Inferoa
An inference-native tokenmaxxing agent harness that keeps engineering loops cache-friendly and self-correcting until the work is proven.
- Source repo
- agentic-in/inferoa
- Stars
- ★ 566
- Last updated
- 4mo ago
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 53/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changes
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer runs /loop to keep improving a repository until the improvements are proven by tests.
- Not a fit if
- Users who just want one-shot Q&A without long recursive loops
- Teams unwilling to operate vLLM or self-hosted inference paths
- Users who need stable releases rather than the @dev channel
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
Inferoa is an open-source (Apache-2.0) command-line agent harness from the agentic-in organization built around Loop Engineering: give the model an objective, feedback, verification, memory, and tools, and let it self-correct until the work is proven. It treats every loop as an inference workload, protecting reusable prefix caches through prompt epochs, deterministic tool schemas, and bounded system sections while exposing token pressure. Its core interface is a TUI where /loop runs recursive long-horizon loops that carry objectives, verification evidence, decisions, and recovery across loop tasks. It is built on the vLLM ecosystem and can choose model paths by cost, safety, privacy, capability, and session pressure across the vLLM Semantic Router, vLLM Engine, vLLM Omni, and self-hosted paths. Installation is a global npm package (inferoa@dev), making this a developer-facing terminal tool.
After inferoa setup configures the endpoint, model, vault-backed API key, and Omni configuration, inferoa opens the TUI; passing a prompt starts a session and submits it as the first user turn (e.g. inferoa "Inspect this repository and list the test entrypoints."). /loop starts a recursive long-horizon loop that maintains the objective, attempts, verification evidence, decisions, and completion evidence across loop tasks; /plan turns ambiguous scope into an inspectable plan; /tokenmaxxing shows token and cost pressure across prefix-cache reuse, context savings, recent turn usage, and model-selection pressure. During runs it applies memory and context control (compression, summaries, graph-shaped repo context, bounded history and tool output) and uses CodeGraph and RTK to select evidence and shrink mutable context. --print runs a single non-interactive request.
- A developer runs /loop to keep improving a repository until the improvements are proven by tests.
- A researcher uses Loop Research to run long-horizon research loops with verification metrics.
- A self-hosting team keeps model routing inside their own vLLM infrastructure while gaining prefix-cache discipline.
- An engineer uses /plan to turn ambiguous requirements into an inspectable plan before execution.
- An operator watches token and cost pressure across multimodal and multi-model paths via /tokenmaxxing.
How do you install or deploy this agent?
Install globally via npm. The @dev dist-tag tracks the latest build published from main; the npm latest dist-tag is reserved for stable releases:
npm install -g inferoa@devThen run the setup wizard to configure the endpoint, model, vault-backed API key, and Omni configuration:
inferoa setupSupported endpoint types and credential formats are documented at https://inferoa.agentic-in.ai/docs/quickstart.
How do you use this agent?
Open the TUI:
inferoaStart a session with a first user turn:
inferoa "Inspect this repository and list the test entrypoints."Start a recursive long-horizon loop from inside the TUI:
/loop Improve this repository and prove it with tests.Run a single non-interactive request:
inferoa --print "Summarize the README in one paragraph."Other core commands: /plan (inspectable plan before execution) and /tokenmaxxing (token and cost pressure).
What are this agent's strengths and limitations?
- Loop Engineering at the core: /loop retains objectives, verification evidence, decisions, and recovery across tasks instead of stopping after one answer.
- Prefix-cache discipline (prompt epochs, deterministic tool schemas, bounded system sections) directly reduces multi-turn inference cost.
- Deep vLLM ecosystem integration (Engine, Omni, Semantic Router) with self-hosted model paths and routing by cost, safety, privacy, and capability.
- Independent feedback surfaces (plans, tests, tool results, research metrics, completion evidence) give the loop something concrete to improve against.
- Core value is tied to the vLLM ecosystem; non-vLLM users must adapt or forgo routing and caching capabilities.
- No stable npm channel is documented yet — the documented install path is the @dev build channel.
- Multi-step setup (endpoint, model, vault-backed API key, Omni) raises onboarding cost versus single-command tools.
- Loop reliability depends on the model and test evidence; the README offers no benchmarks substantiating tokenmaxxing gains.
How does this agent compare with similar options?
The README positions Inferoa at the Loop Engineering / Agent Harness layer, distinct from one-shot coding assistants: /loop persists verification evidence and decisions across tasks, grounded in the vLLM ecosystem (Engine, Semantic Router, Omni) plus CodeGraph and RTK for context optimization.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Inferoa This agent | 53 · Major gaps | CLIFree + model costs | ★ 566 | 4mo ago | TypeScript | — |
| Entroly | 68 · Some gaps | CLIFree + model costs | ★ 480 | 7d ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| CorvinOS | 67 · Some gaps | CLIFree + model costs | ★ 20 | today | Python | Codex · Claude Code |
| Pilotfish Multi-Model Orchestration | 81 · Good | Agent plugin / skillFree + model costs | ★ 711 | 16d ago | Python | Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Permission modes, managed_installs: ask, and web_search: off appear in test configs, but no enforcement code is visible, so least_privilege is 1; the clarify tool's cancel semantics and install confirmation support user_confirmation 2; the cache footer exposing usage and hit rates supports data_flow_transparency 2; only a 'vault-backed API key' mention exists with no credential-handling detail, so sensitive_data_handling is 1; the dependency set is tiny and publishing uses OIDC/restricted tokens, giving dependency_security 2; the loop promises autonomous repo edits while rollback and external-effect controls are absent from visible files, so external_effects and rollback each drop to 1; acknowledgements and ecosystem links make source_attribution 2.
Tests are internally consistent and cover acceptance preflight plus tool failure paths, so self_consistency is 2; CI runs npm test on Node 24 with pinned deps, giving dependency_availability 2; structured error codes (clarify_unavailable, clarify_cancelled) are asserted in tests, so failure_messages is 2; the acceptance test enforces a direct vLLM endpoint, and any user-facing consequence of that constraint is undocumented — noted rather than scored down further.
The README targets Loop Engineering practitioners and provides /loop, /plan, and --print scenarios, so audience_and_scenarios is 2; boundary claims like 'until the work is proven' lack reviewable criteria in the visible files, so capability_boundaries is 1; slash command semantics are precise, so trigger_precision is 2; engines >=24 and rejection of external-only providers give environment_fit 2.
Docs-site structure is well linked, so information_architecture is 2; install and quickstart are concrete, but only the @dev tag is recommended and there is no troubleshooting, so install_notes is 2; naming (inferoa, tokenmaxxing, /loop) is consistent, so naming_stability is 2; examples exist but there is no FAQ, so examples_and_faq is 1; no visible file lists limitations, so known_limitations is 1; LICENSE is full Apache-2.0 and matches package.json, so license is 3; version 0.14.19, dist-tag policy, and a release-coordinates script exist but there is no CHANGELOG, so versioning_changelog is 2; a CI publish pipeline exists but no maintainer or contribution guidance is visible, so maintenance_responsibility is 1.
The TUI, cache footer, and --print mode indicate polished output usability, so output_usability is 2; the tokenmaxxing/prefix-cache angle is differentiated, giving marginal_value 2; cost benefit rests on promotional narrative rather than visible evidence such as savings data, so cost_benefit is 1.
Core selling points (long-horizon loops, proven completion, cache discipline) have no traceable source code or data in the provided files, so claim_traceability is 1; all corroboration comes from this repository alone and external links/GIFs cannot be statically verified, so cross_source_corroboration is 1; the README's marketing tone and the tests' factual content are not explicitly separated, so fact_inference_separation is 1.
- Publisher identity is unverified; confirm repository ownership and supply-chain integrity before adoption.
- Core capabilities (long-horizon loops, 'proven' completion, cache savings) lack traceable evidence in the provided files and remain unverified claims.
- Loop mode autonomously edits codebases by design, while rollback and external-effect controls are not demonstrated; trial in a sandbox first.
- The recommended install target is the rolling @dev dist-tag rather than a stable release, weakening version reproducibility.
- Sensitive-data handling rests on a single 'vault-backed API key' mention; verify the key storage and transport implementation.