Manifold Grace

RESEARCH · OPEN TECHNOLOGY · PUBLIC-INTEREST AI

Free & Low-Cost AI Infrastructure Register

Developed as a cost-first reference for people building their own AI systems. The register distinguishes truly local software, limited free cloud access, commercially hosted services and existing Manifold Grace infrastructure. It supports the AI Public Adviser and the wider AI systems roadmap.

Audit status: 10 November 2026. 63 tool entries are documented; these are candidates, not installed products. No new accounts, API keys, deployments or external provider integrations have been created by this register. Free-tier usage limits and licences can change. Not all free software performs like paid frontier services, and computer hardware still costs money.

Three ways to begin

Entirely local

Ollama or llama.cpp + Open WebUI + SQLite + locally stored documents.

No model API subscription; requires a capable computer and locally licensed model.

Cloud free allocation

Gemini or GroqCloud + Cloudflare Workers AI + Supabase, subject to current quotas.

Useful when local hardware is insufficient; credentials and privacy conditions require attention.

Research-first hybrid

Existing GitHub Pages and GitHub Actions + permitted feeds + source verification + SQLite/Qdrant + model fallback.

Preserves a free public information layer and avoids premature dependency on one provider.

Searchable tool and integration inventory

Expanded 2026-11-10: 36 further entries added to the original 27. New candidates include free hosted inference, local runtime alternatives, evidence ingestion, document parsing, workflow builders, evaluation, security, hosting, accessible speech and data management. No additional software has been installed or connected. Free tiers are limited and can change.

Read the overview first, then filter by technology category or candidate. The full table can be scrolled horizontally without truncating records. Each tool now includes a risk field; these are risk categories, not accusations or independently verified current weaknesses in any named organisation.

63 entries displayed

No.Tool and official linkFunctionAvailabilityAccount requirementFree limits and real costsPotential useNext validationSecurity, geopolitical and information risks
1OllamaModel runtimeLocal open-sourceNo account required for local modelsNo recurring software licence; hardware costsLocal execution; simple model servingEvaluateEndpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
2llama.cppModel runtimeLocal open-sourceNoFree code; hardware costsQuantised models, CPU/GPU and local APIEvaluateEndpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
3JanModel runtimeLocal open-sourceNo for local modelsFree desktop software; hardware costsLocal AI desktop interfaceAlternativeEndpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
4LM StudioModel runtimeFree desktop softwareNo for local useFree software, not all components open sourceLocal model exploration and serverAlternativeEndpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
5Open WebUIAI interfaceFree self-hosted editionNo hosted signup for self-hostSelf-managed storage and maintenanceUser interface, RAG and local/cloud connectorsEvaluateExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
6Google AI Studio / Gemini APIHosted AILimited free API tierGoogle account + API credentialSelected models free; usage limits and free-tier data useCloud reasoning benchmark or low-volume fallbackEvaluateExternal processing, retention, lawful access and provider outages; verify controls before handling sensitive data
7GroqCloudHosted AILimited free serviceAccount + API keyRequests/tokens per day vary by modelFast inference alternativesEvaluateExternal processing, retention, lawful access and provider outages; verify controls before handling sensitive data
8Cloudflare Workers AIHosted AIDaily free allocationCloudflare account10,000 free neurons daily; supported models onlySmall cloud inference and backend functionsEvaluateExternal processing, retention, lawful access and provider outages; verify controls before handling sensitive data
9Hugging FaceHosted AIProvider-dependent credit termsAccountCredit availability and eligibility must be verified; not unlimitedModel discovery and trial provider accessResearch onlyExternal processing, retention, lawful access and provider outages; verify controls before handling sensitive data
10LiteLLMModel routingOpen-source library/gatewayNo for local gatewayThird-party inference still billedProvider-neutral switching, logging and limitsEvaluatePrivilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; Routing to alternate providers can change data jurisdiction and usage conditions; verify controls before handling sensitive data
11Crawl4AIResearch retrievalApache-licensed self-host + hosted free tierNo locally; account/key for cloudLocal compute; hosted 1,000 monthly credits advertisedExtract accessible public sources for RAGEvaluateExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Publisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
12GPT ResearcherResearch agentOpen-source implementationNo for code; some providers require keysSearch and inference may costResearch planning and multi-source synthesisEvaluatePublisher terms, manipulated pages, malicious content, unsupported citations and tracking; Privilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
13GDELTInformation feedsPublic research dataVaries by endpointAccess policies, volume and licensing varyGlobal news/event indicatorsEvaluatePublisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
14RSS and official feedsInformation feedsGenerally public feedsTypically none for public RSSPublisher rights and access rules varyLow-cost dated primary-source ingestEvaluatePublisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
15SupabaseStructured databaseFree hosted development tierAccount + project500MB database and other caps; inactivity pausesEvidence register, audit trail and metadataEvaluateExternal processing, retention, lawful access and provider outages; Misconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
16SQLiteStructured databasePublic-domain libraryNoDisk capacity and operational maintenanceSmall durable local claim storeEvaluateMisconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
17QdrantVector searchFree local/open-source or hosted trial clusterNo local; account for cloudHosted free 1GB RAM, 4GB disk; inactivity policySemantic source retrievalEvaluateExternal processing, retention, lawful access and provider outages; Publisher terms, manipulated pages, malicious content, unsupported citations and tracking; Misconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
18ChromaVector searchOpen-source localNoLocal compute/storageLocal prototyping and retrievalAlternativePublisher terms, manipulated pages, malicious content, unsupported citations and tracking; Misconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
19pytestEvaluationOpen-sourceNoCompute and maintenanceDeterministic correctness testsAvailablePrompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
20RagasEvaluationOpen-source frameworkNo for codeModel evaluations may require paid inferenceEvaluate RAG groundedness and answer qualityEvaluatePrompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
21GitHub / GitHub ActionsCoding and deploymentFree public-repository automationGitHub account already connectedPublic actions free on standard runners; secrets must be protectedCI/CD, scheduled public metadata sync, source archiveExistingExternal processing, retention, lawful access and provider outages; Supply-chain compromise, CI secrets, maintainer permissions and update integrity; Public repositories and CI logs can disclose source code or secrets; verify controls before handling sensitive data
22GitHub PagesStatic hostingFree public siteGitHub account already connectedStatic only; cannot safely expose private API keysPublic research pages and approved demosExistingPublic repositories and CI logs can disclose source code or secrets; verify controls before handling sensitive data
23GraphifyCode graphOpen-source code intelligenceNo for local versionCompute costsMap code dependencies and retrieve related symbolsExistingLicensing, data custody, permissions, dependencies and provider continuity; verify controls before handling sensitive data
24n8nWorkflow automationSource-available self-hosted editionNo SaaS signup if self-hostHosting costs; licensing conditions applyScheduled retrieval, review queues and approvalsEvaluateExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Privilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
25LangGraphWorkflow engineOpen-source frameworkNo for codeModels and hosting separateStateful, auditable multi-step agent executionEvaluatePrivilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
26OpenTelemetryMonitoringOpen-sourceNo for self-hostStorage and service costsTraces, measurements, performance logsEvaluatePrompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
27Google TimesFMForecastingOpen-weight research modelNo for public code accessHardware/model licence and data requirementsTime-series forecasts and baseline comparisonsResearchIncorrect forecasts, missing data, uncertain calibration and consequential over-reliance; verify controls before handling sensitive data
28LocalAILocal inferenceOpen-source self-hostedNo local accountModel download, compute and maintenanceOpenAI-compatible local endpoints, multiple model typesCompare endpoint security, VRAM and throughputExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
29vLLMLocal inferenceOpen-source self-hostedNoSubstantial GPU may be required; electricity/hostingHigh-throughput inference serverBenchmark hardware requirements and concurrent requestsExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
30SGLangLocal inferenceOpen-source self-hostedNoGPU and deployment costsFast structured generation and model servingCheck compatible model and hardwareExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
31MLX LMLocal inferenceOpen-source localNoRequires compatible Apple Silicon hardwareInference and adaptation on MacsCompare local throughput and memoryExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
32AirLLMLocal inferenceOpen-source localNoDisk I/O can make large-model inference slowLayer-streamed large-model inferenceMeasure tokens per second, disk usage and qualityExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
33ColibrìLocal inferenceOpen-source candidateNo for sourceVery low speed possible on laptop RAMExperimental large-model CPU inferenceConfirm published speed on real hardware and reproduce performanceExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
34AnythingLLMLocal interfaceOpen-source/self-hostedNo local accountModel, embeddings and machine costsPrivate document Q&A workspaceCheck source provenance and local data retentionExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
35LibreChatLocal interfaceOpen-source self-hostedNo provider login for self-hostDatabase, model gateway and server maintenanceMulti-provider chat UI, controls and workflowsReview permissions and integrationsExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; verify controls before handling sensitive data
36OpenRouter free modelsHosted inferenceFree tier with request capsAccount and API key50 requests/day on free account; models/provider data policies varyCompare multiple free models via one gatewayRecord free model selection, latency and data termsExternal processing, retention, lawful access and provider outages; Routing to alternate providers can change data jurisdiction and usage conditions; verify controls before handling sensitive data
37Cerebras InferenceHosted inferenceIntroductory free creditAccount and API keyOfficial pricing currently advertises $5 starter credit, not an unlimited tierShort benchmark of high-speed inferenceCredit exhaustion, model availability and privacyExternal processing, retention, lawful access and provider outages; verify controls before handling sensitive data
38NVIDIA NIM APIsHosted inferenceDeveloper access or limited trialNVIDIA account and credentialProvider quota and eligibility vary; production usage may costModel API comparisonVerify account-specific terms and limitsExternal processing, retention, lawful access and provider outages; verify controls before handling sensitive data
39SearXNGSearch and ingestionOpen-source self-hostedNo local signupSearch engines may block or restrict automated queriesMetasearch discovery and source diversityRespect provider access terms; measure recallExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Publisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
40TrafilaturaSearch and ingestionOpen-source localNoWeb access subject to publisher restrictionsClean article text extraction and metadataPreserve original URLs and datesPublisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
41ScrapySearch and ingestionOpen-source localNoRobots/access policies, compute and storage costsPermissioned structured extractionTest crawl politeness and rate limitsPublisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
42Apache TikaSearch and ingestionOpen-source localNoLocal compute, file-processing security risksExtract text and metadata from public documentsSandbox uploaded or untrusted filesPublisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
43FeedparserSearch and ingestionOpen-source localNoPublisher feeds vary in availability and termsRSS/Atom ingest of dated news itemsTrack original URL, author, update and correctionPublisher terms, manipulated pages, malicious content, unsupported citations and tracking; verify controls before handling sensitive data
44DifyOrchestrationSource-available self-hostedNo SaaS signup for self-hostCheck licence/restrictions; model and hosting costsVisual AI workflows and retrievalAssess licence, permissions and data retentionExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Privilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
45FlowiseOrchestrationOpen-source self-hostedNo for local deploymentModel calls and hosting cost extraVisual workflow and agent designBenchmark logs, approvals and securityExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Privilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
46HaystackOrchestrationOpen-source frameworkNoModels and infrastructure separateComposable evidence-first RAG pipelinesMeasure citation fidelity and retrieval completenessPrivilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
47PydanticAIOrchestrationOpen-source frameworkNo for libraryUnderlying model APIs may costType-checked structured outputs and agent toolsTest schema validation and failure modesPrivilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
48LlamaIndexOrchestrationOpen-source framework componentsNo for sourceCloud and model services may chargeDocument indexing, retrieval and tool integrationsCheck package licence and citationsPrivilege escalation, prompt injection, unexpected tool calls and cross-provider data forwarding; verify controls before handling sensitive data
49LangfuseObservabilityOpen-source self-hostedNo for self-hostStorage, hosting and telemetry privacy costsTrace model calls and prompt versionsProtect confidential prompts and personal dataExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Prompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
50Arize PhoenixObservabilityOpen-source self-hostedNoCompute and evaluator-model costsTrace/evaluate answers and evidence retrievalAudit evaluation reproducibilityExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Prompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
51MLflowObservabilityOpen-source self-hostedNoDatabase and storage hostingExperiment and model version trackingDocument data and model provenanceExternal processing, retention, lawful access and provider outages; Endpoint security, model provenance, unreviewed packages and local data exposure; Prompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
52PromptfooEvaluationOpen-source local CLINo for local testingTarget API usage and some security tests may costRegression evaluation and safe red-team testsTest hallucinations and prompt injection using synthetic casesPrompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
53DeepEvalEvaluationOpen-source localNo for core libraryEvaluator API calls may costLLM output evaluation with test fixturesValidate judging metrics with humansPrompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
54Evidently AIEvaluationOpen-source localNo local accountMonitoring storage and infrastructure separateData quality, drift and monitoringCompare metrics to real-world errorsPrompt/trace leakage, model-judge error, sensitive logs and telemetry retention; verify controls before handling sensitive data
55NVIDIA SkillSpectorSecurityOpen-source scannerNo for local toolMay not detect all malicious behaviourScan third-party AI skills before useManual code and permission review still requiredSupply-chain compromise, CI secrets, maintainer permissions and update integrity; Static checks miss some malicious behaviour; independent code review remains required; verify controls before handling sensitive data
56OWASP GenAI Security ProjectSecurityOpen security guidanceNoAssessment and mitigations require workThreat models and AI application checklistsMap controls to own agent privilegesSupply-chain compromise, CI secrets, maintainer permissions and update integrity; verify controls before handling sensitive data
57Cloudflare WorkersDeploymentFree usage tierCloudflare accountFree request/CPU limits; no unattended unlimited inferenceServer-side API gateway to protect secretsApply quotas, access rules and budget limitsExternal processing, retention, lawful access and provider outages; Supply-chain compromise, CI secrets, maintainer permissions and update integrity; verify controls before handling sensitive data
58Hugging Face SpacesDeploymentFree basic CPU hosting tierHugging Face accountGPU/higher memory are paid or quota-controlledPublic demo front-end or limited CPU prototypeCheck uptime, data privacy and limitsExternal processing, retention, lawful access and provider outages; Supply-chain compromise, CI secrets, maintainer permissions and update integrity; verify controls before handling sensitive data
59DuckDBData managementOpen-source embedded analyticsNoLocal disk, compute and backupsAnalyse large sets of news events and indicatorsTest reproducible data transformsMisconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
60PostgreSQLData managementOpen-source databaseNo for localHosting, backups and admin timeDurable structured evidence and audit recordsAccess controls and schema validationMisconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
61FAISSData managementOpen-source localNoLocal vector-index memory and CPU/GPU costsSemantic similarity and duplicate detectionTest false matches and attributionMisconfigured access, backups, cross-user leakage and retention; verify controls before handling sensitive data
62Whisper.cppSpeech and accessibilityOpen-source localNoModel licences, compute and recording permissionsTranscribe approved audio/video sourcesCheck language accuracy, privacy and consentRecording consent, transcript confidentiality and voice model licences; verify controls before handling sensitive data
63PiperSpeech and accessibilityOpen-source localNoGPL-3.0 implementation; voice model licences and local computeText-to-speech for accessibilityActive successor to archived rhasspy/piper; review GPL requirements and voicesRecording consent, transcript confidentiality and voice model licences; verify controls before handling sensitive data

Security and trust must be assessed before integration

Every technology is listed with foreseeable technical and information-handling risks. This does not imply that a named provider or its staff have acted improperly. Nationality, employee background or geography alone does not establish intent; assess published data-handling policies, jurisdiction, access controls, incident response, audits and verifiable contractual terms. Open-source software may reduce dependence on a service provider but does not eliminate malicious packages, inaccurate models or local security problems.

Data handling default: keep sensitive or unpublished information local where feasible, restrict tokens to the minimum permissions, avoid embedding credentials in publicly deployed pages, document third-party processing and test a supplier's revocation and deletion process. Never send personal case files or confidential research to a free model API solely because it is free.

Review 20 documented cases of software compromise, data exposure, third-party risk and governance failure.

Risk review questions and evidence requirements
  • Who processes, stores, trains on, reviews, logs and deletes prompts, attachments, model outputs and metadata?
  • What contractual protections, permitted subprocessors, processing locations and lawful disclosure requirements apply?
  • Can access to accounts, API tokens, support functions and agent tools be independently controlled, logged and revoked?
  • Are the developer, model weights, software dependencies and update processes auditable?
  • Can the project be exported or moved if a company changes ownership, licensing or access?
  • What happens after an incident: notifications, corrections, backups, independent investigations and proven recovery?

Registration and integration sequence

Stage 0 — No new accounts needed

Retain existing GitHub infrastructure; audit current pages and existing Graphify CI; prepare source schemas, synthetic evaluation cases and a consent-based test dataset. Experiment with open-source local tools on suitable hardware before creating subscriptions.

Stage 1 — Optional account-based trials

Evaluate one free model API, one extraction provider only if needed, and one cloud storage/search service only if the existing environment is insufficient. Each connection must have documented purpose, allowed data, quota, provider terms, key custody and a revocation procedure. Do not publish API keys on GitHub Pages.

Stage 2 — Integration pilot

Implement a server-side model gateway, local source cache, explicit citation record, independent-source deduplication, request budget caps, failure fallback, data retention limits and human review for any consequential action. Compare providers using identical prompts and datasets.

Stage 3 — Public demonstration

Deploy only once evidence coverage, cost per request, latency, factual accuracy, uncertainty calibration, accessibility and responsible-use controls have been measured. Public-safety information remains accessible independent of commercial subscriptions.

How to determine whether a free tool is nearly as capable

A free alternative can only be described as comparable after task-specific evaluation. For the AI Public Adviser, measure source coverage, claim attribution accuracy, factual error rate, contradiction detection, refusal to guess when evidence is inadequate, response latency, privacy safeguards and total monthly operating cost.

Recommended comparison: local model vs free hosted model vs a stronger benchmark model, using the same set of public questions, evidence passages and blinded scoring. Publish individual failure cases and date all results.

Official reference sources

Research hub · AI demonstrations · Societal resilience programme · Productivity log