The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
provider
Native TinyAgents ChatModel construction plus cloud/local inference policy,
auth, error taxonomy, and RPC helpers for every chat-model transport OpenHuman
supports. This is the chat-model half of OpenHuman's pluggable-engine story
(see docs/gitbooks/en/developing/engines.md):
the model a turn runs on is chosen by config, not by code change. Was
previously providers/ (pre-consolidation single-crate layout); see
../README.md for how this fits into the wider inference domain.
Public surface
- Factory (
factory.rs+factory/:routing.rs(RouteSourcehost adapter),subprocess_providers.rs,access_gates.rs,chat_model.rs,cloud_slug.rs,credentials.rs,local_runtime.rs,managed_backend.rs),create_chat_model,create_chat_model_from_string[_with_model_id],create_chat_model_with_model_id,provider_for_role,probe_inference_readiness,BYOK_INCOMPLETE_SENTINEL. Parses the provider-string grammar (openhuman,cloud,ollama:<model>,lmstudio:<model>,mlx:<model>,omlx:<model>,local-openai:<model>,claude_agent_sdk[:<model>],claude-code:<model>,<slug>:<model>[@<temp>]) and applies the BYOK sentinel, Privacy-ModeLocalOnly(enforce_local_only_inference), and managed-session (verify_session_active) gates before building a model. See docs/gitbooks/en/features/model-routing/local-and-byok-models.md for the user-facing side of local and BYOK routing. - Models: providers are registered chat models
(
inference::registry); TinyHumans supplies its hosted model vocabulary and request policy. OpenAI-compatible and Anthropic builders live intinyinference_llm::providersand are called directly. - Provider response and usage types:
ChatResponse,ProviderDelta, andBilledUsagecome fromtinyinference_llm; the hosted turn output cap lives with the OpenHuman turn policy. - Error classifiers: reusable classifiers live in
tinyinference_llm::classification; this directory retains OpenHuman managed-backend and telemetry policy.
Transports
| Transport | File | Provider-string prefix |
|---|---|---|
| Managed OpenHuman backend | registered chat model (openhuman-tinyhumans, src/inference/) |
openhuman / cloud (session JWT + billing metadata) |
| OpenAI-compatible (BYOK cloud slugs, local runtimes) | tinyinference_llm::providers::openai |
<slug>:<model>, ollama:<model>, lmstudio:<model>, mlx:<model>, omlx:<model>, local-openai:<model> |
| Anthropic Messages API (prompt caching) | tinyinference_llm::providers::anthropic |
<slug>:<model> whose endpoint is the first-party Messages API and native tool calling is on |
| Codex OAuth / Responses API | Host OAuth selection plus tinyinference_llm::providers::openai::codex routing metadata |
the openai cloud slug once Codex OAuth tokens exist |
| Claude Agent SDK subprocess | tinyagents_harness::providers::claude_agent_sdk |
claude_agent_sdk / claude_agent_sdk:<model> |
| Claude Code CLI subprocess | tinyagents_harness::providers::claude_code (see the Claude Code provider guide) |
claude-code:<model> |
Calls into
tinyinference_llm::model::ChatModel(vendor/tinyagents/vendor/tinyinference): the trait every transport implements.crate::config: cloud-provider schema (AuthStyle, slug reservation),Config::claude_agent_sdk, abstract tier model constants.crate::security::credentials: auth-profile store for BYOK keys and OAuth tokens.crate::agent::tinyagents::host: explicit run explicit run-thread plumbing consumed while building a managed model.crate::security::live_policy+crate::security::egress: Privacy-ModeLocalOnlyrefusal andEgressDescriptoremission at the factory chokepoint (factory/access_gates.rs).crate::inference::host_runtime:profile::is_local_provider_string, Ollama / LM Studio base-url resolution for local provider strings.crate::inference::auth_error_registry: surfaces OpenHuman per-provider auth errors back to the UI.crate::core::bus(BUS.publish) /crate::core::events::DomainEvent:http_error.rs::publish_backend_session_expiredand the registered managed model (openhuman-tinyhumans) publishDomainEvent::SessionExpiredwhen the managed backend reports an auth failure, so the credentials layer can clear/refresh the session;http_error.rsalso publishesDomainEvent::ProviderApiKeyRejectedthe first time a BYO key is rejected.crate::mcp::server::local(via the vendoredclaude_codedriver), the Claude Code provider points the sandboxedclaudesubprocess at the in-process MCP server so it can reach OpenHuman's memory/tools over loopback without the MCP server inheriting CC's OS jail.
Called by
grep -rn 'inference::provider::' crates/openhuman-core/src shows the main
consumers: the agent harness (agent/session_host/builder/factory.rs,
agent/session_host/runtime*.rs, agent/subagent_host/ops/*,
agent/tinyagents/host/model_resolver.rs), web_chat/session.rs and
web_chat/web_errors/ (classify.rs, budget.rs, retry.rs, timeout.rs,
backend_error_code.rs, provider_detail.rs, response_predicates.rs), voice/factory/{helpers,mod}.rs,
inference/ops.rs / inference/schemas/ /
inference/http/backend.rs,
flows/tinyflows/caps/{llm,prompt,agent}.rs, cron/scheduler/failure_classification.rs
(is_budget_exhausted_message), and threads/ops/usage.rs (UsageInfo).
Sub-modules
http_error.rsowns provider HTTP error classification, Sentry routing, andapi_error;models/owns model catalog listing and local-runtime entries. Delegation'sProviderRuntimeOptionslive withtools/delegate.rs.- Claude Code CLI provider:
tinyagents_harness::providers::claude_codeinvendor/tinyagents; see the Claude Code provider guide.
Tests
factory_tests.rs,factory_crate_native_tests.rs,factory_egress_fallback_tests.rs,factory_route_resolution_tests.rs,factory_test_provider_override_tests.rs: provider-string parsing, access gates, and model construction.http_error_tests.rs,models/mod_tests.rs, and provider model-listing tests: error classification and model listing.- Claude Code provider tests live with the provider in
vendor/tinyagents/crates/tinyagents-harness/src/providers/claude_code/. http_error_tests.rs: generic host error handling with provider ownership from the installed classifier. builder and Codex tests live in TinyInference and TinyAgents.
Local runtime chat providers (Ollama, LM Studio, MLX, oMLX, local-openai) and caller-authenticated Claude Code/Agent SDK subprocesses do not require an OpenHuman session, including when used by channel agents. Claude subprocesses still contact their external provider and remain subject to the LocalOnly privacy gate. Managed inference and custom cloud routes retain their session checks.
Flow readiness checks the initial provider role used by a known harness agent, including node model overrides, custom registry pins, and the session builder's configured default. It shares the builder's role resolver before deciding whether a session is required. A separate summarization provider does not determine the session requirement for that agent.