aiMCPGate

Русская версия — README_RU.md.

A gateway / proxy for MCP servers (Model Context Protocol) written in Go. It presents itself to an MCP client (Claude Code, Cursor, etc.) as one MCP server, while under the hood it multiplexes calls across several upstream MCP servers, aggregates their tools, prompts and resources into one catalog, and logs every call.

Status: MVP complete (Stages 0–6) + post-MVP Stages 7–18 shipped, latest release v0.5.0. Phase 1 — multiplexing stdio upstreams behind a stdio endpoint with a call log; Phase 2 — HTTP/SSE client-facing transport, HTTP upstreams, a CLI log viewer (mcp-gate logs); release pipeline (goreleaser, cross-compiled for linux/darwin/windows × amd64/arm64, no CGO). Post-MVP added upstream auto-restart, hot config reload, tool filtering/renaming, doctor, and — in v0.3.0 — full prompts/resources/ resources/templates/completion aggregation, ping, progress forwarding and real cancellation, logging/setLevel fan-out, per-upstream call limits (rate limit / concurrency / result truncation / timeout), a lazy catalog and tools/list pagination, and SSE server→client streams on both the client and the upstream side. v0.4.0 completed the server→client direction: all three server-initiated methods — elicitation/create, sampling/createMessage and roots/list — are proxied in all four transport combinations (stdio or HTTP on the client side × stdio or HTTP on the upstream side); the gateway now declares to an upstream exactly the capabilities its own client declared instead of a blanket {}; the HTTP transport gained server-side Mcp-Session-Id sessions with DELETE /mcp termination. v0.5.0 adds operator observability (Stage 18): eight event kinds — upstream start failures and supervisor give-ups, dropped notifications and server→client requests, an HTTP upstream with no GET SSE, catalog collisions and bad URI templates, and a result that silently bypassed max_result_bytes — now land in the call journal (mcp-gate logs) instead of a stderr an MCP client usually owns; config parsing became strict (unknown/misspelled keys are fatal). It also closes the client-facing half of the guard/truncation story: a tools/call refused by the rate-limit or concurrency guard now returns its own JSON-RPC error code -32029 with machine-readable data: {"retryable":true,"reason":...} instead of an indistinguishable -32603, and a non-text result that bypassed max_result_bytes carries a result._meta marker (content stays byte-for-byte untouched). Finally, auth_token referencing an unset environment variable now refuses to start the gateway instead of silently disabling HTTP authentication.

Upgrading to v0.5.0 — three behaviour changes, none touch the config file format itself:

• Config parsing is now strict. A config with an unknown or misspelled top-level or per-upstream key, which used to be silently ignored, now fails to load. Fix the key name (the error names it) or remove it.

• auth_token: ${VAR} with an unset VAR now refuses to start, naming the variable. Before, it silently became an empty token — which, on an HTTP gateway, disabled the bearer check entirely with no warning. Set the variable (or pass --env-file), or remove auth_token to run without authentication on purpose.

• The call journal (log_file / calls.jsonl) gained a second record kind, "kind":"event", alongside the existing call records. A binary at v0.4.0 or older reading a v0.5.0 journal renders an event line as a sparse ERR entry rather than failing — read a journal with the same or a newer binary than the one that wrote it.

Upgrading to v0.4.0: no config-file change, but two observable HTTP-mode behaviour changes — a session id is now mandatory on POST /mcp after initialize (the header is returned by the initialize response), and the upstream registry starts lazily on the first real MCP request instead of at process start.

Not implemented: a per-client access policy.

Releases

Cross-platform binaries are built via goreleaser (.goreleaser.yaml): linux/darwin/windows × amd64/arm64, no CGO, the version is baked in via -ldflags -X main.version=..., checksums land in SHA256SUMS. Local dry run: goreleaser release --snapshot --clean.

Related MCP server: mcpproxy-go

Install from MCP registry

Besides the raw release binaries, the gateway ships as an OCI image on GitHub Container Registry and as an npm wrapper package — the two formats MCP registries install from.

Docker:

docker run --rm -i -v $(pwd)/config.yaml:/config.yaml ghcr.io/akomyagin/aimcpgate serve

-i is mandatory: the gateway talks MCP over stdio, so the client must keep stdin open (without it the container sees EOF and exits immediately). The image has no config of its own, so mount yours — the example above mounts it onto the default path /config.yaml; any other path works with serve -c.

To reproduce a registry sandbox check (Glama.ai etc.) without any real upstream, use the demo config baked into the image — this exact command is what a sandbox should run:

docker run --rm -i ghcr.io/akomyagin/aimcpgate serve -c /demo.config.yaml

npx (downloads the prebuilt binary for your platform on first install and verifies its SHA256 checksum):

npx aimcpgate serve -c ./config.yaml

Image policy: the OCI image contains only the mcp-gate binary — no runtimes for stdio upstreams (no node/npx, python, shells). If your config launches stdio upstream servers, extend the image yourself and install what they need; HTTP upstreams work out of the box (CA certificates are included).

Demo config: demo.config.yaml and the hidden __demo-echo subcommand exist only so registry sandboxes (Glama.ai) can introspect the gateway without any real upstream — never use them in a real deployment.

Running CLI commands inside a container

doctor, catalog, call and logs are how an operator inspects a deployment. Three facts decide how they must be invoked inside a container:

1. The binary is /mcp-gate and it is NOT in $PATH. The Dockerfile does COPY mcp-gate /mcp-gate and ENTRYPOINT ["/mcp-gate"] — nothing puts it on a search path (check the Dockerfile if this ever looks wrong). So the obvious form fails:

   $ docker exec mcp-gate mcp-gate catalog -c /config.yaml
   OCI runtime exec failed: exec failed: unable to start container process: exec: "mcp-gate": executable file not found in $PATH

   Use the absolute path instead — that is the only difference.

2. The image is distroless, so there is no shell at all. The base is gcr.io/distroless/static-debian12:nonroot, which ships the binary and CA certificates and nothing else. docker exec mcp-gate sh -c '…' fails the same way sh is simply not there, and there is no ls/cat to look around with. Keep pipes, globbing and redirection on the HOST side of the command.

3. docker exec starts a NEW process; it does not query the running serve. doctor, catalog and call build their own registry, open their own connections to the upstreams, report and exit. Their output is therefore upstream reachability right now, not the state of the live gateway: if the running process lost an upstream and dropped it from its catalog, these commands will not show that. They also keep the call journal clean — they run with journaling disabled, so a call you make this way does not appear in logs.

docker exec mcp-gate /mcp-gate version
docker exec mcp-gate /mcp-gate doctor  -c /config.yaml
docker exec mcp-gate /mcp-gate catalog -c /config.yaml
docker exec mcp-gate /mcp-gate call demo__echo '{"text":"hi"}' -c /config.yaml
docker exec mcp-gate /mcp-gate logs    -c /config.yaml --tail 50

The commands assume a container started detached and named, e.g. docker run -d --name mcp-gate … — unlike the foreground docker run --rm -i … example above, which exits as soon as its stdio client disconnects and leaves nothing for docker exec to reach. The config is assumed mounted on the default path /config.yaml, as in that same example; demo__echo stands in for a tool from your own catalog. A few caveats:

• logs is the exception to fact 3: it reads the journal FILE the running gateway writes, so it does reflect the live process. That requires log_file in the mounted config to point at a path visible inside the container, and a volume mounted there — otherwise the journal goes to the container's stderr (i.e. to docker logs) and mcp-gate logs has nothing to read. -c is what tells it where the journal is; --file overrides it.

• This is really about HTTP mode. In stdio mode the MCP client spawns and owns the container, so there is usually no long-lived container to exec into. A gateway you can inspect is one started separately (docker run -d --name mcp-gate …) with transport: http.

• HTTP mode needs a non-default listen_addr. The default is 127.0.0.1:28080 — loopback INSIDE the container, unreachable from the host even with -p. Set listen_addr: 0.0.0.0:<port> in the config; the gateway then refuses to start without an auth_token, on purpose ("the HTTP endpoint would be reachable from the network without authentication").

Why

An active MCP user typically has several servers configured (filesystem, GitHub, search, custom ones), each one duplicated in every client's own config. aiMCPGate gives you:

• One entry point — a single MCP endpoint instead of N entries in the client config.

• One catalog — every upstream server's tools and prompts merged together (namespaced as <upstream>__<tool> so names never collide), plus their resources and resource templates (addressed by URI, so never renamed).

• A call log — which upstream, which tool, when, success/failure. This is the value added on top of "just a proxy".

Solo pet project: the priority is learning Go (concurrency, os/exec, JSON-RPC 2.0, the stdio and HTTP/SSE transports). Cost — $0/month by default (a local process), no telemetry.

How it works (short version)

MCP client ──stdio/HTTP──▶ aiMCPGate ──JSON-RPC──▶ upstream A (stdio)
                              │        ├─────────▶ upstream B (stdio)
                          call log     └─────────▶ upstream C (http, Phase 2)

MVP (two phases)

• Phase 1 — multiplexing 2+ stdio upstreams behind one stdio endpoint (the same transport Claude Code speaks) plus basic logging.

• Phase 2 — HTTP/SSE transport, HTTP upstream servers, a log viewer (the CLI one was built; the web view was deliberately dropped), optionally an access policy — that one was considered and declined.

Build

export PATH="$HOME/sdk/go/bin:$PATH"   # if go isn't already on PATH
go build ./...
go vet ./...
go test -race ./...

go run ./cmd version

Usage

# stdio mode (the client launches the gateway as a subprocess):
mcp-gate serve --config ./config.yaml

# http mode (transport: http in the config) — endpoint at http://<listen_addr>/mcp;
# every request after initialize carries the issued Mcp-Session-Id (see below):
mcp-gate serve --config ./config-http.yaml

# check every enabled upstream once (launch → handshake → tools/list) and print
# a per-upstream OK/FAIL table; exit code is non-zero if any upstream failed
# (scriptable for CI/cron), no auto-restart, no call logging — one pass then exit:
mcp-gate doctor --config ./config.yaml

# call one aggregated tool once from the shell (single bring-up, no supervisor —
# the fastest way to debug a config, a filter or a rename without a live client):
mcp-gate call github__search_repositories '{"query":"mcp"}' --config ./config.yaml

# report the aggregated catalog size per upstream (tools / bytes / ~tokens) plus
# the heaviest individual tools — the data behind allow-list / strip decisions:
mcp-gate catalog --config ./config.yaml --top 20

# view the journal — tool calls AND operator events (last 50 lines; filter by
# upstream/tool/status):
mcp-gate logs --file ./logs/calls.jsonl --tail 50
mcp-gate logs --config ./config.yaml --upstream github --status err
# show ONLY the operator events (see "Operator events" below):
mcp-gate logs --config ./config.yaml --events
# keep watching the log as it grows, or aggregate it instead of listing records
# (--follow and --stats are mutually exclusive):
mcp-gate logs --config ./config.yaml --follow
mcp-gate logs --config ./config.yaml --stats

# generate a random auth token (for the HTTP transport) and see how to wire it in:
mcp-gate token --generate
# print the auth token currently set in the config:
mcp-gate token --config ./config-http.yaml

# print ready-to-paste MCP client config snippets (Claude Code / Cursor / Claude
# Desktop) for whichever transport the config uses: a launch command for stdio, or
# the endpoint URL plus the Bearer header (when auth_token is set) for http:
mcp-gate client-config --config ./config.yaml

# print a SKILL.md teaching an agent how to use the aggregated catalog
# (built-in text by default; overridable via skill_file in the config):
mcp-gate skill > .claude/skills/mcp-gate/SKILL.md

# shell completions (cobra's built-in command; the release archives also ship
# pre-generated ones):
mcp-gate completion bash > /etc/bash_completion.d/mcp-gate

All commands except token --generate, completion and skill (which falls back to a built-in guide) load the config: pass --config, or drop a config.yaml next to the binary (see Configuration below).

serve, doctor, call and catalog also accept --env-file ./.env — a minimal KEY=VALUE parser applied before the config is loaded, so ${VAR} references inside the config resolve from that file. The real process environment always wins over the file.

HTTP sessions (Mcp-Session-Id)

In http mode the gateway runs Streamable HTTP sessions: the reply to initialize carries an Mcp-Session-Id header, and every request after it — POST, the GET SSE stream, DELETE — must send that header back. Without it the answer is 400; with an unknown or expired id, 404, which tells the client to initialize again. A session is released by DELETE /mcp (204), or after 30 minutes with no requests — an open SSE stream counts as activity and keeps it alive.

MCP clients do all of this for you. For hand-made curl calls, take the header from the initialize response and echo it back:

SID=$(curl -sD - -o /dev/null -X POST http://127.0.0.1:28080/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1"}}}' \
  | tr -d '\r' | awk -F': ' '/^[Mm]cp-[Ss]ession-[Ii]d/{print $2}')

curl -s -X POST http://127.0.0.1:28080/mcp \
  -H 'Content-Type: application/json' -H "Mcp-Session-Id: $SID" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

curl -s -X DELETE http://127.0.0.1:28080/mcp -H "Mcp-Session-Id: $SID"

The session also makes the call log honest: every call is audited under the clientInfo of the session that made it, so several HTTP clients are told apart in calls.jsonl instead of sharing one blank client field.

Server→client requests over HTTP (elicitation, sampling, roots)

When an upstream asks something mid-call — elicitation/create, sampling/createMessage, roots/list — the question is delivered as an SSE event on the GET /mcp stream of one session, and the client answers with an ordinary POST carrying a JSON-RPC response with the same id and the same Mcp-Session-Id. Only the session the question was put to may answer it; an answer from any other session is ignored. If nobody has declared the capability with a stream open, the upstream is refused right away in the shape the spec prescribes ({"action":"decline"} for elicitation, -32601 for the other two) rather than being left to time out — and the same happens if the session is terminated while a question is outstanding.

Three consequences worth knowing:

• The upstreams are told about the capabilities of the FIRST client that initializes, and that set is fixed for the life of the process. MCP 2025-06-18 has no re-negotiation, so a second client declaring more cannot change handshakes that already happened — an upstream is never promised a capability on behalf of a client it was not told about.

• The upstreams start on the first request that needs them, not when the gateway binds its port. That is what makes the declaration above possible at all: the handshake has to happen after a client has said what it supports. If the upstreams cannot start, the client gets a JSON-RPC -32603 and the gateway exits with the error, as it did when it started them eagerly.

• The question goes to a client that declared the capability — not necessarily to the one whose call provoked it. Routing is by declared capability, and among the matching sessions the most recently active one wins; an upstream request carries nothing that says which caller it belongs to. With a single client (the normal case) this is invisible, but run two and a form raised by one client's tools/call can surface in the other's UI.

The upstream side of the same exchange works over HTTP too: a remote MCP server reached with url: may ask its question as an SSE frame — either on its long-lived GET stream or interleaved into the stream answering one of the gateway's own POSTs, which is where SDK servers put an elicitation/create raised inside a tools/call. The gateway proxies it through the same pipeline and sends the client's answer back as one ordinary POST carrying a JSON-RPC response under the server's own request id. Such an upstream is told the gateway's client capabilities by the same honest policy as a stdio one — a capability is offered only when the gateway's own client declared it, and doctor/call/catalog, which have no client at all, keep declaring exactly {}. The answer POST is not retried: an upstream that does not get it falls back on its own timeout.

Operator events in the journal

The journal at log_file holds two kinds of line: one per tool call, and one per operator event — a gateway state you would otherwise never learn about. In stdio mode the MCP client owns the terminal, so the gateway's stderr is invisible to you, and several of these conditions were only ever logged at debug level. They are now written to the same file mcp-gate logs reads:

Events show up inline with calls, marked EVT; mcp-gate logs --events shows only them, and --stats gains a per-event table. --tool and --status are call-only filters, so events are excluded while either is set (--upstream applies to both). One consequence worth knowing: notification_dropped names no upstream — a drop is a property of the subscriber whose buffer was full, not of whoever sent the notification — so --upstream X never shows it. Look for it without that filter. Repeated drops are coalesced — the first one is written at once, further ones within a minute are counted into the count= of the next line for that key, and the remainder is flushed at shutdown. A line that carries such a backlog says so in its detail=, naming the time of the oldest occurrence it folds in — the line's own timestamp is the newest one, so the two together bound when the burst actually happened.

Two practical notes:

• Set log_file. With it empty the journal goes to stderr, which in stdio mode belongs to the MCP client — the events would be written where you cannot see them.

• Read a journal with the same (or a newer) binary that wrote it. Events carry a "kind" field older versions do not know, so mcp-gate logs from ≤ v0.4.0 renders them as sparse, mostly empty records.

Nothing about this is visible to the MCP client: no error codes, result bodies or capabilities changed — the events go to the journal only.

A call the gateway could not route is not an event — it is an ordinary failed CALL line. A client asking for a tool name no upstream provides gets a CallRecord like any other, with its upstream column set to the sentinel (unrouted); mcp-gate logs --upstream '(unrouted)' selects exactly those lines and nothing else. A second, distinct case looks almost the same but names a real upstream instead: the route exists (the tool is in the catalog) but the upstream's connection is gone (restarting or dropped) — that line carries the real upstream name, so filter for it with --upstream <name> as usual rather than the sentinel.

Reloading config (SIGHUP)

The gateway reloads its configuration live on SIGHUP — no restart, no dropped client connection. Edit config.yaml and send the signal:

kill -HUP $(pgrep -f 'mcp-gate serve')

On reload the gateway diffs the new config against the running upstreams and applies the minimum change: newly added upstreams are launched, removed (or enabled: false) ones are shut down, upstreams whose launch fields (command/args/url/env/headers) changed are relaunched, and upstreams where only the tool filter changed (allow/deny/rename, or the catalog projection rules strip_annotations/strip_output_schema/max_description/ describe) are re-projected without any restart. Call limits (rate_limit, max_concurrent, max_result_bytes, call_timeout — global or per-upstream) are also applied live: they never require a relaunch, the next call simply uses the new values. Unchanged upstreams keep running untouched. A bad edit (invalid YAML, failed validation) is logged and ignored — the currently running config stays live, so a typo never takes the gateway down.

Behavioural note: since the gateway installs a SIGHUP handler, SIGHUP no longer terminates the process the way the OS default would. To stop the gateway use Ctrl-C, SIGINT, or SIGTERM.

SIGHUP is Unix-only. On Windows — or anywhere you would rather not send signals — use the opt-in polling alternative instead:

mcp-gate serve --config ./config.yaml --watch-config        # bare flag = poll every 2s
mcp-gate serve --config ./config.yaml --watch-config=10s    # note the "=", not a space

It fingerprints the config file on that interval and applies the same reload path SIGHUP takes. Running it alongside the SIGHUP handler is safe.

The watcher compares the file's mtime and size, and waits for that fingerprint to repeat on the next tick before it reads the file. That is what makes a two-step save (truncate, then fill) safe in practice: a writer has to hold the file in a half-written state for longer than a full polling interval to fool the check. The price is latency — a reload lands within up to two polling intervals (up to 4s on the default 2s).

On stdio the upstreams come up on the client's first request, so an edit made before any client has connected cannot be applied yet. The watcher keeps that edit and re-tries it every poll until the gateway is up, then applies it — you never have to save the file a second time to make it take. An edit refused for good (unparseable YAML, or the no-upstreams guard below) is reported once and not retried.

As a backstop on both triggers, a reload whose new config declares no upstreams at all is refused and logged: that is the signature of a half-written file, and applying it would tear down every running upstream. To remove all upstreams deliberately, restart the gateway. An explicit enabled: false is unaffected — disabling the last upstream still applies.

Configuration

Without --config, the gateway looks for config.yaml next to its own binary (e.g. if mcp-gate is installed at /etc/gate/, it looks for /etc/gate/config.yaml — regardless of the working directory it was launched from). If that file doesn't exist and --config wasn't passed either, it errors explicitly instead of starting an empty gateway. Relative paths inside the config (log_file, skill_file, debug_payload_log) resolve against the config file's own directory, not the current working directory.

Unknown keys are a startup error. The config is parsed strictly: a misspelled or unrecognized key stops the gateway with the key name and its line number, instead of being silently ignored as it once was. The concrete win: a typo in enabled can no longer leave an upstream quietly running. Custom x- scratch keys are rejected too — to share a block, put a YAML anchor on the first real upstream and merge it (<<: *anchor) into the others; anchors and merge keys work as usual.

An upstream is enabled by default: omit enabled: entirely and it is launched like any other. To keep one out of the gateway without deleting its config, disable it explicitly with enabled: false — it then appears neither in tools/list nor in mcp-gate doctor's table. Careful: a valueless enabled: (or enabled: null) reads as omitted, so commenting the value out leaves the upstream running — only the literal false disables it.

Note: the "next to the binary" lookup uses the path of the running executable. Under go run ./cmd ... that executable is a throwaway build in a temp directory, so the default lookup will not find your config.yaml — pass --config explicitly when using go run, or run a built binary.

Full example with every field — config.example.yaml. The set of upstream servers is declared in YAML; secrets (tokens) go through env/.env (${VAR} expansion at load time), never committed in the config. Each upstream sets exactly one of command (stdio subprocess) or url (HTTP server, Streamable HTTP) — the connection kind is inferred automatically.

Unset ${VAR} references behave differently per field:

• auth_token referencing an unset variable fails startup, naming the variable — an empty auth_token would silently disable the HTTP bearer check, so this is never allowed to happen quietly. To run without authentication, remove the auth_token key entirely.

• An unset variable in an upstream's env/headers is not an error: the value becomes empty and the missing secret surfaces later as a 401 from that upstream. The gateway reports it ahead of time — an unresolved_secret_var event in the journal (mcp-gate logs) and a WARN line in mcp-gate doctor.

• In stdio mode mcp-gate client-config warns (on stderr) that the operator's environment variables are not inherited by the MCP client, which launches the gateway in its own environment — set them where the client runs it.

transport: stdio            # stdio (Phase 1) | http (Phase 2)
listen_addr: "127.0.0.1:28080"  # only used for transport: http; loopback by default
# auth_token: ${AIMCPGATE_TOKEN}  # required if you widen listen_addr past loopback;
#                                 # the variable must be set or startup fails
log_file: ./logs/calls.jsonl
# debug_payload_log: ./logs/payloads.jsonl  # OPT-IN, off by default: logs raw
#                                   # arguments AND results — can contain secrets
# Optional global call limits (each can be overridden per upstream):
# rate_limit: { rps: 5, burst: 2 }  # token bucket per upstream for tools/call
#                                   # (refusal → client error -32029, retryable)
# max_result_bytes: 65536           # truncate oversized textual results (0 = off;
#                                   # non-text over-limit results get a _meta marker)
# call_timeout: 30s                 # bounds one upstream request
# How the catalog is presented to the client (both hot-reloadable):
# catalog_mode: lazy                # normal (default) | lazy: the client sees only
#                                   # gate_search_tools / gate_describe / gate_call
# page_size: 50                     # paginate tools/list (0/omitted = whole catalog;
#                                   # ignored in lazy mode)
# Auto-restart policy for crashed stdio upstreams (defaults: on, 1s→30s, 5 tries):
# restart: { enabled: true, initial_backoff: 1s, max_backoff: 30s, max_attempts: 5 }
upstreams:
  - name: filesystem        # stdio upstream
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/home/user"]
    enabled: true
  - name: github
    command: github-mcp-server
    env:
      GITHUB_TOKEN: ${GITHUB_TOKEN}   # from the environment, not hardcoded
    enabled: true
    # Optional per-upstream tool filter / catalog projection (keys are ORIGINAL
    # tool names; all editable live via SIGHUP with no upstream restart):
    # tools:
    #   allow: ["search_repositories"]  # if non-empty, only these survive
    #   deny: ["delete_repository"]     # always subtracted, even from allow
    #   rename: { search_repositories: "gh_search" }
    #   strip_annotations: true         # drop heavyweight catalog fields
    #   strip_output_schema: true
    #   max_description: 200            # truncate descriptions to N runes
    #   describe: { get_issue: "Fetch one issue." }   # replace wholesale
    # Optional per-upstream call limits (override the globals for this upstream):
    # rate_limit: { rps: 1, burst: 1 }  # rps: 0 disables the global limit here
    #                                   # (refusal → client error -32029, retryable)
    # max_concurrent: 4                 # cap on simultaneous in-flight calls
    #                                   # (refusal → client error -32029, retryable)
    # max_result_bytes: 32768           # 0 disables the global cap here
    # call_timeout: 120s                # this upstream is slow — give it longer
  - name: remote            # http upstream (Phase 2)
    url: https://mcp.example.com/mcp
    headers:
      Authorization: "Bearer ${REMOTE_MCP_TOKEN}"   # secret, never logged
    enabled: true

What the client sees when a call limit bites

Two of the call limits above surface to the MCP client (agent), not just to the operator journal:

• Guard refusals (rate_limit / max_concurrent). When the gateway turns a tools/call away because the per-upstream rate limiter or concurrency cap could not admit it, the client gets a JSON-RPC error with the gateway's own code -32029 and machine-readable data: {"retryable": true, "reason": "rate_limit" | "concurrency_limit"}. The call never reached the upstream, so an agent may wait and retry without risking double execution. Ordinary transport/routing failures keep the historical -32603, and an error an upstream itself returns is forwarded verbatim, code and data untouched — a -32029 from an upstream is not a gateway signal.

• Oversized results that cannot be truncated (max_result_bytes). Text results are shrunk with an in-content [truncated by mcp-gate: …] marker. A non-text / non-standard result that exceeds the limit but has no truncatable text (e.g. images only) is passed through whole and byte-for-byte — its content[] is never altered — but the result's _meta gains the gateway key io.github.akomyagin.aimcpgate/result-over-limit with {"limitBytes": N, "resultBytes": M} so an agent can tell the limit was bypassed. A client that does not know the key simply ignores it. The operator result_truncation_skipped journal event still fires as before.

License

MIT — see LICENSE.