What we've shipped and where we're headed. Tap any item for detail.
63% of the current roadmap is already shipped.
OSC 8 hyperlinks in tool results — click a file path to open it in your editor.
Unified-diff preview with accept / reject / edit-before-apply on every file write. Cluster-B research finding made visible: review is the bottleneck, not generation.
First-party git tools (status, log, diff, commit, branch, push, rebase, worktree) with explicit permission gating on destructive operations and token-efficient shaped output for small-context models.
Empirically-tuned edit formats, tool-call fallback prompts, and context compaction defaults for the small open-weight class. Benchmark-driven; results persisted in the capabilities registry.
PRs, issues, releases, and CI-status checks as first-party tools. Permission-gated merges. GitHub first; GitLab and Bitbucket follow on demand.
Source-controlled, agent-readable project memory with a typed schema for requirements / design / decisions / glossary. AGENTS.md interop included. Memory lives in your repo, not in someone else's cloud.
A thin VS Code extension that launches copair CLI sessions from the editor, passes file context, and surfaces session history. Not an in-editor agent — the CLI remains canonical. Distribution play for the VS Code marketplace audience.
Spawn multiple agents in isolated git worktrees, coordinated by a single review surface. Permission inheritance prevents escalation; review gates defend against autonomy maximalism.
A thin JetBrains-platform plugin covering the full IntelliJ IDE family (IDEA, PyCharm, WebStorm, GoLand, Rider, etc.) with the same shape as the VS Code companion: launch CLI sessions, pass file context, session-history sidebar. Single plugin across the whole IDE family.
Run copair non-interactively — a prompt in, a structured JSON result out, no TTY. A binary exit-code contract, a versioned result document and JSONL event stream (both shipped as committed JSON Schema), and config-driven harness toggles for one-at-a-time ablation. The foundation for automated runs and reproducible small-model benchmarking.
A single getCapabilities() contract resolves tier, context window, max output, native tool-call quality, and preferred format for any model — generic logic + sparse shipped data + per-model config overrides, with zero per-model code. Unknown models are declared explicitly (strict, with a clear error) instead of silently guessed; ≤22B models engage the small-model harness. Bundles harness hardening — a result-aware loop guard, tool-call format-error repair, and tool-aware output overflow — plus --explain-model for debugging.
Built-in tier classifier auto-engages a tuned harness for small open-weight models: ask_user / task_complete tools, max-turn cap, per-turn format reminders. Covers 250+ model families across 22 providers. Empirical per-model tuning + benchmark shipping under spec 032; the harness is operational, not yet battle-tested.
YAML-defined workflows with sequential and conditional steps, input parameters, loop/retry logic, and progress tracking. Engine validation in progress under spec 033.
Connect to Model Context Protocol servers for extended capabilities and external tool integration.
Install Copair via Homebrew for seamless macOS and Linux setup with automatic updates.
Extensible tool system allowing users to create and share custom tools for domain-specific workflows.
Rich terminal rendering with markdown formatting, syntax highlighting, and git diff display.
Trust-first project initialisation, per-project config in .copair/, and a structured COPAIR_KNOWLEDGE.md codebase index injected at session start.
Automatic branch creation, meaningful commit messages, and diff review built into the agent loop.
Session history and context are preserved across conversations, enabling long-running workflows.
User-defined commands with prompt templates, variable interpolation, and chaining for automating repetitive tasks.
Connect to any OpenAI-compatible API: Ollama, vLLM, LM Studio, llama.cpp, and more.
8 tools out of the box: Read, Write, Edit, Grep, Glob, Bash, Git, and WebSearch.
Switch between Claude, GPT, Gemini, and local models within the same session using the /model command.
Copair is built and maintained by one indie developer. Quarter estimates assume solo-dev pace and may slip when real life intrudes — we ship when it's right, not when the calendar says we promised.
Three shipped capabilities you won't find together on any other open-source coding agent. Closest peers named honestly; parity gaps surfaced below.
Structured ask / auto / deny with allowlists. The cross-repo guard is enforced in the tool resolver, not the system prompt — no model can talk its way around it.
Closest peer: Cline's per-step approval is the closest peer, but it's IDE-bounded and instruction-based. Codex CLI ships OS-level sandboxing, which is adjacent. Copair's guard is structured, terminal-native, and code-enforced.
Author multi-step pipelines with conditional jumps, loops, and on-max-iterations recovery handlers. Check them into your repo alongside your code.
Honest caveat: Engine validation is in progress under spec 033 — known unknowns exist. Three engine bugs were caught and fixed post-launch in spec 028 Phase C; more are likely.
Closest peer: Plandex auto-generates plans; Cline's Plan/Act and OpenCode's Plan/Build are agent personalities you toggle. Copair is the only one where you author a YAML pipeline and version-control it.
`brew install copair` works. SEA binaries for macOS, Linux, and Windows. SLSA v1 provenance on every npm release. No Python, no Node toolchain at runtime, no VS Code dependency.
Closest peer: Codex CLI is also a single-file binary (Rust) but is OpenAI-locked in practice. Aider needs Python; Cline / Roo Code / Continue need VS Code; OpenCode runs a daemon. Copair is the only fully model-agnostic agent with this distribution profile.
Last updated July 1, 2026