ai.plith/plith

Measured conclusion: Alive, not deep-tested. L2 outcome: not_attempted.

AI agent infrastructure: dedup, cost prediction, validation, governance, failure intelligence.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "plith": {
      "url": "https://plith.ai/api/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (15)

• dedupq_check

  Before executing any LLM task, check if an identical or semantically similar task has already been completed. Returns cached result on hit, saving one LLM call. On a miss, execute your task and call dedupq_complete to cache the result for future hits. Costs 1 credit.

  Parameters:

  • content (required) — string — The task content to check for duplicates. This is hashed and embedded for matching.
  • task_id — string — Optional caller task ID for tracing and cross-referencing with BurnRate.
  • hash_only — boolean — If true, skip vector similarity search and use exact hash matching only. Default: false.
  • similarity_threshold — number — Cosine similarity threshold for semantic matching, 0.0 to 1.0. Default: 0.80.

  Hints: read-only no · destructive no · idempotent yes · open-world yes

• dedupq_complete

  After executing a task, store the result so future identical or similar tasks return a cache hit via dedupq_check. Costs 2 credits.

  Parameters:

  • content (required) — string — Original task content. Used to compute hash and embedding for future matching.
  • result (required) — The task result to cache. Can be any JSON value.
  • task_id — string — Optional task ID. Used as the database row ID if provided.
  • hash_only — boolean — If true, skip embedding generation. Default: false.

  Hints: read-only no · destructive no · idempotent yes · open-world yes

• burnrate_estimate

  Before executing a multi-step agent plan, estimate the total LLM cost. Returns per-step breakdown and optimization suggestions. If the estimate exceeds your budget, pipe the same plan into burnrate_optimize. Costs 1 credit.

  Parameters:

  • plan (required) — array — Array of plan steps with provider, model, and token estimates.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• burnrate_track

  Log the actual cost of an LLM call after execution. Call this after every LLM request to build calibration data that improves burnrate_estimate accuracy over time. Free — no credits charged. Returns the recorded cost entry with computed margin versus the prior estimate when one exists for this model and token range.

  Parameters:

  • provider (required) — string — LLM provider identifier. Supported: anthropic, openai, google, mistral, cohere, deepseek, together, fireworks, groq. Must match the provider of the model used.
  • model (required) — string — Model identifier as returned by the provider. Examples: claude-sonnet-4-6, gpt-4o, gemini-2.0-flash, mistral-large-latest. Unknown models are accepted but cost may show as $0.
  • input_tokens (required) — number — Actual prompt tokens used. Must be >= 0.
  • output_tokens (required) — number — Actual completion tokens used. Must be >= 0.
  • task_id — string — Optional task ID for cross-referencing spend with DedupQ deduplication results. Use the same task_id passed to dedupq_check to link cost tracking with deduplication.
  • cache_read_tokens — number — Optional. Cache-read tokens.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• burnrate_optimize

  Get a cheaper equivalent plan by substituting models with lower-cost alternatives. Call after burnrate_estimate if the estimated cost exceeds your budget. Returns the optimized plan with substituted models, new per-step costs, total savings, and whether the target_budget is met. Optionally set target_budget to constrain the optimization. Costs 1 credit.

  Parameters:

  • plan (required) — array — Array of plan steps. Same schema as burnrate_estimate: each step needs step, provider, model, estimated_input_tokens, estimated_output_tokens.
  • target_budget — number — Optional. Target total cost in USD.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• burnrate_budget read-only

  Get today's tracked LLM spend, per-model breakdown, projection, and budget alerts. Free — no credits charged.

  Parameters:

  • daily_limit — number — Optional. Daily budget in USD (e.g., 10.0 for a $10/day cap). Enables budget alerts and remaining-balance calculation.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• qualitygate_validate

  After your agent generates output, validate it against your rules before shipping. Runs deterministic checks (regex, JSON schema, syntax) plus optional LLM-powered tone and factual analysis. Returns a structured verdict (pass, warn, or fail) with a 0-100 score and per-check issue details. Use qualitygate_trends to spot recurring failure patterns over time. Variable cost: 1 credit per deterministic check, 8 credits per LLM check.

  Parameters:

  • output (required) — string — The agent output text to validate.
  • schema — object — JSON Schema to validate output against.
  • language — string — Code language for syntax check: json, python, javascript, typescript.
  • override — boolean — Force pass. Requires override_reason.
  • directives — array — Directive objects. Types: must_include, must_not_include, must_match, must_not_match, must_contain, must_not_contain, min_length, max_length.
  • check_types — array — Checks to run. Auto-inferred if omitted.
  • override_reason — string — Required when override is true.

  Hints: read-only no · destructive no · idempotent yes · open-world yes

• guardrail_check

  Evaluate a proposed agent action against your governance policies. Returns allow or deny with the matched policy reason. Requires at least one active policy created via guardrail_create_policy. Deterministic rule evaluation — no LLM. Costs 1 credit.

  Parameters:

  • agent_id (required) — string — Agent identifier.
  • proposed_action (required) — object — Action to evaluate. Must contain a 'type' field. Example: {"type": "http_request", "url": "https://external.example.com"} or {"type": "file_write", "path": "/etc/config"}.

  Hints: read-only no · destructive no · idempotent yes · open-world no

• guardrail_create_policy

  Create a persistent governance policy that guardrail_check evaluates on every subsequent call. Define rules using and/or/not operators over action types, resource patterns, and budget thresholds. Call this before using guardrail_check — checks require at least one active policy. Policies persist until explicitly deleted. Duplicate policy names return an error. Returns the created policy with its ID and active status.

  Parameters:

  • name (required) — string — Unique policy name per org. Examples: 'no-delete-in-prod', 'budget-cap-50', 'pii-block'.
  • rules (required) — array — Array of rule objects evaluated against the proposed_action in guardrail_check. Leaf operators: eq, starts_with, contains, gt, lt (compare field to value). Compound operators: and, or, not (nest sub-rules in a rules array). Example: [{operator:'eq', field:'type', value:'file_write'}] blocks all file writes. Nested example: [{operator:'and', rules:[{operator:'eq',field:'type',value:'api_call'},{operator:'contains',field:'url',value:'prod'}]}] blocks prod API calls.
  • priority — number — Optional. Evaluation order. Default: 0.
  • description — string — Optional human-readable summary of what this policy enforces. Returned in guardrail_check responses and guardrail_list_policies output for auditability.
  • action_types — array — Optional. Restrict this policy to only evaluate when proposed_action.type matches one of these values. Examples: ['file_write', 'api_call', 'db_delete']. Omit to apply the policy to all action types regardless of type field.

  Hints: read-only no · destructive no · idempotent no · open-world no

• pitfalldb_query

  Check for known failure patterns before executing a task type. Returns pitfalls with severity, fix suggestions, and confidence scores. After your agent runs, submit failures via pitfalldb_report so others benefit. Costs 2 credits.

  Parameters:

  • task_type (required) — string — Task category: code_generation, web_search, data_analysis, etc.
  • filters — object — Optional filters.
  • task_description — string — Optional. Natural-language task description for semantic search.

  Hints: read-only no · destructive no · idempotent yes · open-world yes

• pitfalldb_report

  Report an agent failure. PII-scrubbed before storage. Linked to existing pitfalls if similar. Free — no credits charged.

  Parameters:

  • task_type (required) — string — Task category.
  • task_description (required) — string — Description of the failed task.
  • failure (required) — object — Failure details.

  Hints: read-only no · destructive no · idempotent yes · open-world yes

• rigor_plan read-only

  Before executing a complex task, get a structured workflow plan with per-step cost estimates. Classifies your task, selects the optimal framework sequence, and returns the full plan without executing anything. The response's allowed_modes tells you whether this plan is eligible for direct execution. Free — no credits charged.

  Parameters:

  • task_description (required) — string — Natural language description of the task, at most 4,000 characters. Be specific — include what you want produced, constraints, and context. Example: 'Design a caching layer for our API gateway with Redis integration.'
  • task_type — string — Optional hint to bypass automatic classification. Passing it also removes the slowest classification tiers from the critical path, so send it whenever you know the shape. Multi-step deliverable types: solution_design, requirements_analysis, code_implementation, code_review, bug_fix, root_cause_analysis, incident_response, deployment_execution, competitive_scan, financial_analysis, research_task, documentation, governance_change, compliance_audit, data_security_assessment, performance_optimization, user_story_definition, implementation_prompt_generation. Atomic single-call types, which auto-select direct execution: tag, score, rerank, compose, extract_entities, parse_query, quick_research, quick_classification, quick_extraction, quick_scoring. Call GET /api/rigor/task-types for the full vocabulary with each type's shape.
  • preferences — object — Optional workflow preferences.
  • context_fields — array — Optional, with prior_workflow_id: the fields to carry over. Pass ["full_step_results"] for every step output in full.
  • prior_workflow_id — string — Optional: the workflow_id (wr_…) of a finished workflow in your workspace to build on. Its results are passed to this run as context. It is charged as a new run at the normal quote.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• rigor_execute

  Execute a structured workflow end-to-end. Call rigor_plan first (free) to preview the step sequence and cost estimate before committing credits. Classifies the task, selects the optimal tool sequence, and executes each step with the right LLM model. Returns a complete deliverable — solution designs, competitive analyses, governance documents, and more. Supports SSE streaming for real-time progress, webhook callback, or polling. For atomic work — classification, scoring, ranking, entity extraction, query parsing — set preferences.execution to 'direct' and declare preferences.output_contract to get validated JSON records from a single call, routed to the cheapest model that holds the schema.

  Parameters:

  • task_description (required) — string — Natural language description of the task, at most 4,000 characters. Be specific — include what you want produced, constraints, and context. Example: 'Design a caching layer for our API gateway with Redis integration.'
  • context — object — Additional context for the workflow.
  • delivery — object — Delivery method. Default: polling (MCP clients typically can't handle SSE).
  • task_type — string — Optional hint to bypass automatic classification. Passing it also removes the slowest classification tiers from the critical path, so send it whenever you know the shape. Multi-step deliverable types: solution_design, requirements_analysis, code_implementation, code_review, bug_fix, root_cause_analysis, incident_response, deployment_execution, competitive_scan, financial_analysis, research_task, documentation, governance_change, compliance_audit, data_security_assessment, performance_optimization, user_story_definition, implementation_prompt_generation. Atomic single-call types, which auto-select direct execution: tag, score, rerank, compose, extract_entities, parse_query, quick_research, quick_classification, quick_extraction, quick_scoring. Call GET /api/rigor/task-types for the full vocabulary with each type's shape.
  • preferences — object — Optional workflow preferences.
  • context_fields — array — Optional, with prior_workflow_id: the fields to carry over. Pass ["full_step_results"] for every step output in full.
  • prior_workflow_id — string — Optional: the workflow_id (wr_…) of a finished workflow in your workspace to build on. Its results are passed to this run as context. It is charged as a new run at the normal quote.

  Hints: read-only no · destructive no · idempotent no · open-world yes

• rigor_status read-only

  Check the status of a running or completed Rigor workflow. Returns progress, step results, and the full deliverable when complete. Use after rigor_execute with polling delivery to retrieve results.

  Parameters:

  • workflow_id (required) — string — The workflow ID returned by rigor_execute (format: wr_xxx).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• rigor_workflows read-only

  List and search Rigor workflows for your organization, with filtering and pagination. Returns status, progress, capacity usage, and available actions per workflow. Use to monitor workflow state, understand concurrent limit usage, identify stuck or completed workflows, and — via q — find prior work on a subject before commissioning it again. Pair a q hit with rigor_status to read that workflow's deliverable.

  Parameters:

  • q — string — Search the workflow title and task description. Every whitespace-separated term must appear in one or the other, as a case-insensitive substring — so "vector search postgres" matches a task described as "add vector search to an existing Postgres-backed SaaS app". Substring matching, not full-text: there is no stemming and no ranking, so "migrating" does not match "migration". Max 200 characters and 8 terms; punctuation is treated as a separator and wildcards are not supported.
  • limit — number — Page size (default 20, max 100)
  • cursor — string — Pagination cursor (created_at timestamp from previous page)
  • status — string — Filter by status (comma-separated). Valid values: executing, step_executing, completed, failed, halted, pending_approval, cancelled. E.g. "halted,failed,pending_approval"
  • folder_id — string — Filter by folder ID. Pass "unassigned" for workflows in no folder
  • task_type — string — Filter by classified task type
  • created_after — string — ISO timestamp — only workflows created after this time
  • created_before — string — ISO timestamp — only workflows created before this time
  • counts_toward_limit — string — Filter to workflows counting toward the concurrent limit

  Hints: read-only yes · destructive no · idempotent yes · open-world no

Repository

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.