Corvid 0.1.0: the first release

A local-first agent you can install with one line: the decision fabric, approvals and sandbox, Vault, memory, skills, the scheduler, and front ends for the terminal, editors, HTTP and chat apps.

Corvid started from one question. What if an agent stopped sending every small decision through one big model? Small, calibrated models on your own machine could decide which tool to use, whether a step is risky, and whether to ask or escalate. The generator would then only have to write. Version 0.1.0 is the first build of that idea you can install and run.

curl -fsSL https://raw.githubusercontent.com/starlinghq/corvid/main/install.sh | sh
corvid doctor
corvid model download qwen3.5-0.8b gliner2-base-v1 embeddinggemma-2 kev-0.8b
corvid chat

Add sh -s -- --cuda to the installer for the CUDA build, which runs Kev on an NVIDIA GPU. Shell tools also need bubblewrap. The default models are about 10 GB, and nothing downloads until you ask.

What's in it

• The decision fabric. Kev, GLiNER2 and the embedding models answer each turn's questions locally, against calibrated floors. A confident answer acts, and an unsure one asks or escalates to a bigger generator, local or cloud.
• Approvals and the sandbox. Shell commands are parsed, their targets extracted and checked against a deny list before any model is asked. They run under corvid-sandbox (bubblewrap, Landlock and seccomp). Outside content taints the session, and twelve detection rules watch the log.
• Corvid Vault. A separate, model-free process holds every key. The agent sees placeholders, and Vault swaps in the real value only for the host the secret is bound to. Sandboxed scripts can use a token they never see, through Vault's per-sandbox proxy.
• Checkpoints. /undo and /rollback N restore from a shadow git store, and a three-way merge protects your own edits.
• Memory, lessons and skills. New facts are staged for your approval. Skills are drafted from successful tasks and kept only if a replay passes.
• A durable scheduler. cron_create, /cron runs, and pause and resume.
• Ways in. corvid chat, corvid acp for editors, corvid serve (OpenAI-compatible), corvid gateway for Telegram, Discord and Slack, and corvid mcp serve.

The release notes list every change.

Measured, including what didn't work

Phase 0 of the plan was about proving the fabric with numbers before building on it. All five spikes are written up in the repository, including the ones that failed. Three of them shaped this release most:

• S1: Kev-0.8B answers a four-question ingress checkpoint in about a quarter of a second on an RTX 3090 (220–260 ms at the median across the quiet runs). That passes the 300 ms target, so the fabric can run on every turn.
• S2: showing the model a shortlist of tools instead of all 28 cut a median turn on the 3090 from 8.1 s to about 2.2 s. But the full pipeline, a shortlist plus Kev's pick, chose the right tool 17.8 points less often than the all-tools baseline, failing the spike. The write-up traces most of the gap to an over-eager check on tool descriptions, which has since been changed (C-037). With that check out of the way, the pipeline came out ahead.
• S3: GLiNER2 isn't good enough at finding the targets of shell commands. The structural parser stays the authority for approvals.

What's next

The next big piece is the terminal UI. It renders inline by default and never clears your scrollback, with fullscreen and screen-reader modes built on the same components. The mockups on the home page come from that design. After that comes Windows and macOS support.

Corvid is pre-release. If something breaks, open an issue. New posts land here and in the feed.