"filesystem" matching MCP tools:

• query_docs_filesystem_honeydew_documentation

  Honeydew AI Documentation

  Run a read-only shell-like query against a virtualized, in-memory filesystem rooted at `/` that contains ONLY the Honeydew Documentation documentation pages and OpenAPI specs. This is NOT a shell on any real machine — nothing runs on the user's computer, the server host, or any network. The filesystem is a sandbox backed by documentation chunks. This is how you read documentation pages: there is no separate "get page" tool. To read a page, pass its `.mdx` path (e.g. `/quickstart.mdx`, `/api-reference/create-customer.mdx`) to `head` or `cat`. To search the docs with exact keyword or regex matches, use `rg`. To understand the docs structure, use `tree` or `ls`. **Workflow:** Start with the search tool for broad or conceptual queries like "how to authenticate" or "rate limiting". Use this tool when you need exact keyword/regex matching, structural exploration, or to read the full content of a specific page by path. Supported commands: rg (ripgrep), grep, find, tree, ls, cat, head, tail, stat, wc, sort, uniq, cut, sed, awk, jq, plus basic text utilities. No writes, no network, no process control. Run `--help` on any command for usage. Each call is STATELESS: the working directory always resets to `/` and no shell variables, aliases, or history carry over between calls. If you need to operate in a subdirectory, chain commands in one call with `&&` or pass absolute paths (e.g., `cd /api-reference && ls` or `ls /api-reference`). Do NOT assume that `cd` in one call affects the next call. Examples: - `tree / -L 2` — see the top-level directory layout - `rg -il "rate limit" /` — find all files mentioning "rate limit" - `rg -C 3 "apiKey" /api-reference/` — show matches with 3 lines of context around each hit - `head -80 /quickstart.mdx` — read the top 80 lines of a specific page - `head -80 /quickstart.mdx /installation.mdx /guides/first-deploy.mdx` — read multiple pages in one call - `cat /api-reference/create-customer.mdx` — read a full page when you need everything - `cat /openapi/spec.json | jq '.paths | keys'` — list OpenAPI endpoints Output is truncated to 30KB per call. Prefer targeted `rg -C` or `head -N` over broad `cat` on large files. To read only the relevant sections of a large file, use `rg -C 3 "pattern" /path/file.mdx`. Batch multiple file reads into a single `head` or `cat` call whenever possible. When referencing pages in your response to the user, convert filesystem paths to URL paths by removing the `.mdx` extension. For example, `/quickstart.mdx` becomes `/quickstart` and `/api-reference/overview.mdx` becomes `/api-reference/overview`.

  Connector

• create_tests

  3TG Test Generation

  Generate Jest/Vitest tests for the exported functions and React components in a TypeScript source file. Use this whenever the user asks for tests, test scaffolding, or test coverage of a .ts or .tsx file. Returns the generated test (and any companion .3tg.md / __mocks__) file contents, with paths already translated to the user's `.3tg/` mirror convention. Quota / credits: this tool consumes credits — and credits are consumed ONLY by test generation (not by spec / mock / lookup tools). The accounting is exactly **1 credit per generated test case** (i.e. per `test(...)` / `it(...)` block 3TG emits inside the returned `.test.ts` / `.test.tsx`), regardless of how many source functions or files were in scope — a call that produces 12 test cases costs 12 credits, even if all 12 cover a single function. Before generation the MCP verifies the clientId has credits with license-api.coding-creed.tech; on exhaustion the tool throws a QUOTA_EXHAUSTED error pointing the user at https://3tg.dev. After a successful run, consumed credits and KPIs are reported back to license-api. Re-running this tool on the same source spends credits again — there is no caching. When the previous call returned `enrichment.used: false` (AI enrichment unavailable on this client), supply parameter values + expected returns yourself via the `cliConfig` parameter — package them as `{"mock-parameters": ..., "function-returns": ...}` (same shape AI enrichment would produce) and pass them on a retry call. **Do NOT autonomously write `.3tg/config.3tg.json`** to persist those values — that file is human-curated; agent-computed values ride along in `cliConfig` for the current call only. (Explicit user requests to edit the file are fine — handle those normally.) See the cliConfig parameter description below for the full pattern. CRITICAL POST-CALL ACTION — write returned files to disk: The MCP server does NOT touch the user's filesystem. It returns the generated file CONTENTS in the response's `files` array. After this tool returns, you MUST iterate over `files` and write each entry's `content` verbatim to its `path` using your native file-write capability (e.g. Write / edit_file / create_file — whatever your client exposes). Create parent directories as needed. Returned paths are project-root-relative and already translated to the `.3tg/` mirror convention where applicable (e.g. specs land under `.3tg/<source-path>.3tg.md`; tests / mocks travel through unchanged). Write each path verbatim. Do NOT claim "Generated test file: <path>" unless you have actually written the file. The user will assume the MCP wrote it and waste time looking for a non-existent file. If you can't write for some reason (permission denied, no write capability in this client), return the contents inline in your message so the user can copy-paste them. Never report success silently when the write didn't happen.

  Connector

• generate_illustration

  Logospell

  Generate a single composed illustration from a prompt - a scene, environment, hero image, banner, character portrait, or any standalone picture. (For a SET of separate isolated subjects sharing one style - icons, sprites, asset packs - use generate_image_set for solid-color backgrounds, or generate_transparent_image_set for transparent backgrounds, instead.) Returns a zip download URL: the image plus a prompt.json recording what you asked for, plus an index.html viewer. Each call costs 1 credit. Run generation calls sequentially, never in parallel - only one generation runs at a time per API key. The prompt describes the picture and is used as written (no template, no appended instructions); do not put the pixel size in it - put the size in the width and height arguments. width and height must be one of the supported pixel pairs below; any other pair is rejected. These are the maximums offered - nothing larger is available. Each line is one shape: the first pair is that shape's largest size, the rest are exact proportional downscales of it. 2752x1536, 1376x768, 688x384 2048x2048, 1024x1024, 512x512, 256x256 3168x1344, 1584x672, 792x336 1696x2528, 848x1264, 424x632 2528x1696, 1264x848, 632x424 1792x2400, 896x1200, 448x600 2400x1792, 1200x896, 600x448 1856x2304, 928x1152, 464x576 2304x1856, 1152x928, 576x464 1536x2752, 768x1376, 384x688 Formats: png (default, lossless), jpg, webp. quality (1-100) applies to jpg and webp; default 90. Prompt max length: 2500 characters. A prompt that does not describe a picture is rejected. unpackTo: a directory on your local filesystem to extract the downloaded zip into. filename: name for the image file inside the zip (default illustration.<ext>). If a "Rate limit exceeded" error is returned, wait the suggested number of seconds before retrying. Do not retry immediately.

  Connector

• execute_python

  Python Code Validator

  Everything repair does, and then RUNS the code in a throwaway container — no network, read-only filesystem, killed at options.timeout_s — reporting exit code, stdout and stderr. Any '>>>' examples in the code are run too, and one that does not print what it says is an error the other tools cannot see. This is a side effect: do not submit code you do not want executed. Use it when you need proof that the code runs, or that it does what it says. Alternatives: validate_python for the diagnosis and repair_python for the fix, neither of which runs anything. Auth: a key is required. This call needs a paid key and answers HTTP 402 without one. Credits are bought without an account, 10 per call: GET /v1/pricing says where to send the xDAI. Arguments: code: the whole file, 1..200000 bytes of UTF-8 measured after encoding (empty is refused with 400, larger with 413); a fragment is fine, but line and column numbers in the answer count from 1 in what you sent. language: must be 'python'; anything else is 400, and the field may be omitted. options.max_iterations (1..10, default 3) caps the fix/verify rounds: raise it for a file with several independent faults, leave it for a snippet. options.optimize (default false) additionally folds constants and drops dead code, and is only worth setting when you asked for a rewrite anyway. options.transpile_to (e.g. 'javascript') returns a translation of the *repaired* source in transpiled, not of what you sent. fixed_code is null when nothing could be proven safe to change, so treat null as 'no fix', not as an error. options.timeout_s (seconds, default 5) is the wall clock for the run; the schema allows up to 60 but this deployment caps it at 30 and refuses a larger value with 400. options.expected_output compares stdout byte for byte and adds an 'expected-output' diagnostic (valid=false) when it differs, which is how you ask for 'it did the right thing' rather than 'it ran'. options.examples is the same question for code with no output: pass what you asked for as doctest lines ('>>> total([1, 2])' then '3') or assertions ('assert total([1, 2]) == 3'), and each is run against the code -- one that does not hold is a 'python:example-mismatch' error, and repair looks for a single-token change that makes them all pass. Send it whenever you know what you asked for: without it, code that runs but returns the wrong answer looks perfect from here. The program that runs is the repaired one, so read fixed_code before you trust runtime.stdout, and it runs exactly once however many rounds the repair took. Returns valid, score 0..1, diagnostics (rule, message, line, column), security findings, fixes, fixed_code and runtime; see outputSchema. The code and its verdict are retained to improve the service.

  Connector

• compose_landing

  Layero

  Drive the 3-step quiz flow and produce a complete landing. This is the **main entry point**. Don't ask the user questions in chat before calling this — the tool opens native quiz dialogs in the IDE itself. Call this immediately when the user describes what they want. The three quizzes are: 1. **Motivation** — what's being built, for whom, the desired action. 2. **Look & feel** — palette, tone, optional references. 3. **Final picks** — design system (top 3 matched), where submissions go, project name slug. Returns a ``ComposeResult`` with the file bundle to write to disk. The agent then writes the files using the IDE's filesystem tool and proceeds to integration setup / deploy.

  Connector

• fvs_download_final_video

  Future Video Studio

  Download a completed Future Video Studio final render URL to a local file. Use this only after fvs_get_render_status or fvs_get_paid_render_status returns a final_video_url for a completed render. The tool performs an unauthenticated HTTPS GET to that signed URL and writes the response bytes to output_path on the MCP server's local filesystem. It does not call the FVS Agent API, spend wallet credits, require FVS_AGENT_API_KEY, cancel jobs, or modify remote render state. Side effects and constraints: output_path is a local filesystem path for the MCP server process, parent directories are created, existing files are not replaced unless overwrite is true, and large videos may take minutes to download. The request timeout is 600 seconds. Use a fresh status check to refresh expired signed URLs, and do not pass arbitrary or untrusted URLs.

  Connector

• check_text

  WSC - Writing Style Checker

  Analyze text for writing style issues: weasel words, passive voice, duplicate words, long sentences, nominalizations, hedging, filler adverbs, and research-cited AI tells. Read-only and stateless — text is analyzed in memory on the hosted server and never stored. Returns a plain-text report with each issue's line and column, the matched text, surrounding context, and the reason for AI tells; texts over 100,000 characters return an error message. This hosted server has no filesystem access — the wsc-mcp npm package adds a check_file tool for local files. It only reports issues — to auto-remove duplicate words, follow up with fix_duplicates.

  Connector

• restore_backup

  mcp

  Restores a backup into an existing volume (volumeId), an instance's volume (instanceId), or a new one (volumeName). Poll list_volumes for the restored volume. DISASTER RECOVERY (the original VM is gone, so its volume died with it): restore with volumeName to rebuild the data as a NEW volume. To bring the MACHINE back, pass that volume as create_instance's boot_volume_id - it boots from the restored disk, so the VM returns with its filesystem intact and nothing to mount. Use attach_volume instead only when you want the data as an EXTRA disk on an existing machine (then get_ssh_command -> lsblk -> mount). A restored volume is inert until it is booted from or attached.

  Connector

• query_docs_filesystem_n_top

  docs

  Run a read-only shell-like query against a virtualized, in-memory filesystem rooted at `/` that contains ONLY the nTop documentation pages and OpenAPI specs. This is NOT a shell on any real machine — nothing runs on the user's computer, the server host, or any network. The filesystem is a sandbox backed by documentation chunks. This is how you read documentation pages: there is no separate "get page" tool. To read a page, pass its `.mdx` path (e.g. `/quickstart.mdx`, `/api-reference/create-customer.mdx`) to `head` or `cat`. To search the docs with exact keyword or regex matches, use `rg`. To understand the docs structure, use `tree` or `ls`. **Workflow:** Start with the search tool for broad or conceptual queries like "how to authenticate" or "rate limiting". Use this tool when you need exact keyword/regex matching, structural exploration, or to read the full content of a specific page by path. Supported commands: rg (ripgrep), grep, find, tree, ls, cat, head, tail, stat, wc, sort, uniq, cut, sed, awk, jq, plus basic text utilities. No writes, no network, no process control. Run `--help` on any command for usage. Each call is STATELESS: the working directory always resets to `/` and no shell variables, aliases, or history carry over between calls. If you need to operate in a subdirectory, chain commands in one call with `&&` or pass absolute paths (e.g., `cd /api-reference && ls` or `ls /api-reference`). Do NOT assume that `cd` in one call affects the next call. Examples: - `tree / -L 2` — see the top-level directory layout - `rg -il "rate limit" /` — find all files mentioning "rate limit" - `rg -C 3 "apiKey" /api-reference/` — show matches with 3 lines of context around each hit - `head -80 /quickstart.mdx` — read the top 80 lines of a specific page - `head -80 /quickstart.mdx /installation.mdx /guides/first-deploy.mdx` — read multiple pages in one call - `cat /api-reference/create-customer.mdx` — read a full page when you need everything - `cat /openapi/spec.json | jq '.paths | keys'` — list OpenAPI endpoints Output is truncated to 30KB per call. Prefer targeted `rg -C` or `head -N` over broad `cat` on large files. To read only the relevant sections of a large file, use `rg -C 3 "pattern" /path/file.mdx`. Batch multiple file reads into a single `head` or `cat` call whenever possible. When referencing pages in your response to the user, convert filesystem paths to URL paths by removing the `.mdx` extension. For example, `/quickstart.mdx` becomes `/quickstart` and `/api-reference/overview.mdx` becomes `/api-reference/overview`.

  Connector

• create_spec

  3TG Test Generation

  Generate a functional-requirements spec (`.3tg.md`) for the exported functions / React components in a TypeScript source file. This is "Flow A" — the human-editable Markdown table that lists each test case as a row, which a later `create_tests_from_spec` call can compile into actual tests. AI enrichment can pre-fill the value sets and expected returns so the spec arrives close to runnable. IMPORTANT — never hand-author a `.3tg.md` yourself. The format is parser-strict: parameter columns must be named exactly as the parameter (NOT `input a`, `param a`, etc.), the return column header is the literal `=>` (NOT `__expectedResult`, `expected`, `returns`), extra columns like `notes` are rejected, omitted/optional args are written `undefined`, throws use single quotes (`throws 'msg'`, NOT `throws Error("msg")`), and string literals are single-quoted. Always call this tool to emit the scaffold; the user can then edit rows. The returned `.3tg.md` is reported under the project's `.3tg/` mirror (e.g. source `src/foo/bar.ts` → spec `.3tg/src/foo/bar.3tg.md`). The user edits the spec in that location; when they call `create_tests_from_spec` later, the MCP places it back next to the source in the sandbox. Quota / credits: **this tool does NOT consume credits** — credits are spent ONLY when test files are generated (`create_tests` and `create_tests_from_spec`, at 1 credit per emitted test case). Spec generation is free; iterate on the scaffold as often as needed. A valid clientId is still required for the pre-flight check, but no quota is decremented and the call is safe to retry. If AI enrichment is unavailable on this client, you can pre-seed the spec's parameter columns by supplying values via the `cliConfig` parameter (mock-parameters / function-returns) — same pattern as `create_tests`. **Do NOT autonomously write `.3tg/config.3tg.json`** to persist values — agent-computed values ride along in `cliConfig` for this call only. (Explicit user requests to edit the file are fine — handle those normally.) See the cliConfig parameter description for the full shape. CRITICAL POST-CALL ACTION — write returned files to disk: The MCP server does NOT touch the user's filesystem. It returns the generated file CONTENTS in the response's `files` array. After this tool returns, you MUST iterate over `files` and write each entry's `content` verbatim to its `path` using your native file-write capability (e.g. Write / edit_file / create_file — whatever your client exposes). Create parent directories as needed. Returned paths are project-root-relative and already translated to the `.3tg/` mirror convention where applicable (e.g. specs land under `.3tg/<source-path>.3tg.md`; tests / mocks travel through unchanged). Write each path verbatim. Do NOT claim "Generated test file: <path>" unless you have actually written the file. The user will assume the MCP wrote it and waste time looking for a non-existent file. If you can't write for some reason (permission denied, no write capability in this client), return the contents inline in your message so the user can copy-paste them. Never report success silently when the write didn't happen.

  Connector

• create_tests_from_spec

  3TG Test Generation

  Compile a hand-edited functional-requirements spec (`.3tg.md`) into actual Jest/Vitest tests. This is "Flow B" — the user has already authored or reviewed the `.3tg.md` and is ready to materialise the rows into a runnable test file. Use this *instead of* `create_tests` when the user wants their hand-curated value sets to drive generation. Inputs: the source code plus the spec content (the spec lives at `.3tg/<sourceDir>/<basename>.3tg.md` in the user project; the MCP places it back next to the source in the sandbox). AI enrichment is NOT run — the spec is authoritative. 3TG also writes a `<basename>.md.3tg.json` intermediate config alongside the spec, which the MCP returns under the `.3tg/` mirror so the user can inspect what the spec compiled to. Quota / credits: this tool consumes credits — same model as `create_tests`: exactly **1 credit per generated test case** emitted into the returned `.test.ts` / `.test.tsx`. The number of rows in your `.3tg.md` table is therefore a reliable upper bound on what the call will cost. Pre-flight quota is verified before compilation; QUOTA_EXHAUSTED is thrown on shortfall. **Flow B cliConfig caveat — spec-authoritative keys are STRIPPED.** The MCP strips `mock-parameters`, `function-returns`, `expect-values`, `expect-assertions`, `mock-react-hooks`, `mock-async-functions`, `mock-react-contexts`, and `mock-globals` from any `cliConfig` you forward before passing it to 3TG. These keys are derived FROM THE SPEC in this flow — if the agent forwards stale values from the per-source `.md.3tg.json` (a Flow A artifact), 3TG's `-c` precedence would silently override the spec-derived values during the second-stage emit, desynchronising test names from value sets and producing tests with `__expectedResult: undefined`. For Flow B, forward ONLY global/structural config keys (`rules.*`, `creationMode`, `mockAsFunction`, `no-rule-default-true`, `ignore`, `package.json.type`, …) — the spec owns the test-value plan. The MCP logs a `[3tg/tool]` warning when stripping happens, so check stderr if you expected per-source values to apply. CRITICAL POST-CALL ACTION — write returned files to disk: The MCP server does NOT touch the user's filesystem. It returns the generated file CONTENTS in the response's `files` array. After this tool returns, you MUST iterate over `files` and write each entry's `content` verbatim to its `path` using your native file-write capability (e.g. Write / edit_file / create_file — whatever your client exposes). Create parent directories as needed. Returned paths are project-root-relative and already translated to the `.3tg/` mirror convention where applicable (e.g. specs land under `.3tg/<source-path>.3tg.md`; tests / mocks travel through unchanged). Write each path verbatim. Do NOT claim "Generated test file: <path>" unless you have actually written the file. The user will assume the MCP wrote it and waste time looking for a non-existent file. If you can't write for some reason (permission denied, no write capability in this client), return the contents inline in your message so the user can copy-paste them. Never report success silently when the write didn't happen.

  Connector

• run

  Bowmark

  **Executes the task on the real websites** (the search, the price check, the availability lookup, the configurator, the booking flow) and returns what came back. Runs a script you authored against the `get_library` vocabulary, on the live sites, and returns `{ ok, result, logs, error, ms }`. Call `get_library` FIRST — it gives the exact function names, argument shapes, and return types; this description is the LANGUAGE + how-to (get_library is just the vocabulary). THE LANGUAGE — plain async JavaScript: • `bowmark` is a ready global (no import). Call capabilities off it — `await bowmark.<capability>.<method>(...)` — always `await`, they're async. • Individual sites are callable too, at `await bowmark.providers.<provider>.<fn>(...)`. Use one when you specifically want THAT site; otherwise prefer the capability, which fans out across sites and routes around failures. • Real control flow: `await`, `if`, loops, array methods (`map`/`filter`/`sort`/`slice`), and `Promise.all` for fan-out. • `return` a value to get it back (JSON-serialized). `log(...)` for progress lines. • `bowmark` is the ONLY I/O — no `fetch`, `process`, filesystem, or `import`/`require`. Write a plain async body, not a wrapping function. • Keep scripts small and deterministic — no infinite loops. Runs in a hard sandbox with CPU + memory + wall-clock limits. SENDING IT: pass the script text as `run({ script })` — `script` is the only argument (there is no `site` argument; the library exposes every capability under `bowmark`). `result` is whatever you returned; `logs` are your `log()` lines in order; on a throw/timeout `ok:false` and `error` is set. CHECK `status` BEFORE `ok`. It is `ok` | `error` | `partial` | `needs_user`. • `partial` means the script RAN and `result` is real and usable, but some of what it called never answered — so the result is narrower than what you asked for. `ok` is still `true`; this is not a failure. `incomplete.summary` says what happened in one sentence, `incomplete.failures` names each call that threw and what the site said, and `incomplete.degraded` names each call that answered while reporting its OWN results thin. You MUST say so when you present the result: name what was missed, and do not describe it as complete, exhaustive, or 'all' of anything. A `partial` you report as whole is a wrong answer, not a slightly smaller right one. • Before you conclude a `partial` is final, check `incomplete.failures[].fixable`. `fixable: true` means YOUR ARGUMENT was rejected, not the site — the error text names what that function actually takes, so re-read it in `get_library`, fix the argument and run again; that recovers the whole answer. For any other failure re-running usually returns the same thing. • `needs_user` means a site needs the USER signed in — it is NOT a failure and NOT something you can fix by editing the script. `needs` lists the sites; `meta.handoff.url` is a single-use link that expires (`meta.handoff.expiresAt`). Give the user that URL, say which sites it covers, and WAIT. When they tell you they're done, send the SAME script again unchanged. Do NOT retry before then — it will stop at the same place and cost another run. Do NOT try to log in yourself, ask them for a password, or work around it with a different site. • Logged-in runs need a Bowmark API key on the connection; if you get `needs_user` saying so, tell the user to add one rather than retrying. `trace` is the execution trace — every capability you called and the providers it fanned out to under the hood: `[{ kind:'capability', capability:'flights', method:'search', ms }, { kind:'provider', capability:'flights', provider:'google_flights', fn:'search', results, status, ms }, …]`. The script never visits websites — it calls capabilities that route to providers, and the trace is the receipt. Composition is the point — call a method MULTIPLE times and combine results. To sweep a date range, call the search per date inside `Promise.all` and sort/filter the merged array (each flight result carries its `date`, so you can tell the runs apart). See the `get_library` examples for the exact shape. SOME capabilities return their rows alongside a `warnings` array — `{ flights, warnings }`, `{ hotels, warnings }`, `{ cars, warnings }`. Others return a bare array. The signature in `get_library` tells you which; go by it rather than assuming. Where there IS a `warnings` array it names any site dropped from the fan-out, and the rows themselves look identical with or without it. Read it, and pass on anything it says rather than quoting a 'cheapest' that only ranks the sites that happened to answer. Dropping `warnings` from what you return does not hide it — the run comes back `status: 'partial'` regardless, because the runtime counts what your script CALLED, not what it chose to report.

  Connector

• goal-add-evidence-text

  Planner — provable "done" for AI agents

  SUBORDINATE / supplementary path — does NOT close an acceptance criterion. Adds a text-only note (URL to a permanent external source like CI run / GitHub commit / issue, or a description of a manual scenario) as extra context alongside the real proof. The path that actually covers an AC and closes a Grove goal is goal-attach-evidence — use that one for every criterion. Plain evidence NEVER counts toward AC coverage no matter how many you add; it is only a complement to an attached file. NOT for bytes — screenshots, logs, API responses, exports all go through goal-attach-evidence. NOT for filesystem paths — those need goal-attach-evidence with the actual file.

  Connector

• query_docs_filesystem_political_comms

  docs

  Run a read-only shell-like query against a virtualized, in-memory filesystem rooted at `/` that contains ONLY the Political Comms documentation pages and OpenAPI specs. This is NOT a shell on any real machine — nothing runs on the user's computer, the server host, or any network. The filesystem is a sandbox backed by documentation chunks. This is how you read documentation pages: there is no separate "get page" tool. To read a page, pass its `.mdx` path (e.g. `/quickstart.mdx`, `/api-reference/create-customer.mdx`) to `head` or `cat`. To search the docs with exact keyword or regex matches, use `rg`. To understand the docs structure, use `tree` or `ls`. **Workflow:** Start with the search tool for broad or conceptual queries like "how to authenticate" or "rate limiting". Use this tool when you need exact keyword/regex matching, structural exploration, or to read the full content of a specific page by path. Supported commands: rg (ripgrep), grep, find, tree, ls, cat, head, tail, stat, wc, sort, uniq, cut, sed, awk, jq, plus basic text utilities. No writes, no network, no process control. Run `--help` on any command for usage. Each call is STATELESS: the working directory always resets to `/` and no shell variables, aliases, or history carry over between calls. If you need to operate in a subdirectory, chain commands in one call with `&&` or pass absolute paths (e.g., `cd /api-reference && ls` or `ls /api-reference`). Do NOT assume that `cd` in one call affects the next call. Examples: - `tree / -L 2` — see the top-level directory layout - `rg -il "rate limit" /` — find all files mentioning "rate limit" - `rg -C 3 "apiKey" /api-reference/` — show matches with 3 lines of context around each hit - `head -80 /quickstart.mdx` — read the top 80 lines of a specific page - `head -80 /quickstart.mdx /installation.mdx /guides/first-deploy.mdx` — read multiple pages in one call - `cat /api-reference/create-customer.mdx` — read a full page when you need everything - `cat /openapi/spec.json | jq '.paths | keys'` — list OpenAPI endpoints Output is truncated to 30KB per call. Prefer targeted `rg -C` or `head -N` over broad `cat` on large files. To read only the relevant sections of a large file, use `rg -C 3 "pattern" /path/file.mdx`. Batch multiple file reads into a single `head` or `cat` call whenever possible. When referencing pages in your response to the user, convert filesystem paths to URL paths by removing the `.mdx` extension. For example, `/quickstart.mdx` becomes `/quickstart` and `/api-reference/overview.mdx` becomes `/api-reference/overview`.

  Connector

• run_code

  CoreModels

  Execute a JavaScript program that orchestrates this server's tools, and return only its result. Prefer this over many individual tool calls when a task needs several steps, looping, filtering, or combining data: intermediate results stay in the sandbox, so only what you return reaches the model. Inside the script: - listTools() -> [{name, summary}] discover available tools - getToolDoc(name) -> {name, description, parameters, required} inspect one tool's inputs - tools.<name>(args) -> parsed result call a tool (graphProjectId is injected automatically; do NOT pass it) - console.log(...) captured and returned alongside the result - return <value> JSON-serialized and returned Environment: sandboxed JavaScript, no network or filesystem, with limits on time, memory, statements and number of tool calls. Currently only read-only tools are callable from code.

  Connector

• goal-attach-evidence

  Planner (proof-of-done)

  PRIMARY path to close a Grove goal: this is the ONLY tool that covers an acceptance criterion. Attach binary evidence (screenshot, log dump, API response, export) to an AC — call it once per criterion to satisfy the close gate. The subordinate goal-add-evidence-text only adds context for proofs with NO bytes (URLs to permanent external sources, manual repro descriptions) and does NOT cover an AC. Caption is optional but strongly recommended: state what the file captures and the reproduction conditions (URL/commit/session/inputs) so a third reviewer can reproduce. ⚠ PICK THE RIGHT TRANSPORT BEFORE YOU CALL THIS TOOL ⚠ • BEST for ANY file > ~1 KB raw — and the ONLY no-token path, so use it in a claude.ai / hosted-agent session that has no raw X-Auth-Token → call the sibling MCP tool `goal-request-upload` with this same criterionId. It returns a one-time {uploadUrl, expiresAt}; then stream the raw bytes with a single PUT: `curl -sS --fail --upload-file "/abs/path/to/file.png" "<uploadUrl>"` (optionally add -H "X-Content-Sha256: <hex sha256>" so corruption fails fast). No base64, no token — the signed ?t= ticket in the URL is the only credential, single-use, criterion-scoped. The PUT response is the same evidence JSON this tool returns. • ALTERNATIVELY, if you DO have the raw X-Auth-Token in your shell → the `planner-attach.sh` helper (zero-install bash, binary-safe). The MCP base64 path below is unreliable for non-trivial files: long string arguments get truncated or whitespace-corrupted on the agent side BEFORE the JSON-RPC request is sent. Measured 2026-05-20 on prod: a 4 KB PNG arrived at the server as 1874 decoded bytes (file_hash_mismatch); a 2 KB payload arrived with stray whitespace (failed base64_decode). The server itself accepts up to 25 MiB raw — the bottleneck is the agent-side serialisation of contentBase64, NOT the server. planner-attach.sh COPY-PASTE RECIPE (replace 3 placeholders, run in your shell): curl -sS https://planner.monopoly-gold.com/api/cli/planner-attach.sh \ | PLANNER_TOKEN="<same X-Auth-Token you use for MCP>" bash -s -- \ --criterion-id "<CRITERION_UUID>" \ --file "/abs/path/to/file.png" \ --caption "what is captured and the repro conditions" \ --created-by "<your agent id>" Where to get each value: - PLANNER_TOKEN: the very same token that is already in your MCP config under the X-Auth-Token header for the `planner` server. NOT a separate credential. - CRITERION_UUID: the AC id you got from goal-get / goal-list. Same UUID you would pass to this MCP tool. - file path: absolute path on YOUR (agent) machine — the script reads it locally and streams multipart. The planner server never sees your filesystem. The helper computes SHA-256 itself and ships it as `contentSha256`, so any in-flight corruption fails fast with HTTP 400 instead of poisoning the evidence row. Output on stdout is the same JSON shape this MCP tool returns; non-zero exit means HTTP ≥ 400 (stderr explains). Without curl/bash? Fall back to raw multipart: POST https://planner.monopoly-gold.com/api/criteria/<id>/evidence/file, header X-Auth-Token, form fields file=@..., contentSha256=..., caption, createdBy. • File ≤ ~1 KB raw → this MCP tool is fine. ALWAYS pass `contentSha256` (hex SHA-256 of raw bytes BEFORE base64). Without it, a silently truncated PNG looks valid to the MIME sniffer; the server cannot distinguish a truncated 4 KB PNG from a valid 1 KB one and the vision judge burns ~30s on broken bytes. With the hash, the server fast-fails with error=file_hash_mismatch and points back here at the multipart endpoint. Validates MIME whitelist (png/jpeg/webp/gif/pdf/txt/json/zip), per-file size cap (ATTACHMENTS_MAX_FILE_BYTES, default 25 MiB), per-project attachments quota. Returns evidence record + file URL + serverSha256.

  Connector

• firecrawl_parse

  firecrawl-mcp

  Parse one supported document into markdown, HTML, links, summary, targeted answers, or JSON matching a schema. Supported inputs include common HTML, PDF, Word, RTF, OpenDocument, and spreadsheet files; PDF parsing can be bounded with `pdfOptions.maxPages`. Local MCP reads `filePath` from the server filesystem. Hosted MCP uses two calls: first provide `filePath` to receive upload instructions, upload locally, then call again with the returned `uploadRef`; do not send both fields together. Remote web URLs belong in `firecrawl_scrape`. Set `redactPII` to request redaction of personally identifiable information in the returned content. `zeroDataRetention` requires an eligible authenticated account; omit it for anonymous keyless use. Returns upload instructions for hosted phase one or parsed document content for the final call.

  Connector

• upgrade_to_cluster

  mcp

  THE WAY TO MAKE AN EXISTING APP HIGHLY AVAILABLE. If the user asks for high availability, redundancy, no single point of failure, surviving a host or VM failure, or 'stay up if a machine dies', this is the tool: point it at their running VM and pass high_availability:true. It needs no rebuild and no redeploy, and the URL does not change. Turns a single running instance into a horizontally-autoscaling cluster: snapshots the instance and puts your SOURCE VM behind a load balancer as its ALWAYS-ON BASELINE MEMBER, then an autoscaling group adds EXTRA members (booted from that snapshot) on CPU load and scales them back down to zero. FLAVOR — LEAD WITH A MEMORY-OPTIMIZED SMALL-DISK FLAVOR (m1.mem16/m1.mem32) for the cluster, and ideally have the SOURCE VM on one too: clustering snapshots the source VM's FULL disk, so DISK size (not RAM) drives how long the cluster takes to build and to add each member under load — a 40 GB mem flavor is several times faster to snapshot and scale than a 160 GB standard flavor (m1.xlarge) and costs less for the same RAM. If the source VM is on a big-disk flavor, RECOMMEND migrating the app onto a mem-flavor VM first (a root disk cannot be shrunk in place), even though it is extra work — it pays off on every scale-out, and a clustered app is stateless anyway (state in the managed DB) so the lean disk fits. Expect a DRASTICALLY better clustering experience on a mem flavor: far faster cluster-create, snapshots, and every scale-out. The user can override with any flavor. At rest ONLY your source VM serves — there is NO idle extra VM to pay for (the source VM is the cluster's minimum, so the floor is 0 extra members). Use it to LOAD-SCALE a stateless app tier while managed services hold state: it becomes highly-available UNDER LOAD (multiple members behind the LB), but at rest a SINGLE source VM serves — and that source VM is a plain VM, not an autoscaled member, so it is NOT auto-replaced if it fails while idle (only the autoscaled extra members are ASG-managed and self-healed). If you need always-on redundancy, keep the app under enough load to hold >=1 extra member, or use a separate always-on setup. BILLED — at rest it costs just your source VM (which you already run) plus the load balancer; under load it adds up to max_size EXTRA members at the member flavor (flavor_id), billed only while they run. In guided mode show the cost that way (now: source VM already running + the LB; under load: up to max_size x the member flavor) and get the user's explicit go first. redu automatically repoints the extra members from the old single-VM URL to the load-balancer URL across app config. It REFUSES a STATEFUL VM with 409 cluster_needs_stateless unless confirm_stateless:true. To have redu FIX a stateful VM for you instead of refusing, pass auto_restructure:true — for a single_vm Postgres it fully-automatically provisions a managed DB + migrates the data + repoints the members; for a compose-stack DB it provisions the matching managed DB (set restructure_engine, e.g. 'mysql'/'mariadb' for WordPress) and returns migration commands to run from the app VM. WordPress/WooCommerce is not generic autoscaling: managed DB alone is not enough because wp-content/uploads is file state. Use app_profile:'wordpress'/'woocommerce', cluster_media_mode:'media_space', and either media_space_id or create_media_space:true so all members mount the same uploads filesystem; otherwise the backend refuses with 409 cluster_needs_media_space. PUT THE CLUSTER ON THE SAME private network as the managed DB and media space. HA: cluster members are spread across DIFFERENT physical hosts automatically, and an autoscaled member that is destroyed is REBUILT AUTOMATICALLY in 1.5 to 5 minutes depending on how it failed with no action from you (the always-on source/hero VM is a plain VM and is NOT covered by that). CRITICAL for members: the app must start on EVERY boot (systemd unit or container restart policy) - if it only starts from a first-boot cloud-init script, a rebooted or resized member comes back with no app, silently never rejoins the load balancer, and the cluster quietly loses capacity with nothing reporting an error. Pass startup_command if the app does not already auto-start on boot, and have it bind its port only once it is genuinely ready to serve (the health check can only see whether the port is open). SEQUENCING - this catches people: the snapshot is taken IMMEDIATELY, and every member boots from it, so the source VM's app must already be RUNNING before you call this. Clustering a freshly-created VM whose cloud-init has not finished captures an image with no enabled service, and all members then come up ACTIVE while failing the load-balancer health check forever - a cluster that looks built and serves nothing. Verify the app answers on its port first (get_ssh_command, or just fetch the VM's URL). The snapshot upload can take several minutes; poll list_clusters until CREATE_COMPLETE.

  Connector

• attachments

  Homespun

  Binary attachments (images, PDFs, audio, video) referenced from event payloads and input_data via `format: homespun-attachment-id`. Actions: upload, fetch, presign, finalize, download, show, list, delete, mint_token, revoke_token, list_tokens. Choosing an upload path matters for cost. An inline upload with `content_base64` carries the bytes in the tool-call arguments, so they enter the model context at a token cost proportional to file size, paid again on every retry; a few-hundred-KB image is already expensive. Two paths avoid that entirely: fetch, when the bytes are reachable at a URL, and presign plus finalize, when the client can PUT the raw bytes out of band. Inline upload suits small assets and clients that have neither a URL nor an out-of-band PUT. fetch takes { source_url (https), scope } and the relay downloads the URL itself behind an SSRF guard (https only, no private, loopback or metadata hosts, DNS pinned, redirects refused, size-capped and timed out), then runs the same byte-sniff, allowlist, size, quota and scan checks as any upload. It works on any storage backend. upload takes either `content_base64` (base64 bytes, no filesystem) or `file_path` (an absolute path read on the relay host, so it only applies when the file is local to the relay). presign plus finalize is token-free: presign with { mime, size, sha256, scope } returns { put_url, attachment_id }, the caller PUTs the raw bytes to put_url over plain HTTP out of band, then finalize with the attachment_id. At finalize the relay re-reads the stored bytes, sniffs the real type, and enforces the same allowlist, size, sha256, quota and scan checks, so a presign that misstates its mime is caught and never served inline. The presigned path requires the Azure storage backend; a filesystem-backed relay returns a clear not-supported error and fetch or inline upload apply there instead. download writes to an absolute out_path or returns base64. An upload is scoped to agent (the default, reusable) or app. mint_token returns a /b/<token> capability URL, shown once, that a browser can GET without the caller's API key.

  Connector

• create_race_plan

  CourseProfiler

  Primary CourseProfiler tool for creating a personalized race plan and optional PDF. Use this after obtaining both a course input and runner input. Course input may be course_file, course.source with kind=url/artifact/raw_file, or course.crsprof_artifact_id. Do not use raw_json or reconstruct/synthesize a course from roadbooks, checkpoint tables, elevation profiles, aid-station lists, or screenshots; those are enrichment/context only and are not valid course geometry. Runner input may be runner_profile_file, runner.sources, or runner.usrprof_artifact_id. Prefer generate_runner_profile first when the user has GPX/FIT/CRSProf activity evidence or an uploaded .usrprof source but no already-converted usrprof_artifact_id; then call create_race_plan with runner.usrprof_artifact_id. Before generating a runner profile for a race plan, resolve the target course if possible and ask the user whether the profile should be target-race-specific, target-distance/elevation-specific, general trail, or insights-only; select evidence that matches the intended distance, elevation gain/loss, terrain/technicality, altitude, duration, and recency. Do not silently use all files in a local folder. Use runner.usrprof_artifact_id only for an already-converted USRProf artifact; uploaded .usrprof/source_file artifacts from POST /api/artifact-uploads must be passed through runner.sources[{kind:'artifact'}] or through generate_runner_profile. Bare local filesystem paths never work in hosted MCP clients; send inline content/base64 when the MCP client can read the file, or use the REST POST /api/artifact-uploads flow outside MCP, complete the upload, then pass the returned source_file artifact ID. There is no raw-byte MCP upload tool; get_artifact_upload_requirements only explains the REST flow. Use get_runner_profile_requirements when the user needs instructions for creating/exporting a USRProf, using Strava/browser import, or uploading GPX/FIT/CRSProf evidence. If the user provides only a race name, check the CourseProfiler race catalog manifest first, but do not rely on uniqueness alone: confirm the catalog match is the same event/course/location before using its assetPath CRSProf URL. Auto-resolution only uses exact/strong catalog matches; weak matches require explicit confirmation or an explicit course.source URL/artifact. Use search_race_catalog first when the user may need to choose among multiple matching catalog courses or when the match confidence is not clearly exact/strong. If no catalog match exists, try to resolve it to a fetchable official GPX/FIT/CRSProf URL or ask the user to upload/provide the course file or URL; do not stop at the race name. Third-party route hosts such as Wikiloc may return 403 to server fetches. If an official or third-party GPX/FIT/CRSProf URL cannot be fetched or is blocked, stop and ask the user to download the official file and upload it through POST /api/artifact-uploads, then continue with the returned source_file artifact ID. Do not create a race plan until a real route file/trusted CRSProf is available. Use import_course to create a CRSProf, enrich_course_waypoints to add structured aid stations/resources/cutoffs, then generate_course_segments before creating the race plan. Catalog CRSProf files may already include official waypoints/resources/cutoffs; if they do not, call enrich_course_waypoints or ask the user for structured aid/resource/cutoff data before segmentation. If official pages and regulation PDFs disagree, or exact aid locations are not fully listed in machine-readable form, ask the user to confirm and include only confirmed aid stations; do not invent missing locations. Route-only plans are incomplete unless the user explicitly accepts missing aid/resource details. If browsing/search is available, prefer the catalog first, then official race sources and direct GPX links over generic home pages; course-only pacing is incomplete without waypoint/resource enrichment. If runner data is missing, ask for a USRProf file/artifact or runner evidence, explain that the profile drives estimated times plus insights such as uphill running limit, durability/fatigue tendencies, downhill sensitivity, and terrain strengths/weaknesses, generate a segment-evidence runner profile with generate_runner_profile after user-approved evidence selection, and do not invent personalized fitness data. This race-plan flow creates a personalized plan/PDF only; it does not submit the course to the public catalog and does not need to. Only call submit_course when the user explicitly asks to submit/add/update a course for catalog review. Returns a top-level job with artifact role metadata; use get_job to poll and get_artifact to fetch outputs.

  Connector