kestrel-feature-mcp

Model Context Protocol (MCP) server management and tool bridging for Kestrel Sovereign. Discover servers in the official MCP Registry, Docker's catalog, or the bundled local catalog. Connect to public HTTPS endpoints, Docker MCP gateways and custom container servers through the native MCP client.

Installation

uv pip install kestrel-feature-mcp

Dependencies

• kestrel-sovereign-sdk[httpx2] (version constraints are declared in pyproject.toml)
• docker>=7.1.0
• mcp>=2.3,<3
• requests>=2.34.2
• anyio>=4.7,<5
• pydantic>=2.12,<3
• jsonschema>=4.26,<5

Usage

Once installed, the MCPAgent feature is automatically discovered by kestrel-sovereign via the kestrel_sovereign.features entry point.

Commands

• !mcp-gateway-start <servers> — Start Docker MCP Gateway
• !mcp-gateway-stop — Stop gateway
• !mcp-load <HTTPS-URL|catalog-name|image> — Connect a remote endpoint or load a Docker server
• !mcp-unload <connection|HTTPS-URL|container> — Disconnect a server and remove its saved connection
• !mcp-list — List running MCP servers
• !mcp-search <query> — Search official registry server names
• !mcp-catalog — List one page of the official registry
• !mcp-info <name> — Inspect an official namespaced server and version
• !mcp-docker-catalog <query> — Search Docker's catalog

Remote HTTPS servers

Use the normal load/list/unload commands for an unauthenticated public server:

!mcp-load https://learn.microsoft.com/api/mcp
!mcp-list
!mcp-unload https://learn.microsoft.com/api/mcp

Remote mode never probes the Docker daemon. Docker connects lazily on the first container operation. HTTPS endpoints use the SDK's public-address validation and DNS-pinned native transport, including certificate/hostname verification; redirects and environment proxy routing are refused. URL credentials, fragments, non-public addresses and mixed public/private DNS answers are rejected. Container command arguments are invalid for a remote endpoint.

Load results include the canonical endpoint, DID-bound connection identifier, negotiated protocol_version, mounted count and persistent status. Unload accepts that identifier or the canonical endpoint returned by load. Enabled connections are restored from the host's agent-scoped enablement store. Restore reads existing remote rows without rewriting them and rechecks saved state before tools become callable. Hosts without that capability explicitly report that connections last for the session. Saved state is read back before persistence or removal is reported. A host that silently ignores an enablement write fails explicitly. Storage errors fail the operation; a failed remote load withdraws handles and closes its transport. Reload, unload and shutdown invalidate old bound callers. If setup fails after confirmed persistence, its error reports that enablement remains saved and will restore after restart; unload it to disable that state. New catalogs become callable only after persistence is confirmed. Unload invalidates the catalog before waiting on storage removal. Configuration changes for one remote connection are serialized without locking ordinary calls or unrelated connections.

Feature shutdown refuses new connection starts and joins admitted Docker starts before reporting that resources are released. Docker configuration changes are serialized per feature; remote connections keep their separate connection gates. If container removal fails, the feature retains its resource record and Docker client for cleanup retry. Re-enabling the same feature instance completes that cleanup before allowing a fresh manager and restoring saved connections. Across connection modes, a host that supports enablement persistence must confirm the saved state and configuration by readback before publication. Gateway start also confirms that its complete saved server set matches the requested replacement. If publication fails after that acknowledgement, the error discloses the retained enablement and, for gateway start, the replaced server set. Unload the saved connection to disable restoration.

This release supports auth="none" only. It does not accept token headers or start OAuth registration/consent when a server challenges for authentication. Saved authentication modes or fields it cannot implement are refused. Vendor OAuth, encrypted grants and refresh after restart remain part of Kestrel #3679, including the native SDK gate in #3690. Connecting to a public documentation server is not the three-vendor OAuth acceptance experiment. Remote write scopes remain gated on #596.

Calling tools

Discovered server tools are mounted in the host's dynamic registry. Invoke the advertised handle through Kestrel's governed tool dispatch; its exact owner and handle determine permissions, approvals and audit records. The generic mcp_call_tool / mcp_gateway_call tools and their command aliases have been removed: allowing a generic dispatcher could otherwise bypass a DENY on its target. A host without the dynamic registry cannot mount MCP tools.

Each handle is bound to the connection that advertised its schema. Reloading or withdrawing that connection invalidates its old handles, including a call still waiting for approval. A fresh handle and approval are required after reload.

Configuration

Discovery sources

mcp_search, mcp_catalog and mcp_server_info default to source="official". Select source="local" explicitly to search or inspect bundled entries by capability/category. Docker discovery retains the separate mcp_docker_catalog tool. A local entry never overrides an official entry with the same name, and an official registry error never becomes a successful local result.

Official catalog and search requests return one page, with limit from 1 to 100 (default 20). Omit limit and cursor for the local source; explicit values are refused there. Search and catalog list latest versions only; detail lookup can request a known exact version. Pass the exact result.metadata.nextCursor back as cursor for the next page. Detail lookup accepts the namespaced server name and an exact version, or latest; results retain the actual version, remote endpoint templates, package definitions and registry metadata. Search results are discovery metadata: they do not install packages, connect to remote endpoints or send credentials. SSE definitions remain visible with an unsupported-transport label. These discovery tools work without Docker.

Registry HTTPS requests use the SDK's canonical URL policy and pinned transport, with native TLS verification, no redirects and no proxy-environment routing. Private registry addresses, malformed page envelopes, failed pagination and HTTP errors fail explicitly. Individual records that violate the published schema are rejected with visible diagnostics and a partial result; the exact next cursor remains available. result retains the unmodified response, including rejected records, and rejected_records identifies their indices, names and schema problems. Only conforming records are displayed as catalog entries. A rejected detail lookup is inspectable with its schema problems; RegistryRecord.require_valid() refuses it. Publisher text is quoted and control characters are escaped in the human-readable rendering. The request deadline and byte limit also apply to custom registry bases.

Schema diagnostics are bounded to 256 characters per path, detail or rejected name, with four problems per rejected record plus an explicit omitted-count summary. Every rejected index remains excluded from conforming results even when diagnostics are omitted. Model-facing confirmation and error text each have an 8192-character budget; complete record blocks and escaped cursor lines are omitted with a visible notice rather than cutting an opaque cursor into a different value. The full response and exact cursor remain in data.result. The next cursor has display priority over record blocks so a normal pagination token remains available when the record display budget is exhausted. Configure these bounds with KESTREL_MCP_REGISTRY_PROBLEM_CHARS, KESTREL_MCP_REGISTRY_PROBLEMS_PER_RECORD and KESTREL_MCP_REGISTRY_TEXT_CHARS. These limits apply to the official source.

Validation follows JSON Schema 2020-12 without a format checker: uri and date-time formats are annotations, not additional validity assertions. Strict JSON syntax, schema types, required fields, enumerations and other assertions are enforced. Discovery does not establish that an advertised endpoint works.

The bundled registry_openapi_2025_12_01.json is an unmodified JSON conversion of the official Registry API contract, commit 11b7b0e274d518b6f449c5ee3518a6151d9848b9. Its upstream YAML SHA-256 is 3ed9ca429754087f2c7de81aa845c1cab3286c33941ac48482f923bbd6f1442d; the converted JSON SHA-256 is 69a092a20321d6616d0e83933b1de39916bd0fa87975d7931ef3b92f825c7edc. The upstream license is included alongside it as registry_contract_LICENSE.txt. This Registry API/schema revision is independent of the MCP wire revision.

tests/fixtures/official_registry_rejected_repository.json was captured from the official registry on 2026-10-17 for ai.agentrapay/agentra version 1.0.0. Its repository: {} violates the published required-field contract. The fixture proves visible rejection and continued pagination; the schema is not relaxed to accept it.

Development

The manual vendor OAuth acceptance spike drives initial consent with the native SDK and a published Kestrel client identity. It has separate discovery and authorization modes and keeps tokens in memory. It does not enable production OAuth connections or establish refresh after restart.

uv pip install -e ".[test]"
uv run pytest

Native MCP protocol

The feature uses the official MCP 2.x Client over Streamable HTTP. Its standard version negotiation selects 2026-07-28 with modern servers and a mutually supported published revision with other servers. Tool listings expose the selected protocol_version. Container servers must expose MCP at /mcp; the standalone HTTP+SSE transport and gateway transport="sse" selection have been removed. Docker MCP Gateway 0.44.1 negotiates 2025-11-25; its lack of the newer revision is tracked in Kestrel #3683. Features requiring newer capabilities must check the negotiated connection. Gateway server lists describe requested configuration; observed tools do not prove readiness of every requested server. A call that loses its response or times out has an unconfirmed execution outcome, because the server may have already produced effects. Its tool error carries data.execution_outcome="unknown". The connection does not replay that call. A retry needs policy approval and an explicit idempotency or safe-repeat guarantee; an error is not evidence that nothing executed. Standard pre-execution JSON-RPC rejections retain their original error. Caller cancellation remains cancellation. Operator-input continuations are not implemented yet. The native client's automatic input loop is disabled; an input-required response produces an unknown-outcome error without another tool request. Managed container and gateway HTTP connections also ignore environment proxies, including for gateway bearer headers. Tool discovery follows pagination and fails on repeated cursors. Negotiation uses the connection deadline, then discovery uses the whole-call deadline. Cancelled startup joins its transport owner before returning. Shutdown has its own graceful deadline (KESTREL_MCP_CLOSE_TIMEOUT, default 15 seconds), including session termination and partially completed SDK entry. Forced cleanup gets one further budget of the same length. The owner timer survives caller cancellation; failures propagate and any unfinished owner is retained for a cleanup retry.

Upgrade the MCP feature before upgrading a host with the Phoenix 20/Pydantic AI 2 dependency set. Feature 0.4's mcp<2 requirement cannot resolve with that set. The host dependency lock must be regenerated and verified as part of the same rollout (Kestrel Sovereign epic #3679).

Container identities bind the agent DID and image, with Docker's implicit :latest tag made explicit by the same planner used for routing and unload. Catalog names resolve to their container image through the load route. New managers require owner_did; they reuse a process only when its ownership and image labels match. Each agent can reload or unload its own server independently. MCP container ports bind only to host loopback. Listings retain failed containers' identifiers so they can be unloaded even when no tools were published. Saved container entries using the former global identifiers require explicit removal with !mcp-unload <old-identifier> and reconnection with !mcp-load <image>. The current unload result does not distinguish saved-configuration removal from live-process removal; outcome reporting is tracked in #26. Shut down the old feature before upgrading to release the processes it owns. Hosts expose their agent DID through agent_id; invalid identities fail startup. The first container operation acquires the Docker client off the event loop. An unavailable local daemon disables that operation and leaves remote mode available. Missing Docker dependencies and programming errors propagate at that first container use, independently of typed daemon connection failures. Saved container connections retain their container mode when restored, even if the Docker catalog changes. Server selectors and image references reject surrounding whitespace before routing.

Container creation is recorded before Docker starts the process, so a rejected command still leaves an identifiable resource that unload can reclaim. Resource configuration transitions are serialized per container; a slow load does not block another container's unload. Model-facing handles use the image reference, while resource ownership keeps the full DID-bound identifier. Ordinary MCP tool calls remain concurrent and hold no configuration lock.

Docker catalog discovery uses the current OCI catalog JSON API. Set KESTREL_MCP_DOCKER_CATALOG to select a catalog reference (default mcp/docker-mcp-catalog). Missing metadata is pulled into Docker's cache; unreadable catalogs return an error. Catalog counts describe available definitions, while gateway status reports this process's requested configuration and observed tools.

The installation CI matrix also runs the production connection through the official MCP conformance harness, pinned to 0.2.0-alpha.12 with an npm integrity lock. It runs tools_call and request-metadata explicitly at 2026-07-28, with no expected-failure baseline. These checks cover native tool exchange, wire schemas, per-request metadata and version negotiation retry. Optional client capabilities that Kestrel does not advertise are reported as skipped by the official scenario. The OAuth, input request and extension scenarios remain outside this initial CI profile; passing these two scenarios does not establish full client conformance (epic #3679 A7).

To run the same profile locally with Python test dependencies and Node 24+:

npm ci --prefix tests/conformance --ignore-scripts --no-audit --no-fund
for scenario in tools_call request-metadata; do
  tests/conformance/node_modules/.bin/conformance client \
    --command '.venv/bin/python tests/conformance/client.py' \
    --scenario "$scenario" --spec-version 2026-07-28 --timeout 60000 \
    --output-dir "tests/conformance/results/$scenario" || exit $?
done

Update the pinned harness and lock together, rerun both scenarios, and review new required scenarios as the remaining host capabilities land.