Loading ToolBench

Reading the current corpus…

forge MCP server · C · ToolBench

forge

A multi-module Go application that provides agent tooling, CLI operations, skills management, plugins, and UI components. This is a monorepo containing forge-core (runtime/tools), forge-cli (command-line interface), forge-skills (skill definitions), forge-plugins (extension plugins), and forge-ui (user interface).

2026-07-28+STREAMABLE_HTTP10 tools

60/100

Definition Quality69

Protocol Readiness35

Supportability60

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence · detected: Dynamic Client Registration

Deprecated protocol patterns detected

Summary

Forge has 10 tools with mixed quality. Naming is generally good (verb-first: shell, browser_navigate, browser_state, browser_click, browser_fill, browser_extract, browser_screenshot, local_shell, local_file_browser, json_parse). Descriptions vary significantly, browser tools are well-documented (150-200 chars, context-rich), but devtools and json_parse are sparse or test-only. Schemas are present for most tools but inconsistent in completeness: browser and shell tools have full input schemas with type definitions and parameter descriptions; devtools tools (local_shell, local_file_browser) show minimal schemas from test files, not production code; json_parse has no visible schema. All tools lack documented output schemas, critical for composition planning. Error handling is largely absent; tools provide no guidance for recovery or next steps. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are visible, despite destructive tools (shell, local_shell) being present.

Tools (10)

browser_clickwritesource verified85/100

Click the interactive element with the given index from the latest digest, then return the resulting page digest. Requires the digest's generation number; if the page changed since, the error includes a fresh digest to act on.

browser_extractread onlysource verified87/100

Extract content from the current page: mode=text (readable markdown, default), mode=links (deduplicated [text](url) lines), or mode=html (outerHTML of a CSS selector — escape hatch, prefer text). Long content is paginated: pass offset from the previous call's header to continue.

browser_fillwritesource verified83/100

Type text into the input/textarea/contenteditable with the given index (replaces existing content), or pick an option of a select by its label. Password and payment fields are refused unless the skill opted in. Set submit=true to press Enter afterwards. Returns the resulting page digest.

browser_navigateread onlysource verified88/100

Load a web page in the managed browser and return its digest: title, URL, indexed interactive elements ([N] lines), and the start of the page text. Use the indices with browser_click/browser_fill. Only http/https URLs; navigation is subject to the agent's egress allowlist.

browser_screenshotread onlysource verified42/100

Capture a screenshot of the current browser page.

browser_state

Issues (8)

CRITICALtool ↗

Output schemas absent across all 10 tools

shellbrowser_navigatebrowser_statebrowser_clickbrowser_fillbrowser_extractbrowser_screenshotlocal_shelllocal_file_browserjson_parse

HIGHtool ↗

Three tools sourced from test files, not production code (local_shell, local_file_browser, json_parse)

local_shelllocal_file_browserjson_parse

HIGHconfirmation-request ↗

Destructive tools lack confirmation or dry-run patterns

shelllocal_shell

MEDIUMtool ↗

No tool annotations (readOnlyHint, destructiveHint, idempotentHint) visible

shellbrowser_navigatebrowser_statebrowser_clickbrowser_fillbrowser_extract

Recommendations

ADD OUTPUT SCHEMAS to all 10 tools. Specify exact field names, types, and meanings. Example for browser_navigate: {type: object, properties: {title: string, url: string, elements: array of {index: int, text: string, type: string}, digest_generation: int}}
MOVE tool definitions from test files (local_shell, local_file_browser, json_parse) to production code with formal tool registration. Include complete input/output schemas and parameter descriptions.
PROMOTE safety annotations from description text to formal MCP tool annotations. Use readOnlyHint: true for browser_extract, browser_screenshot, browser_state, browser_navigate. Use destructiveHint: true for shell, local_shell. Use idempotentHint: true where applicable.
ADD CONFIRMATION PATTERN for destructive tools. Implement confirm_before_execute via MRTR result {status: 'input_required', description: 'Delete file /path/to/file?', choices: ['yes', 'no']} before shell rm operations.
DOCUMENT ERROR CASES for each tool with actionable recovery. Example: shell: 'If exit code 127, command not found, check PATH or use absolute path. If exit code 1, check command syntax with `help COMMAND`.'; browser_navigate: 'If timeout after 15s, page may be slow, retry with increased wait_ms or try a different URL.'
EXPAND browser_screenshot description from 39 to 150+ chars. Specify input parameters (selector? delay? viewport size?) and output format (base64 PNG? path to temp file?).
CLARIFY json_parse purpose and scope. If for agent internal use, document input/output format and error handling for malformed JSON. If for users, provide examples and limits (max JSON size?).

Score history

Overall score trend

↑ 25 points across a rubric change (v1 → v2)

60/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-22	C	60	2026-07-28+	v2
2026-03-09	F	35	-	v1