit outside the frozen exit-code table in hhstreamd.protocol, because nothing they do reaches the broker and no agent calls them. Two of the codes are chosen by analogy with it anyway: 0 done, 2 usage, 3 no token with that id (the analogue of NO_SUCH_SESSION), 4 the store itself is unusable (the analogue of BAD_STATE).

$HHSTREAM_MCP_LOG sets the log level on stderr. Nothing else affects behaviour; in particular the server holds no state and caches no lease.

On the HTTP transport a token may carry its own toolsets, and the intersection with the server's own applies — a token cannot widen what the unit was started with. That exists because --toolsets is a launch-time flag and one HTTP process serves many callers: a per-process tool budget stops being a per-caller one the moment the transport is shared, and the token is then the only place the distinction can live. A tool the caller's token excludes is not listed and cannot be reached by name; the refusal names both possible narrowings, because the agent cannot tell them apart from where it stands and neither is its to change.

There is no grouping primitive in MCP, and --toolsets is not pretending to be one. The Tool object carries name, title, description, inputSchema, outputSchema, icons, annotations, _meta and execution — nowhere to put a family. Tools are a flat list in the specification too. So the only honest lever is which tools the server lists, decided by whoever spawns it, and the families below are the grouping that is left: a prefix in the name and a line in this table.

3. The tool surface

One tool per CLI verb that acts, named after it, with _ where the CLI has a space. Twenty-one tools, in four families. The granularity rationale — and the alternative that was rejected, one tool per family with an action enum — is in hhstream_mcp.tools' docstring.

Every schema sets additionalProperties: false, so an argument this server would not pass through is refused by the schema rather than dropped on the way to an argv list.

family tools
frames shot, record, run
sessions session_start, session_touch, session_kill, session_gc, session_adopt, session_container, session_reconcile, session_stream_enable, session_stream_disable
pointer pointer_acquire, pointer_move, pointer_click, pointer_scroll, pointer_down, pointer_up, pointer_release
keyboard keyboard_type, keyboard_key, keyboard_down, keyboard_up

Every tool is in exactly one family — asserted by a unit test, because a tool in none of them would be one --toolsets could never offer.

Frames

Tool Arguments Returns
shot output? image content (PNG or JPEG, by the bytes' own magic) plus text naming the geometry, the media type and the byte count
record seconds?, out?, output?, fps?, work_dir?, keep? the artefact's path and the measured rate — never the video
run argv, settle_s? the application's pid, and that it is unmanaged

shot captures to a temporary file and reads it back, because that is the CLI's only output shape. The media type comes from the bytes and not from the extension: the publisher serves image/jpeg on some machines and hhstream shot writes the payload it was given verbatim.

record hands back a path because an mp4 of a 720p desktop is megabytes and a tool reply is a message in a model's context. The rate is the one the capture measured, read off the CLI's own line.

run is spawned with posix_spawn into its own session and not waited for. settle_s (default 1.5) is how long the server watches it before reporting the pid: a launch that fails — no Wayland socket, no such executable — fails inside that window and comes back as an error, one that survives it comes back as running.

Session lifecycle

Tool Arguments
session_start app, argv?, idle_timeout?, wait_ready?, home?, ephemeral_home?, stream?, stream_phrase?, stream_fps?, request_id?
session_touch id
session_kill id, grace?, signal?, request_id?
session_gc request_id?
session_adopt pid, app?, idle_timeout?, request_id?
session_container id
session_reconcile id?, request_id?
session_stream_enable id, phrase?, fps?
session_stream_disable id

Every one of these runs with --json, so what comes back is the broker's own protocol object: in the text part verbatim, and in structuredContent parsed. The human rendering is a table for a terminal; scraping it would be inventing a second wire format.

home and ephemeral_home are mutually exclusive, and the schema says so — the CLI's argparse group, moved forward to where the refusal is free.

session_stream_enable returns the URL on the first line of its text part, because that is the whole point of the call: it is what gets handed to a person. It publishes the session's own container output and never the main one — that output carries every other session, and putting it behind one session's phrase is the mistake the broker refuses to make (exit 10 with no container). Re-enabling returns the same URL; a different phrase is exit 8, not a silent rotation, because rotating under somebody holding the old URL is the one outcome neither answer can be mistaken for. session_stream_disable is marked destructive: it forgets the phrase, so anyone holding that URL loses it.

session_start with stream: true is one call rather than two, and a stream that cannot be opened does not fail the start — the application is on the screen, and killing it over a web page would be the wrong trade. The reply carries stream_error whenever stream was asked for, null when it worked.

session_container is not marked read-only, and that is not an oversight: with an id the CLI runs container.reconcile, which writes the record back. That is also why it is a tool rather than a resource. The no-id listing, which only reads, is hhstream://containers.

Input — the pointer lease is an argument, never a cache

Tool Arguments
pointer_acquire id, ttl?, no_restore?
pointer_move id, lease, exactly one of to / by (each [x, y])
pointer_click id, lease, button?
pointer_scroll id, lease, axis?, amount?
pointer_down id, lease, button?
pointer_up id, lease, button?
pointer_release id, lease
keyboard_type id, lease, text
keyboard_key id, lease, name
keyboard_down id, lease, name
keyboard_up id, lease, name

lease is required on every gesture. The CLI caches the lease id between invocations (hhstream_session.lease) because its surface is flagless and it is one process per verb; this server passes --lease explicitly on every call and so never reads that cache. The reason is worth stating in full: a cached lease is a lease an agent cannot see, and the failure it produces — a gesture that went to a lease the agent did not know it held — is not one a transcript can explain.

pointer_acquire's reply puts lease_id on the first line of the text part, because that is what the next call needs.

Who holds the lease is a read, so it is hhstream://pointer rather than a tool. That is the resource to read after an exit 8.

Durations (ttl, idle_timeout, wait_ready, grace) are strings in the CLI's own vocabulary: 30s, 5m, 3h, 2d, or a bare number of seconds. They are passed through untouched, so the parsing and the refusal are the CLI's.

Flag form

Every option value is passed as one --name=value token, and a literal -- precedes every positional. Not a style choice: --by -0.1,0.5 and a text beginning with a dash are both values argparse reads as option strings, and one uniform rule is one fewer thing to get wrong per flag.

4. The resource surface

Five fixed resources and three templates, all under one scheme, because they are all views of one machine's one desktop. Every one of them runs the same CLI verb a tool would have, with --json where the verb has it.

URI Body Query
hhstream://status text/plain — the CLI's own report: both compositor units, the frame socket's path and whether this server can read it, the Wayland socket, one probe frame. hhstream status has no JSON form, so neither does this. —
hhstream://sessions application/json — the broker's list object ?all=true includes tombstones
hhstream://pointer application/json — the holder, since when, and the buttons and keys held —
hhstream://containers application/json — every container plus the orphan hh-* outputs. The listing only; see §3. —
hhstream://screen the published main output as an image blob. With containers on this is the whole compositor layout, so it is wider than any single output's configured size. —
hhstream://sessions/{id} application/json — one session with its recent audit events ?events=N
hhstream://sessions/{id}/logs application/json — the application's own captured output ?lines=N
hhstream://outputs/{name} one output captured alone, as an image blob. Exactly that output's configured size, which is the coordinate space the pointer tools aim in. —
hhstream://streams application/json — every session a stream was enabled on, with its URL, rate and connected viewer count, plus available (whether the stream daemon answered) —
hhstream://sessions/{id}/stream application/json — one session's stream, or null. attached is separate from enabled, and the URL is absent rather than stale when it is false. —

The URLs in those two bodies contain their phrases, which are the only thing protecting those streams. Both descriptions say so, because a host that puts hhstream://streams into a model's context has put a credential there — and the one place it must never go is anywhere the streamed session can see.

Rules this surface keeps:

  • A read never writes. That is what excludes session container <id>.
  • Only declared query parameters are accepted. An unknown one is refused, for the reason every tool schema sets additionalProperties: false: a parameter silently ignored is a read that answered a different question from the one asked. A parameter given twice is refused too.
  • A literal route beats a template. hhstream://sessions is the listing, not a session whose id happened to be empty.
  • The media type of a frame is read off the bytes, never off the extension the server asked the CLI for. The entry's declared mimeType is a hint; the ReadResourceContents carries the answer.
  • No subscriptions. Server.get_capabilities in the pinned SDK hardcodes ResourcesCapability(subscribe=False, …) — it never advertises subscribe even if a handler is registered. So resource subscription is not available as the answer to the missing logs --follow, which is a shame, because it would have been the right one. Read it again instead.

5. What is deliberately absent

Not exposed Why
session logs --follow it polls until interrupted, and a tool call that never returns hangs the client that made it. Read hhstream://sessions/{id}/logs again. Resource subscriptions would have been the principled fix; see §4.
session start --no-container the CLI's flag exists in order to refuse: the broker's start verb carries no such field (docs/CONTRACT-sessions.md §4), so the CLI exits 2 and names the disagreement. A tool argument whose only possible outcome is a refusal is a trap.
hhstream --frame-socket / session --socket §2: the environment configures them, in one place.
a per-call timeout §6.
session keyboard with no verb it prints help and exits 2. There is nothing for a tool to do with that.

6. Error mapping

Every refusal the CLI produces on a tool call comes back as a tool error — CallToolResult with isError: true — carrying:

  • the remedy sentence, first, naming the code: exit 9 (NO_LEASE): …;
  • the CLI's own error, reason and hint when it printed a protocol object;
  • everything it said, verbatim, under hhstream said:;
  • structuredContent with ok: false, exit_code, remedy, and the protocol object's own fields.

isError and not a JSON-RPC error, for tools, throughout. A JSON-RPC error reaches the client's log; "another session holds the lease" is something the calling agent has to read and act on. (The SDK's call_tool handler also converts anything raised into isError, so this is the only channel there is in practice.)

On a resource read it is the other way round, and not by preference: ReadResourceResult has no isError, so a refusal is raised as a JSON-RPC error with code -32602 (INVALID_PARAMS), its message carrying the same remedy-then-the-CLI's-own-words text and its data carrying the same {ok, exit_code, error, reason, hint, remedy} object. Returning the broker's failure object as the resource's body would make "no such session" read as a successful read of a session, which is the one outcome a contract has to forbid. The same error is raised for a URI that names nothing and for an undeclared query parameter — in both cases before the CLI is reached at all.

"Arrives as" below is for a tool call; on a resource read every one of them is the protocol error above, with the same remedy.

Exit Name Arrives as The remedy an agent is given
0 OK a result —
1 not in the frozen table tool error the frame verbs' own failure: no frame published, or grim/ffmpeg missing from the server's PATH. Also what an uncaught exception in a handler becomes, with its traceback — which is what a CLI process would have done.
2 USAGE tool error fix the arguments; a retry fails identically
3 NO_SUCH_SESSION tool error read hhstream://sessions to see which ids exist
4 BAD_STATE tool error read hhstream://sessions/<id> to see which state
5 DAEMON_UNREACHABLE tool error an operator problem; the three frame tools and hhstream://status may still work
7 TIMEOUT tool error read the state back, then retry with the same request_id
6 POLICY tool error a configuration decision; no retry changes it
8 BUSY tool error read hhstream://pointer for the holder; wait. Never "acquire one"
9 NO_LEASE tool error pointer_acquire, then pass the new lease_id
10 COMPOSITOR tool error the session is not lost; it is degraded

8 and 9 must never read alike, and a unit test holds them apart: an agent that sees 8 must wait, one that sees 9 must acquire.

Two failures are not the CLI's verdict and are reported in the same channel anyway, because the agent holding the tool is the one who has to say so:

  • a tool name that does not exist — the reply names the ones that do;
  • an argument the schema rejects — the SDK refuses before an argv list exists, so the CLI is never reached.

The sg re-exec is not available here. hhstream recovers from an unreadable frame socket by re-executing itself through the setuid sg wrapper to join the compositor's group. That is an exec, and performing it would replace this server and drop the client's connection. So a frame verb that cannot reach the socket fails with the CLI's own message plus a sentence saying what the CLI would have done and how to avoid needing it: run the server as a member of that group (sg hhstream -c hhstream-mcp in the client's spawn command, or add the account to the group). The hint is added only when the condition is actually true, asked of the CLI's own predicate, and only for the verbs it could be true of.

There is no per-call timeout. In process there is nothing to kill: a thread cannot be interrupted. The deadlines are therefore the ones that already exist — the session client's 60 s socket timeout, the broker's per-verb deadline, and the frame socket's own --timeout — and a call that outlives all of them outlives this server too. A subprocess design could have killed it; that is the one thing given up by calling the CLI in process.

7. Concurrency

Tool calls are serialised. The CLI prints to a process-global sys.stdout, so two concurrent calls would interleave their output into each other's reply. One lock around the whole call is the honest fix, and costs nothing a caller would notice: the broker already serialises every mutation, and the CLI was one process per verb anyway. The event loop stays free, so a ping is answered while a record runs.

8. The two rules, which every relevant tool and resource repeats

These are hhstream_session.ACTIVITY_NOTE and KEYBOARD_LEASE_NOTE restated in terms of this surface. They are on the individual tools and resources and not only in the server instructions, because a client shows a thing's own description at the point of use:

  1. session_touch is the only thing a caller can do that resets a session's idle clock. Reading hhstream://sessions, hhstream://sessions/{id} or its logs does not, session_gc does not, the stream tools do not, and neither does any pointer or keyboard tool. A session being driven is still a session nobody has said they want kept. The resources it applies to say so in their own descriptions, because a description is what a host shows beside the thing it describes.

    The one exception is a connected viewer (docs/CONTRACT-sessions.md §4, v4): about once a minute the broker stamps touch on any session somebody is watching through its live stream. It is not a tool and cannot be invoked — it is the broker noticing that a person is looking, which is the one form of "still wanted" that does not come down the control socket. It cannot hold a session open by itself: an enabled stream with no viewer captures nothing and stamps nothing.

  2. Every keyboard tool needs the pointer lease, and holds only while it holds. There is no keyboard lease: one compositor seat means one keyboard focus. A held non-modifier key floods the window until it is lifted or the lease expires, because the application does the repeating — measured at 25 a second after 600 ms — and this server cannot rate-limit what it does not generate.

tests/unit/test_mcp_tools.py asserts both notes are present on every tool they apply to and holds the CLI's own constants as a tripwire against drift; tests/unit/test_mcp_resources.py does the same for the five reads the first rule applies to.

9. Packaging

packages.hhstream-mcp (nix/mcp.nix), a separate derivation from packages.hhstream because mcp is a third-party dependency and the frame socket, the CLI and the broker are stdlib-only by policy. packages.hhstream must not grow that dependency; the root pyproject.toml excludes hhstream_mcp* from its package discovery so it cannot acquire it by accident.

The stream tools and the two stream resources additionally need services.hyprland-headless-stream.sessions.stream.enable, which runs the hhstream-live unit and gives the broker the socket it switches streams on over. Without it they answer exit 10 (stream_unavailable) — the session is unaffected, which is what that code means.

services.hyprland-headless-stream.mcp.enable (default false) puts the binary on the system PATH. That is the whole of the option: there is no unit, no socket and no state directory, because the stdio server is spawned per client.

services.hyprland-headless-stream.mcp.http.enable (default false) is the other half, and it is a unit — hhstream-mcp.service, running as sessions.user with sessions.group, because the tools it serves are the CLI's own handlers called in process and that CLI opens a 0660 control socket owned by that user and a compositor runtime directory owned by that group. It is Wants/After the broker and not BindsTo: the frame verbs answer with the broker down and the rest answer exit 5, which is a true answer, whereas a listener that vanished with the broker would turn that into a connection refused and tell a client nothing.

The two options are independent and neither implies the other. http.enable also puts the binary on PATH, because token mint is how the credential that unit demands gets issued.

option default
mcp.stateDir /var/lib/hhstream-mcp, created 0700 by a tmpfiles entry
mcp.http.address 127.0.0.1
mcp.http.port 9989
mcp.http.path /mcp
mcp.http.allowedHosts / allowedOrigins [] — hosts are derived from the address
mcp.http.toolsets null, every tool
mcp.http.openFirewall false, and not implied by http.enable
mcp.http.extraArgs []

The unit refuses to start while the authorization table is empty, and retries every five seconds rather than giving up (StartLimitIntervalSec = 0). Both halves are deliberate and neither works without the other: the refusal is there because a listener that 401s every request is almost always a forgotten mint, and "connection works, auth fails" is far worse to debug than a systemctl status that names the remedy — but a refusal that gave up would leave every freshly provisioned machine dead, and minting the first token would then also need a systemctl start nothing told anyone about. With the retry, the unit comes up on its own within one interval of the first mint. checks/vm-mcp-http.nix asserts exactly that, because nothing in the Python can.

Revoking every token does not stop the running process — it keeps its port and refuses everyone. The startup refusal exists to catch a forgotten mint, not to police a deliberate one.

10. What the HTTP transport's token is not

Not MCP's authorization specification. That is OAuth 2.1 plus RFC 9728 protected-resource metadata: a 401 is meant to carry a WWW-Authenticate pointing at a metadata document, the client discovers an authorization server, and a real flow happens. What this serves is a static token an operator minted and pasted into a client's configuration.

Stated rather than discovered, because the consequences are not subtle:

  • a client that can only discover an authorization server will not connect to this at all;
  • a client that lets you set a header works immediately;
  • there is no refresh, no rotation and no revocation list beyond token revoke, which is immediate because the table is read per request — no restart, and an operator who had to restart a service to lock someone out would not have revocation.

The gap is docs/TODO/mcp-http-auth-is-not-oauth.md, beside the two limits the SDK imposes.

A token does not scope resources. §2 keeps resources out of --toolsets on the grounds that they are the host's context rather than the model's budget, and that argument carries over to the server's own flag. It does not carry over cleanly to a token, which is an authorization and not a budget — so a token scoped to frames can still read hhstream://sessions. That is a real gap, not an oversight: docs/TODO/mcp-tokens-do-not-scope-resources.md, and tests/unit/test_mcp_http.py asserts the current behaviour so that fixing it is a visible change rather than a silent one.