If you've been trying out AI coding agents and kept running into the same three questions — "where is my code actually going?", "can I use a local model instead of paying per token?", and "how do I know what it did to my repo?" — I want to show you Tera Pilot.
It's an MIT-licensed, self-hosted coding agent built around one idea: trust through evidence, not marketing claims.
What it actually is
Tera Pilot is a Textual-based TUI-first coding agent (with a web UI, a REST/SSE daemon, and MCP/ACP support too) that runs a classic Plan → Explore → Act → Verify → Report loop against your repo. What makes it different from the usual "wrapper around an LLM API":
- Bring your own key or go fully offline. 16+ providers (Anthropic, OpenAI, Gemini, DeepSeek, Groq, xAI, Mistral, Cerebras, Together, Fireworks, SambaNova, Nvidia NIM, OpenRouter…) or point it at Ollama / LM Studio and never send code to the cloud.
-
Sandboxed by default. Workspace-scoped file access, command allow/deny policy, an OS-level sandbox (macOS
sandbox-exec, Linuxbubblewrap) that denies network and blocks reads of~/.ssh,~/.aws, etc. during command execution. - A "Guardian" that reviews risky tool calls and can approve/reject/modify them — and fails closed on errors instead of silently letting things through.
- Signed, tamper-evident audit trail. Every tool call is logged; you can export it with Ed25519 signatures + a SHA-256 hash chain and verify it on a completely different machine.
- It publishes its own security testing. 671 tests total, 210 of them security/sandbox/policy tests, plus a documented list of five real vulnerabilities they found and fixed via offensive testing (git alias/config shell escapes, an SSRF-adjacent CORS bug, npm script execution, etc.) — with an explicit "what we're NOT claiming" section instead of a hand-wavy security badge. There's also a reproducible eval harness (58 tasks: bug fixes, refactors, test repair, code review, adversarial security tasks) with results checked in, including a run against a fully local 2.6B model in LM Studio.
Why I'm posting this
The project is honest about being in a testing phase — the README says so upfront, no marketing gloss. The maintainer is looking for people to actually run it against real repos, hit edge cases, and file issues. That's a refreshing amount of candor for a project asking for stars, so I wanted to help it find the right audience: people who care about local-first tooling, sandboxing, and not vibing on trust claims.
Quick start
npm install -g tera-pilot
tera-pilot-tui # full-screen terminal UI, the primary way to use it
If npm gives you trouble (it's noted as the rough edge right now), the source install is one extra step and more reliable:
git clone https://github.com/ilyaosovskoi/tera-pilot.git
cd tera-pilot
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
Then run tera-pilot doctor — it checks your Python version, provider keys, local model servers, and workspace in one shot, so you know immediately if something's misconfigured.
Who this is for
- Devs who want a coding agent but don't want their private repo leaving their machine
- Anyone who wants to try local models (Ollama/LM Studio) as a serious agent backend, not a toy
- People who actually read threat models — there's a public one, plus a security report with real numbers, not just a shield emoji in the README
-
Teams evaluating self-hosted alternatives to hosted agent products before committing
How you can help
Clone it, run
tera-pilot doctor, and try a real task on a real repo. Bug reports and rough-edge reports are exactly what the project needs right now.If it does what it says on the tin, a star helps a lot — it's a solo/small-team effort trying to get real-world signal before a public release.
Check the security report and threat model if that's your kind of reading material — feedback from people who actually think about sandboxing is valuable here.
🔗 Repo: https://github.com/ilyaosovskoi/tera-pilot
Would love to hear what breaks for you.
Top comments (0)