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 or eProsima Fast DDS) 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 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.

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 Python 3.11 venv 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.

Install from source

Requires Python 3.11+.

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. See docs/dds-interop-matrix.md for the canonical multi-vendor positioning 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 one or both OSS backends. The v0.4.0 auto-detect framework picks whichever you actually installed:

# Single vendor: install only what you need
pip install topicforge[dds-cyclone]   # Eclipse CycloneDDS only
pip install topicforge[dds-fast]      # eProsima Fast DDS only
pip install topicforge[dds]           # both OSS backends (union)

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

TOPICFORGE_MODE=live TOPICFORGE_DDS_BACKEND=cyclone python -m topicforge
# or:
TOPICFORGE_MODE=live TOPICFORGE_DDS_BACKEND=fast python -m topicforge
# or, auto-select across the 8-vendor priority chain:
#   RTI > OpenSplice > CoreDX > InterCOM (Pro tier, if installed)
#   > OpenDDS > Fast > Cyclone > Dust > Mock (OSS fallback)
TOPICFORGE_MODE=live TOPICFORGE_DDS_BACKEND=auto python -m topicforge

What the OSS install covers. Cyclone and Fast DDS are the two working backends; OpenDDS and Dust are stubs whose upstream Python bindings do not exist yet, and their install extras fail today. 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.

A native RTI Connext adapter exists for the cases where the standard route is not enough (secure domains needing vendor credentials, shared-memory-only deployments, vendor-specific extensions). It requires the RTI Python binding and an RTI license you already hold. See docs/pro.md for commercial support and integration.

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

Composite adapter (v0.4.0 Phase 1+). When TOPICFORGE_MODE=live is paired with a DDS backend (cyclone, fast, ...), TopicForge instantiates both a ROS2 CLI adapter and the chosen DDS adapter and routes per-tool category: the 5 ROS2 tools hit the CLI, the DDS / observability tools 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. The mock backend continues to expose all 11 tools against deterministic fixtures for local development.

peek_dds_samples payload shape (v0.4.0 Phase 1.5). Full-fidelity on the 4 builtin DCPS topics (DCPSParticipant, DCPSSubscription, DCPSPublication). Arbitrary user topics return best-effort decoded samples with a _decode_status annotation : "full" (every IDL field decoded; currently a v0.4.0+ Cyclone XTypes path), "partial" (some fields decoded, others opaque), or "raw" (binding could not resolve the dynamic XTypes). The diagnostic key _decode_note carries a short explanation when the status is non-full. The wire shape is identical across Cyclone and Fast backends. Caveat (v0.5.x): on the current user-topic raw path _raw_bytes_hex is empty. A "raw" status means "topic present on the bus but not decoded", not "here are the serialized bytes to re-decode". Capturing the on-wire CDR bytes into the fallback is roadmapped ; until then use the Cyclone full/partial XTypes path for actual user-topic payloads. The 4 builtin DCPS topics are unaffected (always structured).

RTI Connext is v0.4.0+ Pro tier (BYO license; see docs/pro.md).

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 and the live adapter's pure parsers - they never require a running ROS graph. The full suite completes in well under a second.

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:

powershell

Copy

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). Anything else (including unset) keeps telemetry off.

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.
• 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. Deep anomaly detection is mock-only for now.
• 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
• Extended QoS coverage (Liveliness, Ownership, Partition, TimeBasedFilter, LatencyBudget): v0.5.x patch
• Real RtiConnextAdapter in the Pro tier (BYO RTI Connext license, gated by TOPICFORGE_LICENSE_KEY): the v0.4.0 Phase 1.5 framework is in place ; production binding pending Pro tier launch
• URDF inspector / validator MCP tools (Pro tier)
• Bag anomaly detection (clock jumps, gaps, dropped frames, TF tree health): Pro tier
• Dataset export helpers (rosbag -> COCO / HF Datasets)
• Synthetic data pipeline controller (Blender, Gazebo, Isaac Sim)
• Hosted MCP endpoint with auth (Phase 3; depends on Pro tier traction)

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.