TopicForge

The safety-first read-only MCP for ROS2 robotics: graph introspection, bag analysis, and multi-vendor OMG DDS-RTPS observability. TopicForge lets AI agents inspect your ROS2 graph, recorded bag files, and the raw DDS layer beneath ROS, without ever publishing back to the bus. Eleven typed read-only tools (5 ROS2 graph + 3 DDS + 3 observability/bag) share a single Pydantic envelope so an LLM caller reads one schema across the whole stack. It joins a domain through one OSS Python participant (Eclipse CycloneDDS from PyPI, or eProsima Fast DDS with a binding you build from eProsima's sources) and reads the builtin discovery topics the OMG protocol standardizes, so it observes every conformant vendor on the wire (RTI Connext, OpenDDS, CoreDX, Dust DDS in Rust) with no proprietary binding. See docs/dds-interop-matrix.md for the canonical multi-vendor positioning and the OMG May 2025 interop reference.

TopicForge is a production-minded MCP (Model Context Protocol) server that lets AI agents (such as Claude) inspect ROS2 topics, analyze ROS bag files, and (since v0.2.0) observe the raw DDS layer through a clean, structured tool interface. It is read-only by architecture, not by configuration: there is no write path to misconfigure, no permission system to audit, no liability conversation to have. The MCP client can see the robot stack; it cannot touch it.

This stance matters because the ROS-MCP space is no longer empty: general-purpose ROS-MCP servers exist that let an LLM publish topics, call services, and command robots. That shape is fine for demos; it is untenable for production fleets, defense systems, automotive AUTOSAR Adaptive surfaces, or anything safety-certified. TopicForge is the read-only alternative for those audiences, plus the robotics developers, ML/CV engineers, and teams that want their AI tooling to understand their robotics stack without commanding it.

Why it exists

LLM agents are good at reasoning over text, but ROS2 introspection lives in a CLI + DDS world they cannot directly reach. Without grounding, an LLM will hallucinate topic names, message types, and bag contents. TopicForge bridges that gap with a small, well-typed set of MCP tools, all read-only, all returning frozen Pydantic schemas that a downstream agent can parse without ambiguity:

Outputs are structured, JSON-serializable, and stable across runtime modes - they look the same whether the server is talking to a real robot or to its built-in mock fixtures. Every response except health_check carries a mode_effective field ("live" or "mock") so a downstream LLM can tell a real graph from the demo fixtures without re-reading health_check. health_check itself reports mode (the mode of the adapter actually serving requests) next to requested_mode.

30-second demo without ROS2

The mock adapter ships deterministic fixtures for a small differential robot (LIDAR + RGB camera). You do not need ROS2 installed to try the full tool surface - a clean venv on Python 3.10 or later is enough.

pip install topicforge
TOPICFORGE_MODE=mock python -m topicforge
# Windows PowerShell: $env:TOPICFORGE_MODE="mock"; python -m topicforge

Point any MCP client (Claude Desktop, see below) at this server and ask it to list the topics or analyze /tmp/demo.mcap - every tool returns realistic, typed payloads.

Quickstart

pip install topicforge
python -m topicforge --help
TOPICFORGE_MODE=mock python -m topicforge

Architecture

+----------------------+
|   MCP client (LLM)   |
+----------+-----------+
           |  (stdio, MCP protocol)
           v
+----------+-----------+
|  topicforge.server   |   FastMCP entrypoint, lifecycle, tool registration
+----------+-----------+
           |
           v
+----------+-----------+
|  topicforge.tools    |   Thin handlers - validate, delegate, serialize
+----------+-----------+
           |
           v
+----------+-----------+
| topicforge.services  |   Inspector / Health - orchestration & validation
+----------+-----------+
           |
           v
+----------+-----------+
| topicforge.adapters  |   ros2_live  - subprocess wrappers over `ros2` CLI
|                      |   ros2_mock  - deterministic fixtures
+----------------------+

Layers are strictly separated:

• server/ wires the whole graph and exposes build_app(settings).
• tools/ registers MCP tools on FastMCP. Handlers never call ROS directly.
• services/ validate inputs and orchestrate calls.
• adapters/ are the only code that knows how to talk to a specific backend. New backends (e.g. an rclpy-based adapter) plug in by implementing the RosAdapter protocol.
• models/ holds Pydantic schemas - the contract with MCP clients.
• config/ resolves runtime settings from the environment.

Runtime modes

Mode is selected via the TOPICFORGE_MODE environment variable. live and auto degrade instead of failing: if ros2 is not on PATH and no DDS backend comes up, the server serves the mock fixtures, and health_check reports mode: "mock" next to requested_mode. Only TOPICFORGE_MODE=mock forces fixtures unconditionally. The DDS backend is selected separately, see Multi-vendor DDS support.

Install from source

Requires Python 3.10+ (ROS 2 Humble on Ubuntu 22.04 ships Python 3.10).

git clone https://github.com/yaniswav/TopicForge.git
cd TopicForge
python -m venv .venv
source .venv/bin/activate          # Linux / macOS
# .venv\Scripts\Activate.ps1       # Windows PowerShell
pip install -e ".[dev]"

Or, if you have make:

make dev

Run

Mock mode (no ROS2 required)

TOPICFORGE_MODE=mock python -m topicforge

Or:

make run-mock

Live mode (requires ROS2)

Source your ROS2 distribution first, then:

source /opt/ros/humble/setup.bash
TOPICFORGE_MODE=live python -m topicforge

TopicForge invokes the ros2 CLI under the hood, so it does not require rclpy to be importable. This keeps the live adapter portable across ROS2 distros.

Multi-vendor DDS support (v0.3.0+)

Beyond ROS2 graph introspection, TopicForge observes the raw DDS bus directly via one of two OSS Python adapters (Eclipse CycloneDDS or eProsima Fast DDS) each joining as a read-only DDS-RTPS participant. By the OMG-DDS-RTPS protocol guarantee, both adapters see every conformant participant on the domain (RTI Connext, OpenDDS, CoreDX, Dust DDS in Rust, InterCOM, etc.) regardless of host language. The adapters read discovery data; they do not decode the payload of user topics (see the peek_dds_samples paragraph below).

Validation status. No DDS adapter has been exercised against a live multi-vendor bus yet. The unit tests run against the mock backend and against binding-free helpers; the real-bus rig under scripts/integration/ is a dispatch shell that spawns no publisher and validates nothing (see scripts/integration/README.md). The multi-vendor claim rests on the RTPS protocol guarantee, not on a recorded interop run by this project. For the multi-vendor positioning and its limits, see docs/dds-interop-matrix.md and the OMG May 2025 interop reference.

Useful for non-ROS DDS stacks (defense, aerospace, automotive AUTOSAR Adaptive, industrial integration) and for diagnosing why a ROS2 subscriber isn't receiving when the graph says it should. Same safety-first contract : read-only by architecture. The MiddlewareAdapter protocol does not expose a write method on any backend.

Install the Cyclone backend from PyPI:

pip install topicforge[dds-cyclone]   # Eclipse CycloneDDS
pip install topicforge[dds]           # same thing today ([dds] resolves to [dds-cyclone])

There is no Fast DDS extra. The fastdds Python binding is not published on PyPI, so TopicForge no longer declares it (0.5.3 removed [dds-fast], [dds-opendds], [dds-dust] and [dds-all-oss]; an unclaimed package name in a published extra is a dependency-confusion risk). The Fast DDS adapter still works if you build eProsima's Python binding from their sources and make it importable in the same environment; see docs/DDS_QUICKSTART.md.

Then select a backend (or let auto-detect pick):

TOPICFORGE_DDS_BACKEND=cyclone python -m topicforge
# or, with a source-built Fast DDS binding:
TOPICFORGE_DDS_BACKEND=fast python -m topicforge
# or let TopicForge probe the importable bindings, in this order:
#   fast > cyclone > mock
TOPICFORGE_DDS_BACKEND=auto python -m topicforge

An explicit TOPICFORGE_DDS_BACKEND (cyclone, fast or auto) is honoured whether or not ros2 is on PATH, and in any TOPICFORGE_MODE except mock: on a host without ROS 2 you get a DDS-only adapter, not fixtures. If the chosen binding is missing or the participant cannot start, the server logs a warning that names the cause and falls back to the ROS2 CLI alone, or to the mock fixtures when that is not available either. The default is TOPICFORGE_DDS_BACKEND=mock, which leaves the DDS tools to the mock adapter only.

What the OSS install covers. Cyclone and Fast DDS are the two working backends. opendds and dust remain selectable identifiers but are permanent stubs: their adapters always report unavailable, and no install extra exists for them (pyopendds is on PyPI, but the OpenDDS adapter never uses it; there is no Dust DDS Python binding). That is not a limitation on what you can observe: because discovery runs over the OMG-standardized wire protocol, a Cyclone participant already sees RTI Connext, OpenDDS, CoreDX and Fast endpoints on the same domain. You do not need a commercial adapter to observe a commercial bus.

The values rti, opensplice, coredx and intercom were accepted for TOPICFORGE_DDS_BACKEND until 0.5.2 and are now rejected with an explicit configuration error: the licensed Pro tier is retired and the core no longer reads TOPICFORGE_LICENSE_KEY or loads any topicforge_pro package. A native RTI Connext adapter still exists for the cases where the standard route is not enough (secure domains needing vendor credentials, shared-memory-only deployments, vendor-specific extensions); it needs your own RTI binding and license and is arranged as a support engagement. See docs/pro.md.

Six DDS / observability tools (in addition to the five ROS2 tools above) :

Composite adapter (v0.4.0 Phase 1+). When the ros2 CLI is on PATH and a DDS backend (cyclone or fast) is selected and starts, TopicForge instantiates both a ROS2 CLI adapter and the chosen DDS adapter and routes per tool: the ROS2 graph and bag tools (list_topics, get_topic_info, sample_messages, analyze_bag, peek_bag_samples) hit the CLI half, the DDS tools (list_participants, detect_qos_mismatches, peek_dds_samples, participant_events, topic_metrics) hit the DDS backend. ROS2-only or DDS-only setups still work: the missing half is skipped and the present half serves what it can. In a DDS-only process the ROS2 graph and bag tools raise an error that says so. The mock backend exposes all 11 tools against deterministic fixtures for local development.

peek_dds_samples payload shape. Full-fidelity on the three builtin DCPS topics (DCPSParticipant, DCPSSubscription, DCPSPublication): each sample is a structured discovery payload. User-topic payloads are not decoded (since 0.5.3, on both Cyclone and Fast). Asking for a user topic that is announced on the bus returns a single placeholder sample whose payload carries _decode_status="raw" and a _decode_note saying that decoding is disabled; _raw_bytes_hex is empty. That sample means "topic present on the bus", not "message received": no traffic was read, and nothing is fed to topic_metrics. A topic nobody announces raises an error. The previous decoding path never worked on either backend (Cyclone called the type lookup with the wrong arity and swallowed the error; Fast returned None unconditionally), and re-enabling it needs a real bus to validate against. The "full" and "partial" statuses stay in the schema, and the mock backend still emits a "full" and a "raw" example so the wire shape can be tested, but no live adapter produces them today.

topic_metrics scope. The metrics buffer is filled only when peek_dds_samples surfaces samples, and only the builtin discovery topics surface any. On a user topic the tool returns samples_observed=0. On a builtin topic, frequency_hz_observed reflects how often you call peek_dds_samples, not how often anything publishes: it is a statement about the tool's own cadence. Builtin samples carry no application sequence number and no publish timestamp, so sequence_numbers_available is false and the latency percentiles are null. frequency_hz_declared is always null, nothing populates it. Read this tool as a discovery-layer probe, not as a publish-rate monitor.

Full 5-minute walkthrough (backend selection, the canonical QoS-mismatch debugging scenario, troubleshooting) lives in docs/DDS_QUICKSTART.md. Migration history : v0.2 -> v0.3, v0.3 -> v0.4.

Configure with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "topicforge": {
      "command": "python",
      "args": ["-m", "topicforge"],
      "env": { "TOPICFORGE_MODE": "auto" }
    }
  }
}

Test

pytest
# or
make test

Tests run entirely against the mock adapter, the live adapter's pure parsers and the binding-free DDS helpers - they never require a running ROS graph. Tests that need the cyclonedds or fastdds binding skip themselves when it is absent, and the integration marker (real-bus scenarios) is deselected by default; pytest -m integration selects it, but the runner behind it is not implemented yet, see scripts/integration/README.md.

Lint & format

make lint     # ruff check
make fmt      # ruff format
make check    # both, plus tests (CI bundle)

Windows note. The Makefile uses POSIX shell syntax (VAR=value cmd, find ... -exec). Run it from Git Bash, WSL, or MSYS2. From a plain PowerShell session, invoke the underlying commands directly:

python -m ruff check src tests
python -m ruff format src tests
python -m pytest
$env:TOPICFORGE_MODE = "mock"; python -m topicforge   # equivalent of `make run-mock`

Configuration reference

See .env.example.

Telemetry

TopicForge ships an opt-in, anonymous, minimal telemetry hook. It is off by default and the OFF code path performs zero network calls: pinned by a unit test (tests/test_telemetry.py::test_build_app_off_makes_no_transport_calls).

How to opt in

TOPICFORGE_TELEMETRY=on python -m topicforge
# Windows PowerShell: $env:TOPICFORGE_TELEMETRY="on"; python -m topicforge

Accepted on-values: on, 1, true, yes, enabled (case-insensitive). Accepted off-values: unset or empty, off, 0, false, no, disabled. Any other value is a configuration error, not a silent "off": the server prints topicforge: configuration error: Invalid TOPICFORGE_TELEMETRY=... to stderr and exits with code 2, so a typo such as TOPICFORGE_TELEMETRY=ture cannot leave you believing telemetry is on, or off, when it is not.

How to opt out

Unset the variable, set it to off, or just don't touch it. Opt-out is the default.

Exactly what is sent

When telemetry is on, each MCP tool call emits a single event with only these six fields:

What is never sent

• Topic names, message types, message payloads
• Bag file paths or bag contents
• Hostnames, usernames, IP addresses, ROS distro, environment variables
• Stack traces, error messages, or any free-form text
• Any persistent identifier: session_id is regenerated on every server start

The payload shape is fenced by tests/test_telemetry.py::test_payload_contains_only_whitelisted_keys. Adding a field there requires a matching change in this section.

Where the code lives

The complete telemetry implementation is in src/topicforge/telemetry/. Read it in under five minutes. The default transport is a structured log line (no HTTP endpoint yet); a future S3-backed endpoint will plug into the same Transport callable without touching tool handlers.

Security model

TopicForge is designed for local trust: it runs as a subprocess of your MCP client (Claude Desktop, Claude Code) on a machine you control, and inspects your own ROS2 graph or your own bag files. It is not hardened for adversarial inputs.

• TOPICFORGE_ROS2_BIN accepts an arbitrary path - if you point it at a malicious binary, TopicForge will execute it. Treat the variable the way you treat PATH.
• analyze_bag opens whatever path the MCP client passes (no workspace isolation, no symlink restriction). The threat model assumes the client is your trusted agent acting on your behalf.
• All ros2 CLI invocations use subprocess.run with an argument list - never shell=True. Topic names are validated against a strict allowlist (^/[A-Za-z0-9_/]+$) before being passed to the CLI.
• TopicForge does not load third-party code at startup. Until 0.5.2 the server imported any installed package named topicforge_pro and handed it the MCP server instance; that hook never registered a tool, and it was removed in 0.5.3 because it was an opening for a package of that name to add write tools.
• No outbound network calls by default. Since v0.1.1, opt-in anonymous usage telemetry is available behind TOPICFORGE_TELEMETRY=on: see Telemetry for the exact payload and opt-out instructions. When off (the default), the OFF code path is a verified no-op.

Before exposing TopicForge to untrusted MCP clients (hosted endpoints, shared environments), add path isolation and revisit the TOPICFORGE_ROS2_BIN policy.

MVP limitations

• sample_messages in live mode uses ros2 topic echo --csv --once with a short timeout; topics with no current publisher will return an empty sample. MessageSample.timestamp_ns is the message's header.stamp (publish time) for Header-stamped messages and 0 for headerless types (std_msgs/String, geometry_msgs/Twist, ...); surfacing the rmw receive timestamp for arbitrary types waits on the future rclpy-backed adapter.
• sample_messages silently clamps count to 50 to keep tool output bounded; requests for more than 50 messages return at most 50 (the SampleResult.count field reflects what was actually returned).
• analyze_bag in live mode shells out to ros2 bag info and parses its text output; it does not use rosbags. Deep anomaly detection is mock-only for now. peek_bag_samples is the only bag tool that reads the file itself, through rosbags (pip install topicforge[bags]), and it is served only by the ROS2 CLI adapter (alone or as the ROS half of a composite) or by the mock adapter. In a DDS-only process it raises the "DDS observability only" error.
• peek_dds_samples does not decode user-topic payloads, and topic_metrics only has data for the builtin discovery topics (details in Multi-vendor DDS support).
• No DDS adapter has been validated against a live multi-vendor bus, and DDS Security (authenticated or encrypted domains) is not handled: a participant without credentials sees an empty secure bus.
• detect_qos_mismatches covers four QoS policies (Reliability, Durability, History, Deadline). Liveliness, Ownership and Partition are not checked.
• No streaming / push subscriptions in the MVP. Tools are strictly request/response.
• Live adapter is CLI-based, not rclpy-based - by design, for portability.

Roadmap

See docs/product-plan.md for the full product trajectory.

Near-term additions on the bench:

• rclpy-backed live adapter for faster & richer sampling (per-message rmw receive timestamps, windowed sampling): gated on external user demand
• Validating the Cyclone and Fast adapters against a real multi-vendor bus, then re-enabling user-topic payload decoding on top of that
• Migration to the MCP SDK 2.x API (the mcp>=1.0.0,<2 pin is a stopgap)
• Extended QoS coverage (Liveliness, Ownership, Partition, TimeBasedFilter, LatencyBudget)
• URDF inspector / validator and bag anomaly detection (clock jumps, gaps, dropped frames, TF tree health): candidate engagement deliverables, not a shipped product tier (see docs/product-plan.md)
• Dataset export helpers (rosbag -> COCO / HF Datasets)
• Synthetic data pipeline controller (Blender, Gazebo, Isaac Sim)

The licensed Pro tier is retired. Commercial support and integration work, including the RTI Connext adapter, is handled by direct agreement: see docs/pro.md.

Project layout

topicforge-mcp/
├── README.md                  # You are here
├── Makefile                   # Common developer tasks
├── pyproject.toml             # Build & tooling config
├── .env.example               # Example runtime configuration
├── docs/
│   └── product-plan.md        # Product strategy & roadmap
├── src/topicforge/
│   ├── __main__.py            # `python -m topicforge`
│   ├── server/                # MCP bootstrap & lifecycle
│   ├── tools/                 # MCP tool definitions
│   ├── services/              # Domain orchestration
│   ├── adapters/
│   │   ├── ros2_live/         # `ros2` CLI wrappers
│   │   ├── ros2_mock/         # Deterministic fixtures (incl. DDS)
│   │   ├── dds_cyclone/       # Eclipse CycloneDDS adapter (lazy import)
│   │   ├── dds_fast/          # eProsima Fast DDS adapter (lazy import)
│   │   └── common/            # Vendor-neutral helpers + pure QoS analyzer
│   ├── models/                # Pydantic schemas
│   └── config/                # Settings & mode resolution
└── tests/                     # Pytest suite (mock-only, no ROS2/DDS required)

License

MIT - see LICENSE.