Algenta MCP Server

Estimate the best value for each decision variable given a plain-English objective, and return the per-variable optima. Every variable is sampled uniformly over its [min, max] range; an objective containing 'maximize' favors each variable's max, anything else favors the min, and the returned optimum blends that endpoint with the range midpoint. Use product_decision when you want a proceed/pause/reject recommendation instead of raw optima. Synchronous deterministic compute; nothing is persisted. Returns optimal_values, objective_value, improvement_vs_midpoint (percent), constraints_satisfied, and iterations_run.

Rank caller-supplied documents against a search query and return the top matches with snippets. Scoring is deterministic lexical word-overlap between query and document plus a bonus when the query prefix appears in the document; results sort by relevance_score with rank starting at 1. Provide documents or a collection_id - a call with neither fails with missing_source. Use query_data for analytics over connected datasets instead. Read-only; nothing is stored. Returns results (rank, document_id, content excerpt, relevance_score, snippet) and total_searched.

Search persisted checkpoints across all of the organization's agent runs, paginated (defaults page 1, limit 25). Filter by run_id or checkpoint_id to pinpoint one, or by status, request_hash, policy_snapshot_id, or schema_snapshot_id to audit lineage. Use get_agent_run_checkpoints when you already know the run_id and want its full checkpoint list. Read-only. Returns data, total, page, limit, and pages.

Search canonical mission-event records across all of the organization's agent runs, paginated (defaults page 1, limit 25) and newest first. Filter by run_id or event_type to pinpoint, or by status, request_hash, policy_snapshot_id, or schema_snapshot_id for lineage audits. Use get_agent_run_mission_events when you already know the run_id. Read-only. Returns data, total, page, limit, and pages.

Search runtime telemetry batches across all of the organization's agent runs, paginated (defaults page 1, limit 25). Filter by run_id, telemetry_kind, or module_name to pinpoint, or by status, request_hash, policy_snapshot_id, or schema_snapshot_id for lineage audits. Use get_agent_run_telemetry when you already know the run_id. Read-only. Returns data, total, page, limit, and pages.

Execute several governed exact queries in one API call. Use this for multi-metric prompts after choosing a dataset with list_data and get_data_summary. Each item reuses the same structured query contract as query_data; defaults may provide shared dataset_id, filter, limit, and order. Read-only against the engine; executes under the active API key with no separate per-route rate limit. Returns request_id and a results array with each item's key, data envelope, metadata, or error.

Execute a structured query against connected data sources. Convert the user's question to a structured intent and call this tool — do NOT try to write SQL or parse column names yourself. The engine resolves column meaning from mathematical relationships and statistical structure only. It works on any dataset without configuration. The governed filter shape is a record-predicate contract over normalized rows, not a SQL predicate language, so it also applies to Redis and other non-SQL sources. Structural roles (use in metric.role): - derived_measure: the main financial/operational aggregate (revenue, spend, value) - base_measure: counts, quantities, discrete amounts - unit_measure: per-unit prices, rates - ratio: percentages, margins, fill rates (0-1 range) - metric: let the engine pick the best numeric column If clarification_required is true, or if confidence < 0.85, check the candidates list and ask the user to clarify. Never fabricate column names or SQL. Read-only against the engine; executes under the active API key with no separate per-route rate limit. Returns the query envelope: result, result_type, row_count, confidence, resolved_column, decision_path, and plan, with candidates and clarification_required set when the engine cannot resolve deterministically.

Walk the dependency graph of one persisted repository snapshot from optional file_path/symbol_name seeds and return impacted files and symbols with change-risk scores. Seed scope comes from snapshot_id or a triage workspace_evidence_bundle_ref — one of the two is required. direction inbound follows dependents, outbound follows dependencies, both (default) walks both. Use this before simulate_repository or apply_repository to size the blast radius of a change. Read-only against the repository; persists a lookup artifact. Returns seed files/symbols, direct dependencies and dependents, impacted files/symbols, graph nodes and edges, and top_change_risk_files.

Execute a constrained read-only SQL rowset query over authorized datasets. Use this only for wide reports that do not fit the governed exact-query surface. SQL must be a single SELECT/WITH statement over the provided dataset aliases. Returns columns, rows, row_count, truncated, request_id, and latency_ms.

Compare multiple named actions/options and get a ranked recommendation. Use when you need to choose between two or more alternatives with uncertainty. Synchronous deterministic compute; nothing is persisted and no separate rate limit applies. Returns recommended_action, confidence, rationale, and the ranked action list with expected_value and score per action. Which to use: a ranked pick among 2+ named actions with confidence and rationale; use compare for named scenario deltas, or simulate for one scenario's full envelope.

Close the feedback loop: record what actually happened after a decision was made. Sets actual_outcome and computes outcome_delta = actual - expected. Over time this data measures prediction accuracy and reveals systematic biases. Recording updates the persisted decision record in place; repeat calls with the same value converge. Returns decision_id, chosen_action, expected_value, actual_outcome, outcome_delta, and a summary line.

Issue a compatibility credit batch to a quota-governed managed runtime. This exists for non-Algenta managed plans; Algenta editions are unmetered and do not need execution credits. Requires an API-key identity (api_key_identity_required otherwise) and a registered device_id. credits_used reports consumption since the last refresh and defaults to 0. Returns credits_granted, credits_issued_this_month, monthly_limit (0 means unlimited), monthly_remaining, expires_at, refresh_after, and server_time.

Re-pull a saved dataset from its original database, API, or object-store origin using the stored connection and selection, and return the same envelope as connect_data (status, schema_summary, refreshable). Only datasets created from a live connection can refresh — an inline upload fails with not_refreshable (check the refreshable flag in list_data first), and an unknown dataset_id fails with not_found.

Advanced tool. Register a data source and get full schema profiling + join detection. Profiles every column (type, cardinality, fill rate, distribution). Detects formula relationships (A×B≈C) within the source. Detects join keys to every already-registered source automatically. After registration the source is queryable by name via query_data. Safe to call multiple times — re-registration is a no-op if data is unchanged. Registration persists the source profile under the active API key's organization. Returns source_id and the profiled schema with columns, roles, formulas, and detected join keys. Which to use: the advanced path — full column profiling, formula detection, and join-key detection across registered sources (browse the results with list_data); use onboard_dataset for semantic training, or connect_data for refreshable live connections.

Register a real-time trigger that watches a data source for a threshold condition. The engine evaluates the condition automatically when the source data updates; when the threshold is met it runs the simulation template and POSTs a result summary to webhook_url if set. auto_execute=true additionally dispatches the full DecisionPlan to execution_webhook_url after every successful fire (same payload format as execute_decision, through the execution-policy gates). Examples: 'alert me when monthly revenue drops below $80k', 'simulate expansion if Downtown revenue exceeds $200k'. Use fire_trigger to test it immediately and delete_trigger to remove it. Returns trigger_id, name, status, condition, and created_at.

Remove one member from the caller's organization by user_id (find ids with list_team_members). Requires an admin API key. The member is suspended immediately — their API keys stop authenticating at once — and removing the last active owner is refused (last_owner). Removing is idempotent: repeating the call on an already-removed or never-existing user_id returns success with already_absent: true instead of not_found. Returns removed: true with the removed user_id. Use update_team_member_role to change access without removing the member.

Rank caller-supplied document embeddings against a query embedding with a supported deterministic similarity metric (default embeddings.cosine_similarity), most relevant first. This tool does not embed text — call embeddings first to produce the query and document vectors. Every document embedding must share the query's dimension or the call fails with invalid_embedding_dimensions. Read-only and deterministic. Returns ranked items with rank (starting at 1) and score, plus total_documents and returned_documents counts.

Resolve a Hugging Face artifact path through the Algenta compatibility-ring artifact bridge. Defaults to cache-only lookup and never downloads unless local_files_only=false. Use this only for Hugging Face artifact paths; use list_models for the model catalog. Returns the resolution record: status, backend, artifact_backend, resolved_path, cache_root, revision, auth_env_var_used, and auth_configured.

Run the unified Algenta response envelope over one string or a list of independent strings, each processed as its own single-turn request. The envelope returns id, model, status, an output array with one {index, content} item per input, and usage with the total token count. The content per item depends on the model: tokenization models (default text.tokenizer) return the input's tokens and token_count; embedding models return a deterministic vector of dimensions length; provider-backed chat models advertised by list_models return generated text. Use chat_completions for an ordered multi-role transcript and embeddings when you specifically need vectors. Stateless and non-destructive: no conversation state is created, continued, or stored by this tool. An unsupported model id fails with model_not_supported.

Resume a paused agent run by run_id. A run created with approval_mode=auto executes to completion synchronously and returns completed; a manual-mode run moves to requires_approval and still needs approve_agent_run. Resuming anything that is not paused fails with agent_run_invalid_state; an unknown run_id fails with agent_run_not_found. The transition is audit-logged and checkpointed. Returns the updated run resource.

Re-trigger background semantic training for one dataset and return immediately with status and a confirmation message — the build runs asynchronously, so poll get_dataset_status until model_tier reaches 'schema'. Use after schema changes, alias updates, or to force a fresh model build. epochs (default 80, range 5-500) controls training length. An unknown dataset_id fails with not_found; a dataset whose training backend is unavailable fails with semantic_training_unavailable.

Revoke one API key by id (find ids with list_api_keys). The key stops authenticating and the revocation cannot be undone from this tool. This is a credential revocation, not an idempotent delete: an unknown key_id — including an already-revoked one — fails with api_key_not_found (no silent no-op, so a mistyped id can never mask a live credential). Guardrails: revoking the organization's last active key is refused with cannot_revoke_last_key — create a replacement with create_api_key first. Returns key_id with revoked: true.

Revoke one registered device by registration_id (find ids with list_devices), freeing one device slot. The device loses access on its next license refresh. This is a credential revocation, not an idempotent delete: an unknown registration_id — including an already-revoked one — fails with not_found (no silent no-op, so a mistyped id can never masquerade as a successful revocation). Returns revoked: true with the registration_id.

Pick the best unified capability for a natural-language objective and return the route plan: the selected capability, binding, and kind, the authoritative execution_owner, whether approval is required, confidence and reason, plus ordered fallbacks (max_fallbacks, default 3). The optional filters narrow which catalog entries may be selected. Routing never executes anything — feed the selected_capability_id to execute_capability. This picks a capability for an objective; it is not decision analysis — use plan_decision or simulate for a DecisionPlan or full risk envelope instead. Read-only.

Run the repository pipeline and then apply its result in one call, returning the canonical repository envelope for both stages. pipeline takes run_repository_pipeline's arguments and must complete through simulate (the call fails otherwise); apply takes apply_repository's arguments with mode defaulting to patch_only — the write modes still require the simulation gate to pass and write_permission=true. Use the separate stage tools when you need to review the plan or simulation before anything is written.

Run the whole repository-intelligence chain — snapshot, triage, plan, simulate — in one call and return the canonical repository envelope with every stage's response, stage timings, and the ids (snapshot_id, plan_id, simulation_id) the apply step needs. Pass snapshot_id to reuse an existing snapshot or snapshot to create one inline; stop_after halts the chain early (triage skips the LLM planner, plan also skips the deterministic simulate). Signals, triage bounds, model, runs, and seed mirror the standalone stage tools. Use the single-stage tools when you need to inspect or adjust between stages.

Run one simulation request (the same payload shape as simulate) and return the decision envelope fields plus a composite score with its breakdown. The score blends the normalized expected value and one minus the probability of loss; scoring_weights tunes the blend (expected_value default 0.6, downside_risk default 0.4). Use simulate when you need the full envelope without scoring, and compare to rank several scenarios. Synchronous deterministic compute; nothing is persisted. Returns recommended_action, expected_value, probability_of_loss, score, and score_breakdown. Which to use: one scenario reduced to a single weighted number; use simulate for the full envelope without scoring, or compare and recommend to pick between options.

Run a Monte Carlo simulation and get a structured decision recommendation. Use for: quantifying risk in a decision, comparing expected outcomes, getting probability-weighted recommendations. Synchronous deterministic compute governed by the plan's per-minute rate limit and monthly quota (429 on excess); the run is recorded asynchronously and appears in list_runs. Returns the decision envelope: recommended_action, expected_value, probability_of_loss, confidence, percentiles, and run metadata (run_id, execution_ms, scenarios_run). Which to use: the full decision-envelope analysis on ONE scenario; use compare for named scenario bake-offs, recommend for a ranked pick among actions, score for a single weighted number, plan_decision for just the plan summary, or product_decision for plain-English inputs.

Score the patch risk of a stored repository DecisionPlan with the deterministic simulation engine and return the gated DecisionEnvelope whose apply gate apply_repository checks. snapshot_id is resolved from the plan when omitted. runs pins the scenario count (100 or more; omit for the complexity-adaptive count) and seed (default 42) makes repeated calls reproducible — no LLM is involved in this stage. Call create_repository_decision_plan first; use simulate_repository_patch instead for a patch that has no stored plan.

Simulate the risk of an in-flight unified diff against one persisted snapshot and return the canonical repository envelope — without creating a stored decision plan. Use this to verify a working-tree patch mid-run; use simulate_repository when a stored DecisionPlan already exists. Deterministic; no LLM is involved. Returns the gated DecisionEnvelope including the apply gate verdict.

Submit a long-running async simulation job. Use for n_simulations > 500,000 or when you need a callback. variables is an array of {name, low, high} objects — one uniform range per variable; objective defaults to maximize and n_simulations defaults to 1000000. Returns a job_id — poll with get_job_status. Submitting persists the job under the active API key's organization; when callback_url is set, completion is delivered to it by outbound webhook.

Run a health test on one capability binding and return the outcome. Pass binding_id to test a saved binding, or a full inline definition (provider_id, profile_id, config, ...) to preview-test one that was never saved — nothing is persisted in the preview form. Use this after create_capability_binding or update, before routing traffic to the binding.

Run a real connectivity test against one saved connector's stored config and persist the outcome as its live or error status with last_tested_at. This opens an actual connection to the source. Use preview_test_connector for an unsaved inline definition, and browse_connector once the connector is live. Returns success, message, latency_ms, status, error_type, and recoverable.

Send one real test webhook payload (event webhook.test with a sample message) to a callback URL and return the delivery result. This makes an actual outbound HTTP POST from the Algenta API to the given URL, with no retries. Use it to verify a receiver before wiring callback_url into submit_job or register_trigger. Returns success, the receiver's HTTP status_code, and a message.

Tokenize UTF-8 text into individual tokens with a supported deterministic Algenta tokenizer model (default text.tokenizer; call list_models for every supported model id). Use this when you need the token strings themselves; call count_tokens when you only need the number. Read-only and deterministic: the same input and model always return the same tokens, and nothing is stored. Returns the resolved model id, its tokenizer_kind, the tokens array, and token_count. An unsupported model id fails with model_not_supported.

Condense one repository snapshot into a bounded workspace evidence bundle for the planner: ranked suspect files and symbols with scored, budget-capped snippets. signals seeds the search — pass issue_text, diagnostics, failing_tests, changed_files, and/or workspace_context. The returned workspace_evidence_bundle_ref is the required input to create_repository_decision_plan; use run_repository_pipeline to chain both stages in one call. Read-only against the repository; persists the bundle artifact. Returns suspect_files, suspect_symbols, evidence_items, and token reduction stats.

Partially update one saved connector: only the supplied fields change. Passing a new config replaces the encrypted credentials and resets the connector to untested, so call test_connector again afterwards. Requires manage permission on the connector (access_scope_denied otherwise) and at least one field; an unknown id fails with not_found. Returns the updated connector. Use preview_test_connector to validate a new config before applying it here.

Partially update the organization's autonomous execution policy: only the fields supplied change, the rest keep their values. min_confidence blocks decisions below that confidence, risk_floor blocks decisions whose worst-case (p5) loss exceeds it, require_calibration makes auto-execution wait for enough recorded outcomes, and allow_reexecution is the idempotency gate. Changes take effect immediately, are recorded in the audit log, and write a new policy snapshot (see list_execution_policy_snapshots). Read the current values first with get_execution_policy. Returns the full updated policy.

Update the current user's name and/or the organization name for the active API key; only the supplied fields change. Renaming the organization requires an admin or owner key (access_scope_denied otherwise), and the key must be linked to a user (user_not_found for service keys). At least one of name or org_name is required. Returns the updated identity; read it first with get_me.

Change one organization member's role by user_id (find ids with list_team_members). Requires an admin API key. Guardrails: you cannot change your own role (self_role_change_forbidden), only an owner can grant the owner role (owner_grant_forbidden), and demoting the last active owner is refused (last_owner). An unknown user_id fails with not_found. Returns the updated user_id and a confirmation message. Use remove_team_member to take the member out of the organization instead.