MCP
The whole platform as MCP tools, from the same contracts as the API, the SDKs and the CLI.
The Sylphx MCP server is the platform inside an agent: the same methods as the API reference, the CLI, the SDKs and the Terraform provider, served as MCP tools over stdio. Agents and coding assistants connect a client; this page is what it answers with. To use the hosted server instead, see Connect an MCP client.
Run
npx @sylphx/mcp
sylphx mcp
cargo install sylphx-mcp && sylphx-mcpThe three run the same server — Node, the operator CLI, or the binary. A client starts it over stdio, so a run line is for trying it by hand: the process reads JSON-RPC on standard input and writes it on standard output.
Configure a client
{ "mcpServers": { "sylphx": { "command": "npx", "args": ["-y", "@sylphx/mcp"], "env": { "SYLPHX_API_KEY": "sylphx_sk_…" } } } }The server reads its key from SYLPHX_API_KEY. Without one, search and
describe still answer, and a call explains how to sign in.
Core tools
Core methods come with a tool of their own — {service}_{collection}_{verb},
or {service}_{verb} for a call a service serves itself — each carrying
the request's input schema. The tool's name links to the method it serves:
| Tool | Effect | What it does |
|---|---|---|
access_whoami | read | Returns the caller: the principal and the scope of the key or token that authenticated this request. Clients fill their org, project, and env defaults from it; every valid key may call it, whatever its scopes. |
access_projects_create | write | Creates a project. Access assigns the id. |
agents_agents_create | write | Creates an agent and its version 1. |
agents_agents_run | write | Atomically admits a pinned task session, initial event and pending turn. |
agents_agent_sessions_create | write | Starts a session of an agent, pinned to its current version unless agent_version names another. Requires a secret key until end-user ownership is implemented. The session starts idle; :send starts its first turn. |
agents_agent_sessions_send | write | Sends a user message. On an idle or failed session it starts a new turn; while a turn runs it is refused with INVALID_STATE (use :steer). The same Idempotency-Key returns the same event. |
agents_session_events_list | read | Lists a session's events in sequence order. To follow a session, call again with the filter sequence > N, N the last sequence read; wait holds the call open until a newer event is appended or the wait ends. |
build_builds_get | read | Gets a build. |
build_run | write | Runs one command over an uploaded tree and settles with its exit code, like the local command. A caller that gets a failure before the command started (BUILD_FAILURE_SETUP_FAILED, NO_CAPACITY) may build locally. |
build_image | write | Builds an image, with an SBOM and signed provenance, from a commit or an uploaded tree. |
build_logs_read | read | Reads a page of a build's log lines, oldest first. With wait, an empty page on a running build waits up to that long for new lines, so a client follows the log by passing each next_page_token back. |
build_cache_env | write | Mints the settings a build that runs elsewhere (a laptop, a CI runner) uses to reach the cache: the sccache, Turbo and crates mirror doors and a short-lived token, as environment variables. |
build_caches_purge | destructive | Empties a build cache and keeps the cache. The undo is that the next build refills it. |
data_databases_create | write | Creates a database. The Operation is done when the controller settles it. |
knowledge_remember | write | Records an episode and returns at once (remember); extraction reads it afterwards. |
knowledge_recall | read | Searches the graph (recall): hybrid lexical and semantic search over episodes and facts, fused by reciprocal rank, then at most two hops of expansion from the top entities. Returns a context pack with citations, built only from records the caller may see. |
observability_log_entries_query | read | Queries the environment's log entries over a bounded interval; follow a trace with trace_id = "...". |
observability_traces_query | read | Queries the environment's traces over a bounded interval, returning summaries without spans. |
observability_generations_get | read | Gets a generation: one Sylphx AI call. |
observability_generations_list | read | Lists the environment's Sylphx AI calls, newest first, within the trace retention; filter by request_model, outcome, api_key, start_time, tags.<key> and the other filterable fields. |
workflows_schedule | write | Updates a schedule; with allow_missing it creates the schedule when it does not exist, so its short name schedule is an upsert by name. |
workflows_start | write | Starts a Run of a Workflow, now or after start_delay or at schedule_time. run_id is the caller's key: the same key and intent return the same Run, a changed intent is refused. An end user's session may call it on a Workflow whose callers is USERS when the session holds the Workflow's permission. |
workflows_signal | write | Sends a named signal to a running Run; a wait op for that name resumes with its payload. A signal to a closed Run fails with INVALID_STATE. An end user's session may call it on a Workflow whose callers is USERS when the session holds the Workflow's permission. |
workflows_cancel | destructive | Requests cancellation. Timers and waits stop; a running invocation is not interrupted: the handler receives one final invocation with cancel: true, and the Run ends cancelled when it answers. An end user's session may call it on a Workflow whose callers is USERS when the session holds the Workflow's permission. |
workflows_result | read | Waits up to timeout (at most 60s) for the Run to close, then returns it, closed or not. Its short name is result: the closed Run carries the handler's output or error. An end user's session may call it on a Workflow whose callers is USERS when the session holds the Workflow's permission. |
Meta tools
Three tools reach the rest of the platform, whatever its service:
| Tool | What it does |
|---|---|
sylphx_search_methods | Search every Sylphx API method by service, collection, verb, or description. |
sylphx_describe_method | Return the input JSON Schema, permission, and effect of one method. |
sylphx_call | Call any Sylphx method by id with arguments matching its input schema. Destructive methods require confirm: true. |
The catalog
sylphx_search_methods searches the method catalog — every method this
reference documents except the ones the protocol reserves — and each hit
carries the id, the tool name, the effect, the scope, the HTTP call and the
input schema. sylphx_describe_method returns one entry, and sylphx_call
runs one: its arguments are the method's request, so a call an agent makes is
the call its page documents, and a destructive one needs
confirm: true in the arguments.
The server declares no resources and no prompts: resources/list and
prompts/list answer with empty lists, because everything the platform holds
is reached by calling a method.