vivado-mcp
Automates AMD Vivado FPGA design suite, providing remote session control, project management, synthesis/implementation/bitstream builds, and report analysis over SSH.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@vivado-mcprun synthesis"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
vivado-mcp
vivado-mcp is a remote automation server for AMD Vivado built on the Model Context Protocol (MCP). It allows Claude Code and other MCP clients to use Vivado on a Linux build server over SSH, without requiring the Xilinx/AMD toolchain on the client machine.
The current release provides full Vivado session control, Git workspace synchronization, project management, durable build jobs, build progress and log inspection, and structured analysis of utilization, timing, DRC, CDC, power, and related reports.
Key capabilities
Connects to the build server over SSH without opening additional TCP ports.
Automatically deploys a remote Python agent backed by a persistent Unix socket service.
Maintains long-lived Vivado Tcl sessions so open projects and checkpoints can be reused across commands.
Runs synthesis, implementation, and bitstream builds as independent jobs. Job state survives SSH disconnections and daemon restarts.
Provides cancellable jobs, long polling, and incremental event cursors.
Synchronizes dirty Git worktrees through a temporary index without changing the user's branch, HEAD, or staging area.
Enforces workspace boundaries for every file path and exposes paths as workspace-relative POSIX paths.
Normalizes Vivado reports instead of sending large text reports directly into the model context.
Caches report results by DCP content hash and compares utilization or timing across builds.
Deduplicates Vivado messages by message ID to condense repeated warnings and errors.
Enforces a single-writer lock for each project, preventing multiple Vivado processes from modifying the same
.xprfile concurrently.
Related MCP server: vivado-mcp-agent
Requirements
Client
Python 3.12 or later
uvOpenSSH client
Git
rsync
An MCP client with stdio transport support, such as Claude Code
Build server
Linux
Python 3.12 or later
Git
rsync
Non-interactive SSH access
An installed Vivado release with an accessible
settings64.sh
The remote agent uses only the Python standard library. It does not require the project's Python dependencies or root privileges on the build server.
Installation
Clone the repository on the client and install its dependencies:
uv sync --lockedCreate ~/.config/vivado-mcp/config.toml. To use a different configuration
file, set the VMCP_CONFIG environment variable:
[host]
name = "fpga-builder"
ssh = "fpga-builder"
agent_dir = "~/.vivado-mcp"
max_concurrent_jobs = 1
default_jobs = 16
nice = 10
stall_timeout_s = 900
[[host.tools]]
label = "vivado-2023.2"
kind = "vivado"
settings_sh = "/tools/Xilinx/Vivado/2023.2/settings64.sh"
default = true
[[workspace]]
name = "example-project"
local = "/absolute/path/to/example-project"
host = "fpga-builder"
build = "build"
default = true
[security]
allow_eval = truehost.ssh may be a user@hostname target or a host alias from
~/.ssh/config. Public-key authentication is recommended. Verify that the
following command runs without prompting for input:
ssh fpga-builder python3 --versionMultiple [[host.tools]] entries may be configured to select among Vivado
versions. The tool argument accepted by MCP tools refers to the corresponding
label. When omitted, vivado-mcp uses the entry marked default = true, or the
first entry if no default is declared.
Multiple [[workspace]] entries are also supported. MCP calls may omit the
workspace argument when exactly one workspace is configured or one entry is
marked default = true.
Registering the MCP server
Register the server with Claude Code:
claude mcp add vivado -- uv run --directory /absolute/path/to/vivado-mcp vmcp-mcpOther MCP clients can use the same stdio command:
command: uv
args: run --directory /absolute/path/to/vivado-mcp vmcp-mcpOn the first tool call, the client builds and deploys the remote agent. Later calls verify the protocol version and payload hash. If the running agent needs an upgrade, the server reports that state explicitly instead of closing active Vivado sessions without authorization.
The remote service uses namespaced host resources under agent_dir:
vmcp-agentd.sock, vmcp-agentd.lock, and vmcp-agentd.log. This avoids
collisions with unrelated services that use generic agentd names.
Workspace and path model
workspace.local must point to the root of a local Git repository. For each
configured workspace, vivado-mcp maintains three separate areas on the build
server:
a bare Git repository that receives synchronization commits;
a source worktree checked out at the exact selected commit;
a persistent build directory for non-Git artifacts such as
.xprfiles, run directories, DCPs, bitstreams, and reports.
Project, source, DCP, and ordinary file paths passed to MCP tools are POSIX-style paths relative to the workspace. For example:
rtl/top.sv
constraints/top.xdc
build/demo/demo.xpr
build/flows/j_12345678/post_synth.dcpAbsolute paths, .., backslashes, and symlink escapes are rejected. Paths that
begin with the configured build prefix map to the persistent build directory;
all other paths map to the remote source worktree.
File synchronization deliberately has one transport in each direction:
sync_pushis the only source-upload path. It captures modified, staged, and untracked files through a temporary Git index and a synthetic commit, without altering the user's HEAD, branch, or real index. rsync push, SSHFS, and shared filesystem source backends are not supported.Build artifacts live outside the Git worktree and never enter synchronization commits.
sync_pulluses rsync over SSH to copy explicitly requested remote artifacts, including the complete configured build directory when requested. It never performs deletions.
The local checkout remains the source of truth. Direct remote source edits are
not part of the normal workflow; edit locally and run sync_push again.
Recommended workflows
Project mode
host_status
→ sync_status
→ sync_push
→ project_create or project_open
→ sources_add / set_top / set_generics / set_property
→ project_close
→ session_close
→ build
→ job_wait
→ report_utilization / report_timing_summary / report_drcproject_close closes the project and releases its single-writer lock.
session_close terminates the Vivado process and releases its memory. After
editing a project, call both in that order.
Non-project mode
sync_push
→ flow_run
→ job_wait
→ job_artifacts
→ report_timing_summaryflow_run reads HDL and XDC files directly. It is well suited to small
experiments, quick synthesis runs, and automation that does not need a
long-lived .xpr project.
Iterative optimization
report_utilization(old_job)
→ report_timing_summary(old_job)
→ modify the design and rebuild
→ report_diff(old_job, new_job, kind="utilization")
→ report_diff(old_job, new_job, kind="timing")MCP tool reference
The current release exposes 47 MCP tools. In the signatures below, workspace
and tool may be omitted when the configuration provides an unambiguous
default.
Host and agent
Tool | Parameters | Description |
| None | Returns the agent version and uptime, host CPU/memory/disk information, configured tool probes, active sessions, active jobs, and job-slot usage. This is the recommended first call. |
|
| Deploys or upgrades the remote agent. |
Vivado sessions and Tcl
Tool | Parameters | Description |
|
| Starts a long-lived tool interpreter and returns a |
| None | Lists live sessions, including their |
|
| Closes a session and releases its memory. Interpreter state that has not been written to disk is lost. |
|
| Executes Tcl in an existing session and returns |
|
| Runs |
tcl_eval is an intentionally powerful interface with the same Tcl privileges
as the remote user. If arbitrary Tcl execution is unnecessary, set
security.allow_eval = false.
Workspace synchronization and file access
Tool | Parameters | Description |
|
| Returns the local Git HEAD and dirty-file list together with the exact commit currently active on the build server. |
|
| Synchronizes the local worktree to the build server. Omitting |
|
| Pulls selected remote artifacts or directories into the local workspace using rsync over SSH. Passing the configured build directory pulls the complete build tree. Paths must be workspace-relative, and |
|
| Reads a bounded UTF-8 segment from a remote workspace file. Returns |
|
| Searches the remote source and build trees with a regular expression and returns relative paths, line numbers, and matching text. Results are capped at 1,000 entries, and the number of scanned files is bounded. |
Vivado projects, sources, and IP
Tool | Parameters | Description |
|
| Creates and opens an |
|
| Opens an existing |
|
| Closes the current project and releases its single-writer lock without terminating the Vivado session. |
|
| Returns the project path, part, board, top, target language, source inventory, run status, IP inventory, and residual lock information. File paths are converted to workspace-relative paths. |
|
| Adds HDL, constraint, or data files to the current project. Vivado |
|
| Removes file references from the project without deleting the files from disk. |
|
| Lists files registered with the current project, including their type and library. |
|
| Sets the top-level module or entity for the current source fileset. |
|
| Sets generic or parameter overrides on the current fileset. |
|
| Appends workspace-relative directories to the Verilog/SystemVerilog include path. |
|
| Sets the Vivado strategy for a named synthesis or implementation run. |
|
| Sets one Vivado property on a typed object. |
|
| Creates a Vivado IP instance and optionally sets |
|
| Upgrades selected IP instances to the current catalog version. |
|
| Runs |
|
| Starts a durable non-project job. |
|
| Removes stale Vivado |
The complete Block Design Tcl interface remains available through tcl_eval.
Potentially long-running commands should still be submitted as jobs rather than
blocking an interactive interpreter.
Build jobs and observability
Tool | Parameters | Description |
|
| Starts an independent durable job for an existing |
|
| Lists durable jobs in reverse creation order. |
|
| Returns job state, the current step and phase, a monotonic progress estimate, elapsed time, message counts, the log tail, and the latest event sequence number. |
|
| Reads events after |
|
| Long-polls a job until it reaches a terminal state, emits an important event, or times out. Each call is limited to 3,600 seconds and uses MCP progress notifications to remain visible to the client. |
|
| Cancels a queued or running job and sends a termination signal to its dedicated process group. Calling it on a terminal job is safe. |
|
| Returns a bounded job log tail, optionally filtered by a regular expression. It never returns an entire multi-megabyte Vivado log. |
|
| Aggregates log messages by Vivado message ID and returns the severity, occurrence count, and one example. |
|
| Lists DCP, bitstream, report, XSA, LTX, and related job artifacts. Workspace-backed jobs return workspace-relative paths. |
Job states follow this lifecycle:
queued → starting → running → succeeded
↘ failed
↘ cancelled
↘ timeout
↘ oom
↘ lostlost means the daemon cannot determine the process's final outcome; it is not
reported as an ordinary failure. Use an idempotency_key when retrying a job
submission whose response may have been lost to a network interruption. A
repeated key returns the existing job instead of launching another expensive
build.
Reports and analysis
The source accepted by report tools may be a job_id or a workspace-relative
DCP path. Jobs emit a standard report bundle when they finish. When only a DCP
is available, the agent reuses a read-only Vivado session to generate missing
reports on demand. Normalized results are cached by DCP SHA-256, report type,
and arguments.
Tool | Parameters | Description |
|
| Returns used, available, and percentage values for LUT, LUTRAM, FF, BRAM, URAM, DSP, IO, and BUFG resources. Zero-use resources are omitted by default. Hierarchical output and cell filters are supported. |
|
| Returns WNS, TNS, WHS, THS, pulse-width metrics, failing endpoint counts, per-clock-domain statistics, clock frequencies, and compact worst-path summaries. |
|
| Queries targeted timing paths. Endpoint arguments use Vivado object-matching patterns. |
|
| Returns clock periods, target frequencies, WNS-derived achieved frequencies, and timing metrics for each clock domain. |
|
| Aggregates DRC violations by rule and returns the rule, severity, count, description, and one example. |
|
| Aggregates Vivado methodology findings by rule without returning the full report. |
|
| Aggregates clock-domain crossing findings by rule and severity. |
|
| Returns total on-chip, dynamic, and static power, junction temperature, confidence, and component-level power data. |
|
| Compares normalized reports from two jobs or DCPs. |
Reliability and runtime semantics
vivado-mcp deliberately separates interactive sessions from long-running builds:
Sessions handle fast Tcl queries and project editing. They survive SSH reconnections, but a daemon restart closes them.
Jobs run in independent process groups with state and events persisted to disk. They survive SSH disconnections and daemon restarts.
Job state, events, and JSON indexes are written atomically. Process identity is verified with both the PID and Linux process start time to avoid mistakes caused by PID reuse.
Build artifacts and reports remain in the persistent remote build directory; source synchronization does not clear that directory.
Large logs and reports are always returned in bounded, paginated, or structured form to protect the MCP context window.
Security considerations
Network access inherits the authentication and authorization policy of SSH; the agent does not listen on a TCP port.
Workspace tools reject absolute paths, parent-directory traversal, backslashes, and out-of-bounds symlinks.
Only one active write session may own a project at any time.
sync_pullnever deletes remote or local files.tcl_evalis equivalent to executing Vivado Tcl as the SSH user. When MCP callers are not fully trusted, setsecurity.allow_eval = falseand apply the principle of least privilege to the remote account.agent_ensure(force=true)closes active sessions. Checkhost_statusorsession_listfirst to understand the impact. Running durable jobs are not terminated.
Supported scope
The public interface currently covers project and non-project Vivado flows, including synthesis, implementation, bitstream generation, IP management, and static report analysis. Vitis Unified, Vitis HLS, and hardware programming are not yet exposed through MCP tools.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- FlicenseCqualityCmaintenanceAn AI-centric MCP server that enables automated Xilinx Vivado workflows, including project management, synthesis, implementation, and timing analysis. It allows AI agents to drive hardware design processes while integrating directly with the official Vivado GUI for visual context.Last updated154
- Alicense-qualityCmaintenanceMCP server for deterministic Vivado FPGA automation including project scaffolding, simulation, synthesis, implementation, and gated bitstream generation.Last updated4MIT
- Alicense-qualityAmaintenanceA minimal MCP server that provides 25 tools and 5 hooks to control Xilinx Vivado EDA for FPGA development, including session management, Tcl execution, smart diagnostics, and IP debugging.Last updated85Apache 2.0
- AlicenseBqualityDmaintenanceA Model Context Protocol (MCP) server that enables AI assistants like Claude to directly interact with AMD/Xilinx Vivado FPGA development tools.Last updated4053MIT
Related MCP Connectors
Remote MCP server for XDaLa workflow preparation on XGR.Network.
Remote MCP for Android CLI agent build gate, structured receipts, audit logs, and reviewer-ready evi
A paid remote MCP for ClawManager, built to return verdicts, receipts, usage logs, and audit-ready J
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/kkkaishao/vivado-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server