Big AI owns the model, the servers, and every line you send them. With igni, you own everything. The agent, the model and the harness cost nothing; you pay only to rent the GPU it runs on — roughly $2.50 per developer per month, because one NVIDIA GH200 serves up to 200 developers.
igni 1.0 27B is a fine-tune of Qwen 3.8 27B, trained for tool-calling, diffs and code review rather than chat. It is small enough to serve 200 developers from a single GH200, so the bill is GPU rental rather than a per-seat licence. Deploy it in our racks, your cloud account, your own datacenter, or fully air-gapped — your code never leaves your network, and there is a full audit trail.
An agent with CodeIndex answers questions about a codebase correctly 90.5% of the time against 69.5% with ripgrep, at 176k tokens per answer against 436k. Measured over 82 questions across 18 real repositories in Python, TypeScript and JavaScript, each asked three times per arm, with the correct answers computed mechanically from source rather than from the index. It beats or ties grep on 15 of 18 repositories. Grep ties it on "which files import this one", because that answer is a literal line of text; the index pulls ahead where the answer is a count across the whole repository.
Every push re-indexes incrementally and every pull request or merge request is reviewed against the index, with findings posted as comments before a human opens it — security risks, swallowed errors, and untested modules the change depends on. The review reads the call graph, not just the diff.
Twelve specialist agents — Explorer, Architect, Editor, Reviewer, Security, QA, Debugger, Simplifier, Docs, Visualizer, Plan-researcher and a Data-architect that queries the CodeIndex graph — with five team modes and unlimited nesting. Describe a task and the Orchestrator assembles the agents, picks the mode and executes. Agents are Markdown files with YAML frontmatter, so adding one takes no code.
The software is free. GPU rental works out at about $2.50 per developer per month and is paid to your provider, not to us. CodeIndex, pull-request review, a named engineer and an SLA are an optional $10 per developer per month. Unlimited tokens on every plan, nothing metered, no per-seat licence.
Fine-tune the model on an internal framework, a regulated domain, a support corpus or a library of precedent — none of it limited to code. Fine-tunes ship as LoRA adapters loaded and swapped per request, so many departments can hold many adapters on the same GPU without changing the hardware bill. Point a CodeIndex-style index at whatever your teams work in, connect internal systems as MCP servers, and scope agents, skills and policy per team.
Nine layers: permission tiers, protected paths, command blocking, PII detection, prompt-injection blocking, audit logging, agent isolation, orchestration guards and managed policies.