Repository navigation
Releases
v0.6.7
Fixed
- The Tools column of a set's servers and of the all-servers page takes the
count from the catalog of a running set (GET /api/sets/<id>/tools) instead
of waiting for Check on every row; a probe still replaces it with a live
count, and a stopped set keeps—.
v0.6.6
Fixed
- Reload applies profiles and keys as one generation: the listener receives the
new token list only when the reload needs no restart, so a key that arrived
together with a transport change is refused until the set restarts instead of
being accepted without its profile (D10). - Upstream error text is masked like the payload: a server repeating its own
secret inside an error no longer leaks it into the tool answer, the counters
orraw-*.jsonl, meta-tools included (D12). - Secrets are masked by value rather than by field name: every
envor header
value of 8 bytes or more becomes***—GITHUB_PATandCF_ACCESS
included — so a tool catalog may show***where a harmless path used to be
(D4). - A panic inside the health probe no longer hangs every later request to that
set: the single-flight is released, the probe is retried and oneWarnis
logged (D2). - Deleting a set forgets its webhook transitions, so a new set created under the
same id warns again (D3). - Failed logins: expired windows are dropped on every recorded failure and the
map is capped at 1024 entries, instead of being cleaned only when it had
already overflowed (D5). - The tool-catalog throttle map sweeps marks older than its 5 s window instead
of growing for the lifetime of the process (D1). - The statistics cache evicts the oldest quarter of its files by
mtimeinstead
of dropping the whole map at 4096 entries (D7). - CI: the
acceptancejob runs the 26-step panel acceptance on a freshly built
binary;ACC_DAPbecame optional and its 11 stdio steps are skipped with a
line in the output (D6). GET /statsof a set no longer hands the full snapshot to whoever holds the
set's key: a key now gets counters and latencies with client labels replaced
by(аноним)and upstream error texts emptied. The full snapshot goes to the
panel's own poll only (X-MCPRig-Caller: panel), so/api/callskeeps
showing labels and errors as before.- Release pipeline: the tag name was expanded into the text of
run:blocks in
jobs that hold a write token and the registry password, so a tag containing
shell metacharacters became code. The tag now arrives throughenv:and is
read as"$TAG"; the release checkout keeps no credentials in.git/config
(persist-credentials: false). The Gitea image workflow takesgitea.ref_name
the same way. Who may pushv*tags is a hosting-side protection rule that
this repository cannot confirm. - Server log: the Log dialog of a server showed only the access lines of the
set and was empty in practice. The child's stderr now goes to
sets/<id>/serverlogs/<name>.log, capped at 256 KiB, kept after a stop so the
trace stays readable and cleared on the next start; an automatic restart
appends under a new marker instead of wiping the reason the server died. The
dialog appends that tail to the set lines. - Server log: each run of the process opens with a
— запуск <date time> —
marker, so a restart after a crash is told apart from the first start; a tail
that exists but cannot be read now comes witherrorin the answer and the
dialog prints it instead of showing "empty" as if the server were silent. - Sets:
POST /api/setsrefused to say anything when the name was outside the
id pattern — it renamed the set silently. The id is now taken as typed and a
bad one is400with the rule in the message; the panel reduces the name to
that form in the form and shows the rule instead of sending something else. - Panel: one refusal of the first
/api/statestopped every timer for good —
the panel then sat frozen until a reload; the failure is reported and the
timers start anyway. - Panel: a click that landed while the periodic sync held the panel showed
nothing; it now says an operation is in progress. - Panel: the statistics page loaded its data only on the second tick when opened
by address; it now asks as soon as the sets are known. - Server log: the mode of an existing tail file is brought back to
0600when it
is opened, and a read failure answers with the file name inside the set rather
than the absolute data path. - Test:
scripts/acceptance/panel.cjs— the 37-step panel walk — now lives in
the repository instead of a scratch directory, andweb/ids_test.gofails
whenapp.jsasks for an idindex.htmldoes not have. - Panel: the copy button of the endpoint moved the address out of the DOM for the
second it showed "copied" — switching the language in that window threw
Cannot set properties of null; the feedback now goes to the button label. - Panel: deleting the open set returns to the list of sets; before, the next set
in the rail was opened, which read as "wrong set". - Panel: on a narrow screen the copy feedback shrank the address button from 74 to
45 px; it now shows a mark there and keeps its width. - Panel: statistics opened by address stayed blank for up to 15 s — the load ran
before/api/stateknew the sets and returned early; it now starts as soon as
the set is known. - Panel: the statistics filters (server, tool, client) survive a reload.
- Panel: building a prompt with an empty task showed nothing and left the button
enabled; the "set has no servers" hint sat inside the field that is hidden in
exactly that case and was unreachable. Both are shown now. - Panel: revoking the active key stops the set and takes a few seconds; the keys
tab now says so while it waits instead of going quiet. - Panel: buttons and the key-note field were 27–31 px tall; the tap target
minimum is 32 px, and the Calls checkbox is pressed through its label. - Panel: choosing a set in the rail opens that set from any page — from the host
overview, statistics, all servers; before, the click only changed the window
title, and on a fresh install the tabs of a set could not be reached at all. - Panel: the state badge of a set (sets list, host overview, servers table) took
its colour from a class that does not exist and showed as a black dot —
paintBadgenow accepts both spellings the callers use. - Panel: a direct link to
#/set/<id>/<tab>and a reload on it left the page
half-filled (health, calls and the tool catalogue arrived only with the 5 s
tick, the key profiles never); the first state now loads the set the address
names. - Panel: the saved key label was silently dropped while the overview tick held
the busy flag, and a refused refresh after a revoke said nothing. - Panel: the reason of a failed call and of a stopped set was cut to 80/60
characters in the markup itself, so pressing it on a touch screen could not
reveal the rest; the DOM carries the whole text and only the line clamps it. - Panel: deleting the open set jumped to the neighbour in the address bar before
returning to the list, leaving a foreign set in the history. - CLI:
mcprig logsprinted the child's stderr as is, so escape sequences from a
server could rewrite the operator's terminal; control characters are stripped. - Server log: a failed write, truncate or cleanup of the tail is now reported in
the set log instead of leaving "log is empty" for weeks, and one oversized
write can no longer push the file over its cap. - Auth:
X-Forwarded-*was trusted from any loopback peer once
MCPRIG_PROXY_SECRETwas set, which let that peer skip the login limit by
naming a fresh client address; the request must now carry
x-mcp-proxy-secretas well. - Panel: a deleted set left its rings, locks and nonces in memory, so a new set
with the same id appeared stopped, and a pending operation of a removed set
was never finished. - Panel: a count the panel does not know yet (
tools to approve, servers with an
error) printed0; it shows a dash instead of a wrong number. - CLI:
mcprig logs --serverprinted "log is empty" when the tail could not be
read; the answer carrieserrorand the command prints the reason. - Panel: the Logs tab of a set reported "log is empty" when
mcprig.logexisted
but could not be read; it now names the reason above the box. - Panel: the Log dialog subtitle described the format wrongly (it said "set
log only"); it now says the set lines come first and the process tail after them,
and the same sentence is used by the tab.
v0.6.5
Added
- Panel: an overview page at
#/— sets running and stopped, servers by health,
tools waiting for approval, calls and errors over 7 days, and the newest calls;
the list of sets moved to#/sets. - Panel: every tab of a set has its own address (
#/set/<id>/servers,…/keys,
…/tools,…/calls,…/logs); back and forward walk through tabs, an unknown
tab opens the overview. Statistics became a page of its own (#/stats/<period>).
Changed
- Panel: the set log is a Logs tab, keys and tool approvals are tabs of their
own, and the details drawer is gone. - CodeQL analyses the code only: the generated diagram pages under
docs/diagrams/and the design mockups underdesign/ui/are excluded
(.github/codeql/codeql-config.yml), which is where most of the 112 open
alerts came from.
Fixed
- A path segment taken from a request — a set name in the statistics directory, a
history file — is checked against the registry's name rule before it is joined
into a path (internal/safepath). A bad segment reads as "nothing to read"
instead of reaching outside the data directory. - A mount path of the sandbox is cleaned and rejected when it contains
..,
instead of relying onfilepath.Jointo collapse it later. - A notification webhook is parsed before the request is built: only
httpand
https, a non-empty host, nouser:pass@in the address. - Values that reach a log from a request or from configuration no longer carry
control characters, so a name cannot forge an extra log line
(internal/safelog).
v0.6.4
Fixed
mcprig panel -h/--helplists all twenty-seven environment variables the
panel reads — grouped as paths, network, auth, TLS, proxy and OpenBao — and
exits0; any other argument is a usage error with exit code 2.XDG_STATE_HOME,
BAO_ADDR/OPENBAO_ADDR,BAO_TOKEN/OPENBAO_TOKEN,BAO_MOUNTand
BAO_PREFIXwere missing although the panel already read them, and the count
indocs/{en,ru}/cli.mdmatches again (cmd/mcprig/main.go,
cmd/mcprig/main_test.go).- The panel's Keys dialog says
label, as its API always did: issuing a key
is onePOST /api/sets/<id>/keyswith{"label":…,"profile":…}, and editing a
label sends{"label":…};notesurvives only as the old name the daemon
still accepts (web/app.js,internal/manager/api.go,
internal/registry/types.go). - A corrupt
registry.jsonno longer disappears silently: it still moves to
registry.json.corrupt-<unix>, but startup logs an ERROR naming the backup,
the byte size and the number of set directories left on disk, and the same
line is available throughDropped(). Before, the panel came up with an empty
registry and the log said nothing (internal/registry/store.go,
internal/manager/api.go). PUT /api/servers/<name>with a changedenv/envAllownow names the set
whose restart failed inreloadErrors, like the sibling branch that reloads
keys, instead of leaving the failure in the log behind a 200.- Changing or deleting an
envvalue waits for the restart in the response
(bounded by the existing 60 srestartTimeout) and puts a set that did not
come back intoreloadErrors;restartednow lists sets that actually
restarted. A failed background restart used to leave a set serving with a
revoked secret while the answer promisedrestarted. PUT /api/sets/<id>answers 200 withfailed[setID]when the edit is already
in the registry but the restart or reload could not apply it: a 500 made the
panel treat a committed edit as lost.- The panel shows both of those: a set named under the variables form with the
reason fromreloadErrors, and the same line under the set composition when
failedorreloadErrorcomes back with the edit. A secret rotation that did
not restart the set used to look applied (web/app.js). GET /api/sets/<id>/servers/<name>/authorize/statuschecks the same
conditions as the launch — the server exists, belongs to the set and is
remote — and answers 400 otherwise. A server removed from the set used to
report200 idle, and the panel dialog spun forever.- The tool-out sweep counts files it could not delete, logs them and returns
the aggregated error (не удалось удалить N файлов), while still cleaning the
other directories. Retention that failed silently lettool-outgrow
unbounded (internal/toolout/toolout.go). POST /api/sets/<id>/reloadon a stopped set answers
200 {"reloaded":false,"reason":…}instead of 500 when the set has no active
key: nothing runs, so there is nothing to refuse.servers.jsonis not
written, and the reason names the missing key. A 500 is now only a real
failure to build or write the config (internal/manager/api.go,
internal/manager/setconfig.go).- A stdio server now receives
HTTP_PROXY,HTTPS_PROXY,NO_PROXY, the
lowercase forms andNODE_USE_ENV_PROXYfrom the proxy's environment: behind
a proxy it used to see "no network" and report a fake authentication failure.
The rest of the environment stays whitelisted (internal/downstream/env.go). - A truncated answer stays valid JSON: long string values are shortened and the
fact is recorded inside the document (_truncated), the marker no longer
promises a/tool-out/<name>URL for an open set, and a non-JSON answer says
plainly that it was cut and where the full one lives
(internal/mcpserver/toolout.go). - Timeouts name the server and the command and wrap the cause, so
errors.Is(err, context.DeadlineExceeded)answers; waiting for a slot is
reported as a queue refusal with its limit and wait time instead of a call
timeout that never happened (internal/mcpserver/manager.go,
internal/mcpserver/invokemany.go). list_commandsanddescribe_commandsapplypermissions.deny: a command
the key may not call no longer appears in the catalog or its schemas, while
invoke_commandkeeps answering with the policy refusal.refresh:truedrops the catalog cache at most once per 5 s per caller and
server — the caller is the key's fingerprint, or the client address on an open
set, so one client no longer pays for another's refresh; a throttled refresh
returns the cached catalog with arefreshDeferredmarker instead of an error.- A fan-out budget exists only when the set carries a
requestTimeout: a call
ofinvoke_manythat brought notimeoutMsof its own is no longer cut short
by a neighbour's limit, and the answer names the limit the call actually ran
under instead of0 ms(internal/mcpserver/invokemany.go). - Issuing a key writes both
labeland the oldnotefield, so a daemon from
an earlier release still shows the label after a downgrade
(internal/registry/store.go). - A corrupt
tools.lock.jsonis a refusal, not an empty file: descriptors are
no longer auto-approved behind the quarantine, the corruption is logged and
list_serversshowsapproval: "quarantine"for the set. - The port probe binds the wildcard address a set listens on, not
127.0.0.1: a port taken on the host's own address no longer looks free
(internal/registry/store.go). invoke_manyanswersisErrorwhen every call failed, likeinvoke_command,
and a set can no longer list the same server twice.list_serversnames each server's URL or command, the declared tool count
against what the key actually sees, and whether approvals are enforced.
Changed
- A tag no longer publishes unchecked code:
.github/workflows/release.yml
runsgo vet ./...andgo test ./... -race -count=1 -p 1before building the
archives, becauseci.ymldoes not trigger on tags and the Gitea workflow
that used to gate releases is gone. .gitea/workflows/images.ymlbuilds on av*tag only: the manual dispatch
used to publishlatestfrom an arbitrary branch to the host registry..golangci.ymlstops excludinggosecG301/G302/G306globally; every
write in the tree already uses0600/0700, and the exclusion hid the next
permissive one.invoke_commandand every call ofinvoke_manytake their owntimeoutMs
(1000..600000 ms): the call is bounded by it instead of the set's
requestTimeout, the answer names the limit actually used in
_meta.timeoutMs(timeout_msper batch entry, the batch's budget in its
_meta), and an out-of-range value is refused before anything runs
(internal/mcpserver/manager.go).
Added
list_serverscarries ahintper server — one line on what it solves and
where to start — from a built-in table of the servers this box knows
(internal/mcpserver/hints.go); a set's ownhintfield overrides it, and an
unknown server has no field at all (internal/config/config.go,
internal/registry/validate.go).- Appearance splits into two axes switched from one menu in the header: the
theme (defaultorwin98) and the colour scheme of that theme. The new
win98theme is the Windows 98 look — a navy title bar with the pixel logo
and the three window buttons (decoration; every action stays in the toolbar),
square corners, two-tone bevels, MS Sans Serif, dotted focus inset 3 px,
sunken fields and selects, folder tabs, segmented status bar, grey scrollbars
— inweb/win98.css, with four schemes (Classic, Slate, Wheat and High
Contrast: black body and fields, white text and outlines, yellow accent) that
differ in shades only, and 3 px bevels underprefers-contrast: more. Theme
and scheme are stored inmcprig.themeandmcprig.palette, the Win98 scheme
inmcprig.palette.win98; names written by earlier versions open as the
defaulttheme with the same colours, without a reload
(web/theme.js,web/index.html:20-29,web/app.js:3779-3855).
v0.6.3
Changed
- The image build in
ci.ymlruns on a pull request only, and only when the
Dockerfile,.dockerignore,go.mod/go.sumordocker/changed: a push
tomasterno longer rebuilds an image the tag will build anyway.
Fixed
GET …/authorize/statusanswers404for a set or server that does not
exist, like the launch route, instead of200 idle.mcprig --helpprintsUsage of mcprig:— it used to sayUsage of serve:
even when no subcommand was given.
v0.6.2
Changed
- Checks and releases run in one place: GitHub.
.github/workflows/ci.ymlruns
on every push and pull request (gofmt,go mod tidy -diff,go vet,
go build, cross-build fordarwin/arm64,darwin/amd64andlinux/arm64,
tests with-race,golangci-lintfrom.golangci.yml,govulncheck,
gitleaksover the whole history,docs-check,node --checkand a
docker buildof both image targets),codeql.ymlscans Go and the panel's
JavaScript,dependency-review.ymlfails a pull request that adds a
dependency with a high-severity advisory, and Dependabot proposes updates
weekly. Gitea keeps the code mirror and one workflow,images, which builds
slim/fullon a tag and pushes them to the host registry — GitHub cannot
log in there (403). Itsci.ymlandrelease.ymlare gone, so tests and
release archives are built once, on GitHub, and the release publishes them
together with the GHCR images. - The pipeline is hardened: actions are pinned to commit SHAs, permissions
default tocontents: read, checkout leaves no token behind
(persist-credentials: false), pull requests receive no repository secrets,
and the scanners install their own tools instead of pulling third-party
actions.
Docs
docs/en/dev/release.mdanddocs/ru/dev/release.mddescribe the single
builder, the list of checks and what keeps the pipeline itself safe;
docs/en|ru/dev/development.mdpoints at the same file instead of repeating
the job list.
v0.6.1
Fixed
- A request with
Origin: nulland no session cookie — a sandboxed frame, a
file://page, a client that is not a browser — is served again: the check
refused it before it could be told apart from a cookie-carrying one, which
also brokePOST /api/loginfrom such a page. The refusal stays for the case
it exists for:Origin: nullwith a session cookie.
Docs
- docs/en/upgrading.md and its Russian mirror collect
what answers differently after 0.6.0, with the fix for each symptom, and how
to roll back; the README now carries all three diagrams under spoilers.
v0.6.0
Added
- The panel has pages of its own instead of one long Overview:
#/sets,
#/serversand#/varsare pills in the content head,#/statsand
#/set/<id>keep working as before, and every page has an address that
survives a reload and can be sent to someone else. - All servers is a page of the whole host registry: name with the reason of
a failure, transport, a health badge in colour (healthygreen,failedred,
not checkedneutral — the same badge the set table uses), the tools a check
found, the sets the server belongs to, and per-row check, log,
edit and × that deletes it from the registry after naming the sets it
leaves. The details drawer no longer carries the registry: it keeps the log
tail of the set, so a set's own composition and a host-wide list stop sharing
one screen. - Variables is a page: the names your servers reference as
${NAME}, each
with whether it is set and where it comes from (openbao,file,env), a
form that writes a value — name, value, storeautomatic/OpenBao/file—
and × that forgets it everywhere. The panel says plainly that it never
reads a value back and that a name no server mentions stays out of the list. - The panel composes a set in place: the servers table of the Overview tab
gains a registry picker with add to set and a × that detaches a server.
Both sendPUT /api/sets/<id>with the wholeserverslist and redraw from
GET /api/state; a refusal shows under the block, an empty set says it will
serve no tools, and the hint states that a running set takes the change by
reload, not by restart. - Import / Export is a dialog of its own, opened from the head of a set and
from the empty state: the whole registry or one set as JSON with Download
JSON and Copy, and an import from a JSON file or a pastedmcpServers
object behind a preview that names the servers to be added — sets found in the
file are skipped with a note, because sets are made in the panel. - The head of a set carries a red Delete set button: the confirmation names
the endpoint and the directory that go away, and the panel returns to the
empty state afterwards. - Prompt in the head of the set opens a dialog over the endpoint below: the
task is required, the introduction and the server subset are optional, and the
built text comes back monospaced with copy and download .md;notes
are shown above it, and a stopped set warns beforehand that schemas will be
missing. - The keys block of a set shows the
keyMarkofGET /api/statenext to the
prefix — the very label the statistics print for that key: its label when it
has one, otherwisekey:<mark>; hidden for an open set. POST /api/sets/<id>/promptbuilds one text for an agent out of three
parts:{"task":"…","intro":"…","servers":["…"]}and the answer
{"prompt":"…","notes":["…"]}. The tool list comes from the live catalog of
the set — names, descriptions and one-line input schemas — and a stopped set
falls back to thetools.lock.jsonnames with a note that schemas are
missing. One schema is capped at 4000 characters, the whole prompt at 64 KiB,
both cuts reported innotes; set tokens and variable values never reach the
text.GET /api/stategives every set akeyMark: the label its primary key carries
in the statistics — the key's label when it has one, otherwise thekey:<6 hex>
fingerprint — so a row in the statistics can be matched to the key in the
panel. Empty for an open set;keyPrefixstays as it was.- A set may be created and started with no servers. An empty
serverslist is
now a legal state:POST /api/setsandPUT /api/sets/<id>accept it, in an
update a missingserverskey leaves the list untouched, and a set without
servers starts and answers the meta-tools (list_serversis empty). Deleting
the last server of a set no longer answers400— the set keeps its key and
port with an empty list. - A saved variable that no server of the install references is visible in the
panel and inGET /api/envwith"used": false: the Variables page lists
every stored name, marks the ones nothing references as not used by any
server and lets them be forgotten, instead of hiding a value that stays on
disk. Unreferenced names come last;usedis new in the answer.
Docs
- Two new diagrams next to the architecture one: the same servers, one
endpoint — what a client pays without a proxy and with one, plus what the
endpoint costs — and what one binary does, a capability map
(docs/diagrams/, spec.jsonand PNG per language). docs/en/alternatives.md(and the Russian mirror) compares mcprig with
mcp-proxy, mcp-hub, mcpo, Docker MCP Gateway, MetaMCP, 1mcp, the Lasso gateway
and supergateway, says where each fact came from, and lists where mcprig is
weaker: one operator, no roles, a spoofableX-MCPRig-Caller, a sandbox that
leaves CPU and egress alone.- The README carries a fresh panel screenshot and the corrected count of meta
tools: nine, of which five work with servers.
Changed
- A panel started without a password (
MCPRIG_AUTH=off) now says so at
startup:WARN панель без пароля (MCPRIG_AUTH=off): доступна всем, кто дотянется до порта; задайте MCPRIG_PASSWORD.mcprig doctorprints the same
asпредупреждение:under services when noMCPRIG_PASSWORDor
MCPRIG_PASSWORD_HASHis set — a warning, not a problem, so the exit code
stays0. list_serversno longer answers withpid: the proxy's process id is of no
use to a client.versionstays.GET /api/callsanswers for a stopped set as well: its tail and counters come
from the raw journal on disk (sets/<id>/stats/raw-*.jsonl) over the same
7-day window asperiod=7d, up to the 512 newest rows, and a missing
directory answers empty instead of landing inunreachable.schema,
throttledandunreachablestay properties of a running set; a set without
a token is anonymized on read.- Signing a key (
PATCH /api/sets/<id>/keys/<keyID>) reaches the set by reload
instead of a restart: the label travels to the statistics client names and the
set's pid stays put. - Stopping the panel closes every set in parallel under one shared budget (15 s
for readiness and 5 s for termination in total) instead of one budget per set,
so a dozen sets no longer hold the exit for a minute. - A refused start or update now answers by the kind of refusal instead of its
wording: a busy port, a set that did not come up and a set without an active
key answer409, and anything unclassified answers500with a line in the
log and a generic message in the body. A set with an unfilled${VAR}in its
environment is refused at start with409and the name of the variable
instead of starting without it. mcprig doctorreads an empty install as healthy: a missing config and a
missing registry together mean the setup has not been made yet, so it prints
one line, exits0and no longer reports "3 problems" on a fresh directory.describe_commandsanswers within the caller's profile: the same allow and
deny rules that hide a tool fromlist_commandsnow hide it here as well.invoke_commandtakes the server's slot: a slow upstream no longer lets an
authenticated client stack calls past the per-server concurrency limit.- A registry validation error answers
400with the reason instead of500
with "see the log":internal/registrynow returns a typedErrInvalid
carrying the text of the first broken field (port 9200,env.A), and the
panel shows it as it is. - A dump at
/tool-out/<name>belongs to the key that caused it: the file gets
a<name>.ownerfingerprint next to it and another key gets404. The
lifetime of a dump follows the set policy now (transport.toolOutTtl) instead
of only the panel-wideMCPRIG_TOOLOUT_TTL. GET /api/servers/<id>/authorizeno longer starts the OAuth flow: the flow is
POST, andGET …/authorize/statusreads it. AGETon the starting route
answers405withAllow: POST; the panel polls the status route.X-Forwarded-Host,X-Forwarded-ProtoandX-Forwarded-Forare trusted only
from a loopback peer or from an address inMCPRIG_TRUSTED_PROXIES(IP or
CIDR, comma-separated), and not becauseMCPRIG_PROXY_SECRETis set. A peer
outside the list loses the header silently: the panel answers with its own
origin and the login limit counts the peer's address. A value that parses as
neither an IP nor a CIDR refuses the start with the offending entry.- Behind a trusted proxy the login limit counts the client, not the proxy: five
failed attempts from oneX-Forwarded-Foraddress no longer lock out every
other client of the same proxy. - A release is published by CI itself:
.goreleaser.yamlcarries
release.draft: false, so av*tag produces a public release with no human
step. Every push and pull request runs.gitea/workflows/ci.yml—gofmt,
go vet,go build, the race suite,scripts/docs-check.sh,
node --check web/app.jsandgovulncheck— where before only docs were
checked.
Fixed
PUTandDELETE /api/env/<NAME>no longer stop a running set that would
lack a variable afterwards: the set keeps working on its previous environment
and the answer names it inskipped({set, reason, missing}) next torestarted.
Before, the set was restarted into a refused start and the answer still said
it had been restarted.- Adding a server that brings a
${VAR}the set did not use before restarts the
set: the process environment is fixed at start, so a reload would have left
the variable empty while answering200.needsRestartcompares the variable
references of both compositions, not only the port. POST /api/sets/<id>/promptmasks tool descriptions and schemas li...
v0.5.1
Fixed
TestEnvPutRestartsOnlySetsUsingVarstops the restarted set and waits for it
to exit before the temp directory is removed; the race used to fail the
verify job on CI.- Builds move to Go 1.27.4 and
golang.org/x/netv0.62.0: govulncheck reports
no reachable vulnerabilities (GO-2026-6859 and the x/net advisories).
v0.5.0
Added
-
Downstream resources and prompts reach the MCP client: a server whose config
names them inexpose(["resources"],["prompts"], or both) contributes
them through the newlist_resources,read_resource,list_promptsand
get_prompttools. Withoutexposethe client still sees tools alone, and an
unknown value is a config error — the server does not start. Permissions
address them asresource:<uri>andprompt:<name>, so the existing rules
and profiles apply unchanged; answers are cut atmaxBytes(32768 by default
for a read) and the full one lands intool-out. The listings honour the same
rules: adefault: denykeeps a resource out oflist_resourcesuntil a rule
allows it. -
exposeis editable from the panel: the server form carries two checkboxes
(resources, prompts), the registry reads and validates the field onPOST /api/serversandPUT /api/servers/<name>, and an unknown value is a400
that names what is allowed. Unchecking both removes the field. -
An edit no longer restarts a whole set:
PUT /api/sets/<id>,PUT /api/servers/<name>and the newPOST /api/sets/<id>/reloadwrite
servers.jsonand sendSIGHUPto the set process alone — its stdio children
have no handler and keep running. The process rereads the file and answers in
reload.statusnext to it (nonce,generation,applied,restartNeeded,
error); the nonce the panel puts intoreload.requestbefore the signal
keeps a late answer to an earlier attempt from being taken for this one. Calls
in flight to the other servers keep their sessions, and the set's pid and
restart counter stay put. A transport change (port,open,path,
host,allowedHosts) is still applied by a restart and lands in
restartNeeded; keys are the exception and travel by reload (below);path,hostandallowedHostsin a hand-edited
servers.jsonland there too and the set is restarted by hand. A server's
envandenvAlloware not reloadable either, so an edit of them restarts
the sets using that server. A broken config leaves the process on the previous one and
reports the cause. A failed reload does not lose the edit:PUT /api/sets/<id>
answers 200 withreloadError, andPUT /api/servers/<name>lists the sets it
could not apply inreloadErrors. That answer now names the setsreloaded
instead ofrestarted;restartedstill carries the same list for one
release. -
A set serves several keys at once. Every key carries
id, its token,
createdAt,revokedAt,profileandlabel; the token never leaves the
server and the hash lives only in the listener's memory. Issuing a key no
longer revokes the previous one, and revoking one key no longer stops the
set — only the last active key does. A set keeps up to 8 active keys and the
ninth answers409. The routes areGET /api/sets/<id>/keys(ids, labels,
profiles and dates, never a token),POST /api/sets/<id>/keyswith an optional
{"label":…,"profile":…}(the token is in that answer only) andDELETE /api/sets/<id>/keys/<keyID>;PUT /api/sets/<id>withtokenstill replaces
the primary key, and creates one when the set has no active key. A new key
reaches a running set by reload, not by restart: the transport takes the whole
authTokenslist, compares the presented token with every active key's
sha256 in constant time per element, and takes the caller's fingerprint — and
so its profile and rate limit — from the token that matched. Keys issued
before this release read as a one-element list, and anotebecomes the key's
label. -
The panel states where it is reachable:
MCPRIG_PUBLIC_URLsets the public
address (http(s)://host[:port][/prefix]), and it becomes the only source of
the OAuth redirect (<publicUrl>/oauth/callback, path prefix kept) and of the
decision that the session cookie isSecure. A bad value — no scheme, no
host, credentials, query, fragment or a
trailing/— stops the panel at startup instead of falling back silently.
Without it the redirect is built fromMCPRIG_HOSTandMCPRIG_PORTas
before.mcprig doctorprints the effective redirect and flags one that
cannot receive the code (no host, port0, loopback behindhttps). -
The panel can serve TLS itself:
MCPRIG_TLS_CERTandMCPRIG_TLS_KEYmake it
listen over https. Both are required together, and an unreadable file fails
startup loudly instead of quietly dropping to http. -
A set can cap how often a caller calls its tools:
rateLimit: {"perMinute": 120, "burst": 20, "perTool": {"serena/find_symbol": {"perMinute": 10, "burst": 2}}}in the set's record. Each inner call of
invoke_manycounts on its own, a caller is identified by the fingerprint of
its key (in an open set, by its address), and a refused call is an ordinary tool error that never reaches a
server —/statsreports it asthrottled. No block, or zeros, means no
limit. The field is edited inregistry.jsonor over the API, asnotify. -
A truncated tool answer can be fetched over HTTP:
GET /tool-out/<name>
serves the saved file behind the same guard as/mcp— bearer key,
allowedHosts, Origin — and the truncation marker names that URL when the set
runs over HTTP with a key. The file is never served anonymously, even in an
open set, and a name pastMCPRIG_TOOLOUT_TTL(24h by default) answers410. -
mcprig statusshows panel health and, per set, its state, port, tool count,
key and last call;mcprig logsprints a set's or a server's log tail with
-fand--json. Both take--panel(defaultMCPRIG_PANEL_URL, else
MCPRIG_PORT) and log in withMCPRIG_PASSWORDor the reverse-proxy secret. -
A set can report failures on its own:
notify: {"webhook": "https://…"}in the
set's record sendsset.crashed,set.restart_failedand
server.unhealthyas one JSON POST per transition — 5 s timeout, three
attempts, and never a token,envor command in the body. The address itself
never reaches the log, even when it carries a query token. -
The panel has a Log drawer per server and Show more under the call
history: the drawer reads the set's log filtered by server name, and the
button pages/api/callsbybeforeuntil a page comes back short. Paging
pauses the 5 s refresh and shows resume next to it, which returns to the
first page. -
A key can be narrowed to a profile of its set:
profiles: {"reader": {"deny": ["*/*_write"], "allow": ["*"]}}in the set's record, and a key
issued with{"profile": "reader"}(a400lists the set's profiles, or
says the set has none). The profile applies afterpermissionsand only
narrows — a rule's denial keeps its own message.list_commandshides what
the profile refuses,invoke_commandandinvoke_manyrefuse it as a tool
error naming the profile, and a key without a profile behaves as before.
The field is edited inregistry.jsonor over the API, asnotify. -
Tracing crosses the proxy: a client's W3C
traceparentin_meta(SEP-414)
is validated (version00, non-zero ids, noff), becomes the parent of the
call's span and goes down to the upstream in the_metaofinvoke_command,
read_resourceandget_prompt; nothing else from the client's_metais
forwarded. Spans —mcp.tool.call,mcp.server.connect,mcp.reload, with
names and outcome only, never a key or its fingerprint — are exported over
OTLP/HTTP whenOTEL_EXPORTER_OTLP_ENDPOINTis set, and cost nothing (no
recording, no network) when it is not. -
A changed tool descriptor is quarantined: the set remembers
name,
description,inputSchema,outputSchemaandannotationshashed into
sets/<id>/tools.lock.json, approves a tool the first time it appears
(approvedBy: "auto") and refusesinvoke_commandandinvoke_manyfor that
tool once the descriptor differs —isError,_meta.lazyMcpPolicy: "quarantine"and a message with the old and new hash and the command that
lifts the block. Neighbours on the same server and other calls of the same
batch run as usual;list_commandsmarks the toolchanged: true.
A person approves the new descriptor from the Tools drawer of the panel
(GET /api/sets/<id>/tools,POST /api/sets/<id>/tools/approve,{"all": true}included) or with the newmcprig approve --set <id> <server/tool>…/--all,
which talks to the same routes. The check does not depend on a warm
catalogue: a cold cache is filled first, and a catalogue that cannot be read
refuses the call (list_commands {refresh:true}no longer empties the
neighbours' caches either). The panel's browse session applies the same
record — a changed tool is marked in its listing and refused with403
before the downstream sees the call. The record survives a set restart, and
entries of a server removed from a set are pruned when the set config is
rebuilt. Without a set data directory (MCPRIG_DATA_DIR) there is no record
and no quarantine.
Changed
X-Forwarded-ProtoandX-Forwarded-Hostare trusted only when
MCPRIG_PROXY_SECRETis set. Without it both are ignored for the Host and
Origin checks and for thehttpsmark, so a client can no longer claim
httpsor a different public host by sending a header; the OAuth redirect
never depends on them. The session cookie isSecure
whenever the public address ishttpsor the panel has its own certificate,
no longer only when the request itself arrived over TLS.- Login failures are counted by the socket address and by the address from
X-Forwarded-For: rotating the header behind a trusted proxy no longer
bypasses the five-per-minute limit, and a request is refused when either
counter is over it. - Browse sessions in the panel now apply the server's
permissionsto tool
calls, the same gateinvoke_commanduses: a denied...