A Rust coding-agent proof of concept.
NextEleven Harness edits repositories with sandboxed tools, tracks costs and sessions, runs parallel workers, speaks MCP, and provides local HTTP/SSE interfaces. Users choose the exact provider, model, and fallback order; Harness does not prefer a vendor or silently insert one. The public repository is evidence of implemented systems work—not a claim of enterprise adoption or finished production support.
Why this exists
Harness explores what a maintainable local agent runtime needs beyond a one-shot demo: bounded tools, user-owned provider routing, persistence, documentation, and repeatable checks.
Every project reinvented the agent
New glue, new tool schemas, and new crash modes make cost tracking, recovery, and troubleshooting inconsistent.
A tool should not choose your model stack
Harness records the route the user selects: one exact primary model followed by explicit fallbacks in the same order. Credential detection never becomes product policy.
Agents that write anywhere are a liability
Shell, git, and file edits need allowlists, plan mode, and recoverable runs — not “trust me” on a customer machine.
What Harness does
A terminal-native coding agent with the pieces operators actually need after the demo.
User-owned models and routing
- 18 alphabetized built-in provider configurations plus custom OpenAI chat-completions-compatible endpoints
- One explicit model per provider; the first is primary and the rest are fallbacks in the user's exact order
- No recommended preset, credential-derived ranking, or implicit Ollama fallback
- Setup, browser setup, and scoped route-management commands use the same saved route
Tools on a real repo
- Read / write / patch files, search, shell, git, gh, tests
- LSP helpers, plan mode for destructive steps
- Optional gated tools: browser, docker, notebook, database
Swarm and sessions
- Parallel workers with SQLite registry and timeouts
- Session resume, project memory, semantic recall
- TUI panel for swarm status and results
Ops surface
- harness serve — local HTTP + SSE
- MCP tools/resources, doctor, bench, sync, bridges
- macOS / Linux / Windows CI on the public tree
What's ready, and what isn't
Harness is a proprietary public proof of concept with defined limitations. Passing checks provide evidence for tested paths; they do not establish complete correctness, complete security, adoption, or mature production support. Current versions and check results are documented in the repository.
What you get today
- Interactive TUI and one-shot CLI
- Provider-neutral streaming + tool loop with an exact user-selected route
- Swarm workers, sessions, cost tracking
- Local serve endpoint and MCP path
- Install script + docs in the public repo
What we won't claim
- A finished “stable 1.0” release stamp
- Open-source / MIT licensing — this is proprietary NextEleven LLC software with a public proof tree
- Turnkey SaaS hosting (local-first by design)
- Every optional tool on by default — gated extras stay off until you enable them
License: proprietary · NextEleven LLC. The GitHub repo is for proof and install visibility, not a free software grant. Custom endpoints must use the supported OpenAI chat-completions-compatible HTTP(S) shape; other protocols require a native adapter.
Get started
Install from the public repo, or talk to us if you want a custom desk that sits on this runtime.