English · 简体中文
From an exact ADS selection or blank circuit to checked data and an editable DDS result.
“Start from a blank workspace, build this testbench, run the simulation, export the data, and leave the response plot editable in DDS.”
| Built in ADS | Result left editable in DDS |
|---|---|
 |
 |
The public ADS 2026 Update 2.1 acceptance completed the whole request:
- built a six-instance AC circuit from an empty workspace;
- ran the circuit simulation and returned 31 finite rows;
- exported CSV and created two native DDS pages with rectangular and polar plots;
- saved, closed, and freshly reopened the editable result;
- completed a separate maintained Bridge acceptance for the four-stage Runtime plan in 4.312 seconds.
These are real ADS application-window captures; the DDS curve remains native and editable. The same maintained path can continue from an exact schematic, layout, library-tree, or DDS selection copied with Copy ADS Context, and can run an already-generated Momentum input on a protected sibling copy.
ADS Agent Bridge connects Codex or Pi Agent to the ADS installation and object you actually selected. Version-matched local documentation, the DE/DDS context plug-in, controlled live sessions, and bounded automation stay with ADS on the EDA host; repeated remote work reuses EDA Bridge Runtime instead of rebuilding an SSH command for every action.
New ADS functionality does not require a new Bridge wrapper. The Agent first
uses version-matched official docs and a small packaged experience library,
then runs official ADS Python through a governed workspace transaction.
Maintained operations such as design.apply and dds.create remain available
as asset-bound compiled shortcuts: they save tokens and transcription errors,
but never define the outer limit of ADS capability.
Successful governed native work returns an opaque continuation Context. A later batch can reuse its exact host-private target and content fingerprint; the Bridge still requires a new explicit program, effect, write scope, purpose, idempotency key, and validation, and rejects stale content or conflicting identity. See the continuation contract.
The ADS 2026 Update 2.1 acceptance reused one already-open graphical process. A typed patch created a resistor and labeled wire, read both objects back, and left save or discard as an explicit decision. The observed object transaction took 253 ms; exact replay returned in 3 ms with zero duplicate objects; patch-local rollback took 21 ms and removed only those created objects. Warm parameter edits completed end to end in 93–187 ms.
These are bounded observations from the disposable 0.1.0a48 acceptance on
2026-08-31, not a general latency guarantee. Codex and Pi Agent independently
passed the same create, replay, and rollback contract. See the
sanitized live-edit evidence.
Prerequisites: a licensed ADS installation, Windows or Linux, and Python 3.10 or later for the public command.
pipx install ads-agent-bridge
ads-agent setup
ads-agent quickstartInstalling the package automatically installs its compatible
eda-bridge-runtime Python dependency. You do not need to install a second
Python package by hand. If the Agent runs on another computer, enable the
EDA Bridge Runtime MCP/plugin
there; the ADS-only host does not need the Agent-facing plugin.
setup discovers installed ADS versions, asks for an explicit selection, and
installs the recoverable context add-on and two cooperating Skills.
quickstart passes only after documentation query, add-on registration,
disposable workspace creation, minimal circuit simulation, and dataset readback
have each passed.
If pipx or a suitable Python is not yet available, use the versioned
bootstrap for Linux
or Windows PowerShell.
The bootstrap creates an isolated environment and does not replace an
externally managed system Python.
Then open an exact workspace:
ads-agent --pretty launch --workspace /path/to/MyWorkspace_wrk
ads-agent --pretty status
ads-agent disconnectdisconnect leaves ADS running. shutdown requests native exit only for a
matching Agent-owned session.
| Natural-language request | What the Bridge checks |
|---|---|
| “Which ADS installations are here, and what can this one do?” | Discovers multiple versions, keeps the choice explicit, and probes real capabilities. |
| “Find the correct API for this installed ADS version.” | Searches a private, version-scoped local index and returns focused source evidence. |
| “Prove automation works before touching my project.” | Creates a disposable workspace, runs a minimal AC simulation, and reads the dataset through separate gates. |
| “Use the schematic, layout, cell, folder, or DDS item I selected.” | Resolves the copied ADS_CONTEXT instead of guessing the foreground window. |
| “Open this exact workspace and tell me what ADS is doing.” | Verifies workspace, process, ADS version, display, ownership, visible UI, and blocking dialogs. |
| “Add this part and wire in the design I am watching.” | Applies one typed, idempotent patch inside the exact active ADS design, uses a native ADS transaction, reads the created objects back immediately, and can roll back only that patch without saving. |
| “Apply these schematic edits safely.” | Modifies a non-overwriting copy and accepts it only after save, close, fresh reopen, and exact assertions. |
| “Simulate this circuit, give me the data, and build the DDS plots.” | Generates the netlist, runs the circuit simulator, checks numeric dataset columns, exports CSV, and fresh-reopens a native multi-page DDS report with rectangular or polar plots. |
| “Run this already-generated Momentum input.” | Preserves the source, solves a sibling copy, and checks a complete finite N-port result before promotion. |
| “Disconnect but keep ADS open.” | Separates client disconnect from identity-checked native shutdown. |
The capability matrix gives the maintained evidence and stop rule behind every row.
After ADS restarts, Copy ADS Context is available from supported DE schematic, layout, symbol, folder/library-tree, and DDS selections. The copied text identifies the software, host-local origin, workspace, object kind, selection, and freshness needed by the Agent. It contains no password, live port, or permission to mutate.
This is the normal user interaction:
- Select the intended object in ADS.
- Click Copy ADS Context.
- Paste it into the conversation and describe the task naturally.
- Review the target and evidence returned by the Agent.
See the context interaction contract for exact selection coverage.
The maintained acceptance path uses real ADS installations on Windows and Linux. It separately checks documentation, context capture, live session identity, safe dialog supervision, typed schematic construction, circuit simulation, dataset and CSV readback, native DDS equation and plot creation, generated-input Momentum execution, and reversible live object patches in the already-open design.
The separate maintained Bridge acceptance for the blank-workspace → schematic → simulation → native DDS path passed as one four-stage Runtime plan in 4.312 seconds, with 31 finite rows, a deterministic native dataset, CSV, and a freshly reopened two-page DDS report containing rectangular and polar plots. See the sanitized workflow evidence.
The corrected ADS 2027 comparison uses the current EDA Runtime MCP plus Runtime/ADS Skills—not the historical product CLI—against the official MCP, with Codex and Pi Agent. For the exact headless AC task, Runtime completed 4/6 and the official MCP 6/6; the official route was also faster. Successful Runtime native execution itself had a 2.379 s median, while the much larger Agent-time gap came from discovery, plan construction, and retries. Method and data
The knowledge result points the other way: Runtime completed 18/18 audited K1/K3/K6 runs, versus 12/18 for the official MCP. ADS 2027 does document Python DRC automation; the earlier negative benchmark oracle was wrong, and both surfaces now pass K6 in 12/12 combined runs. The remaining official misses were incomplete K1 route descriptions and non-current K3 geometry calls. Method and data
These are small, isolated regression suites, not a universal product ranking. They do not compare installation, the DE/DDS plug-in, GUI session control, dialog handling, or every ADS solver workflow; this is not a full-product comparison. The audited aggregate data is available as JSON. The v1/v2 files remain available as historical baselines but are superseded for current-product comparison.
On a remote ADS host, repeated Agent operations use:
ads-agent runtime serveEDA Bridge Runtime keeps one SSH stdio process alive, records the purpose and timing of each operation, and never exposes the embedded ADS Bridge port. If Agent and ADS share a machine, register the same service as a local connection; do not bypass Runtime. The engineering behavior and evidence remain the same in both topologies.
- Local documentation, indexes, workspaces, tokens, and results stay on the EDA host unless the user deliberately exports them.
- Live endpoints bind to loopback and require a random session token.
- A reused session must match the selected ADS installation and exact workspace.
- A Context identifies the target but never authorizes editing or simulation.
- Supervised live edits are read back immediately and are never saved or discarded implicitly.
- Structured edits protect the source and require fresh-reopen assertions.
- Dialog actions require a fresh process/window fingerprint, not screen coordinates.
- Shutdown refuses unverified or user-owned sessions and never force-kills ADS.
- Arbitrary embedded Python and dynamic AEL remain disabled by default.
| ADS generation | Support level |
|---|---|
| ADS 2025 and later | Stable target, decided by runtime capability probes |
| ADS 2024 Update 2 | Preview |
| ADS 2023 Update 2 through ADS 2024 Update 1 | Experimental |
| Older installations | Documentation-only when local docs can be discovered |
- governed access to version-matched official Python/AEL so new documented ADS uses do not require one Bridge wrapper per component, plot, or solver option;
- richer RF testbenches, reusable parameterized cells, native DDS, layout, and EM work delivered through that common path and selectively promoted as certified workflows.
- Installation and command reference
- Five public examples
- Capability and evidence matrix
- Architecture and capability growth
- Operation classification
- Session and dialog behavior
- Execution context contract
- Release contract
- Changelog
To remove only this ADS integration while preserving unrelated add-ons:
ads-agent addon uninstall