io.github.YawLabs/mcp

Load one or more installed servers' tools into this session by namespace; tools are prefixed by it (e.g. "gh_create_issue"). Pass `server` for one or `servers` for several. `tools: [...]` (single server only) advertises just those tools; the rest stay callable by name. Refused when YAW_MCP_MIN_COMPLIANCE is set and the server's grade is below it (ungraded always passes); the message names the grade and the variable.

List curated multi-server 'bundles' (`pr-review` = github + linear, `devops-incident` = github + pagerduty + slack, ...) for a known multi-server WORKFLOW, each with a ready-to-run `mcp_connect_activate({ servers: [...] })` call. `action="match"` partitions them against the installed servers into READY (activate now) and PARTIAL (names the missing namespaces; `yaw-mcp add <slug>` installs one, catalog at https://yaw.sh/mcp/catalog/). `action="list"` (default) returns the whole catalog. Static data, no network call.

Unload one or more loaded servers' tools from this session to free context. The server stays installed and can be reloaded with `mcp_connect_activate`. yaw-mcp also auto-unloads a server after a run of calls to other servers (baseline YAW_MCP_IDLE_THRESHOLD, raised for a server used in bursts).

List the MCP servers installed in the user's ~/.yaw-mcp/bundles.json: name, namespace, tool count, an estimated token cost (a tilde marks an estimate from cached metadata), whether it is loaded, any local CLI it shadows, an inline `[A]`-`[F]` compliance grade when scored (ungraded is unmarked, not penalized), and usage hints. Tool-name lists are truncated; pass `server` for one server's full card. Recurring packs get a ready-to-run `mcp_connect_activate` call at the top.

Pick the best-matching installed server(s) for a natural-language task and load their tools in one call. Describe the task ("create a github issue for the login bug", "query the prod postgres"); yaw-mcp ranks the installed servers with BM25 and loads the top match. `budget` defaults to 1; raise it only when one task genuinely spans several servers.

Run a short DECLARATIVE pipeline of upstream tool calls in one round-trip, when you already know the 2-4 calls and one step's output feeds another's args -- e.g. `a = gh_list_prs(); b = gh_get_pr(a[0].number); return b`. No expressions, loops, branches or arithmetic. Steps run in order; the only data flow is `{"$ref": "<stepId>[.path.to.value]"}` (dot keys, `[N]`/`.N` indexing), substituting a prior step's output into a later step's args. A `$ref` reads the output as the server sent it, before the pruning and size cap applied to the echoed copy. Each `tool` is a namespaced upstream name: a loaded server is called directly; an unloaded server whose tools are cached is loaded on first use (its tools join this session). A compliance-floor refusal is decided before step 0 runs and refuses the whole pipeline with nothing done, so a re-run repeats no side effect; a server-cap refusal is only knowable when the step is reached and fails it there, with `partial` holding what already ran. A name neither loaded nor cached, or on a disabled/profile-blocked server, fails its step. Max 16 steps per exec. Refused before step 0: plain `exec: ...` text; a later failure: `{ ok: false, failedStep, error, partial }`. Success returns `{ ok: true, result, steps }`; with a named `return`, `result` is that step's output, `stepKeys` is added, and `steps` is dropped once the intermediates exceed about 4 KB. Name a `return` when one value is enough: it keeps a large intermediate out of your context.

Search every installed server's tools by what they DO ("resize an image", "list pull requests") when you do not know which server has it. Ranks tool names and descriptions across loaded and unloaded servers from cache; nothing is loaded and no server is contacted. On a fresh install a no-match first waits (up to 20 s) for the startup scan still learning a server's tools. A match on a loaded server carries its full input schema; one on a never-loaded server carries name and description only -- `mcp_connect_read_tool` returns its arguments.

Health stats for the servers loaded this session: total calls, error count, average latency, last error. Installed-but-unloaded servers with a poor persisted success rate (<80% across sessions) are listed in a separate block -- do NOT load a server just to see its history, loading resets its in-session counters.

Return one tool's full input schema without loading its server. Free for a loaded server; for an unloaded one yaw-mcp opens a transient connection, reads the schema and closes it -- nothing is added to the session. To call the tool, load its server with `mcp_connect_activate` or call it by name from `mcp_connect_exec`.

Report, per installed server, which local-vault secrets its `${secret:NAME}` references resolve to -- NAMES only, never a value. `injectedSecrets`: names the vault has. `missing`: names it lacks (`yaw-mcp secrets set <name>`). `malformed`: references that do not parse (a space in the name, a missing `}`), quoted in bounded form behind a `<malformed ref>` marker. yaw-mcp refuses to start a server over a missing or malformed reference. Servers with no references are omitted. Decrypts nothing; needs no passphrase.

Surface recurring multi-server 'packs' -- servers used together in short bursts more than once -- each as a ready-to-run `mcp_connect_activate` call with all its namespaces. Observation-only: loads and unloads nothing. Patterns persist across yaw-mcp restarts. Returns a 'no patterns yet' message when nothing has recurred.