$ overbrilliant — the multi-agent CLI coding agent. completely free.
12,685 lines of TypeScript + Bun. Multi-agent orchestration with parallel sub-agents running across isolated worktrees, adaptive routing, and compute-matched eval — backed by FreeLLMAPI, which stacks 16+ providers' free tiers into one endpoint (~1.7B tokens/month, zero cost). Sign up with email, no credit card ever.
01
completely free — no API keys, no billing, no limits
OB-1 ships with FreeLLMAPI built in — a proxy that stacks 16+ LLM providers' free tiers into a single endpoint. ~1.7B tokens/month across the fleet, zero cost. Sign up with email, get a bearer token, start coding. No credit card ever.
→FreeLLMAPI: a single URL that round-robins across Anthropic, OpenAI, OpenRouter, and 13+ other providers — whenever one throttles, the next picks up.
→~1.7B free tokens/month total across the federation. Enough for serious daily use, not just demos.
→/models setup tab: live connection test with auto-model picker — picks the best free model available right now.
→Bring your own key optionally: Full Anthropic, OpenAI, OpenRouter, and any OpenAI-compatible HTTP endpoint supported too.
02
multi-agent orchestration
A sub-agent runtime that spawns, isolates, and coordinates parallel workers. Fan-out N sub-agents across git worktrees, manage long-running tasks with stall detection, and keep token spend tight with an adaptive router that escalates only when Solo fails.
→Worker runtime: each sub-agent gets its own LLM context, worktree copy, and token budget — mutations never collide.
→Fusion: run N parallel candidates → auto-score with syntax/typecheck/compile → synthesize from the passers.
→Adaptive router: runs Solo first, scores it against an objective check, escalates to Fusion/Council only on failure.
→Long-running task management: TaskLedger + ProgressLedger track progress, detect stalls at 3 identical steps, and auto-replan (max 3 replans, max 20 rounds).
→Token optimization: Solo@k math proves whether a heavy mode justifies its cost — if k cheap attempts at the same budget outperform, the router picks Solo.
03
gated agent loop
A ReAct cycle with per-action approval gates, streaming tokens, and auto-verification — everything Claude Code does, with read/write/bash/edit tools and plan/act modes.
→Per-action approval: ask / warn / deny gates based on a configurable policy.json file.
→Auto-verification: after every file-changing turn, OB-1 runs your project's typecheck, test, lint — and fixes what fails.
→Checkpointing with git worktrees: /rewind reverts to a known-good state. Zero lost work.
04
compute-matched eval — built-in honesty
No vague benchmarks — SWE-bench-shaped tasks with objective exit-0 checks. OB-1 proves each orchestration mode justifies its token cost against Solo@k.
→Objective pass/fail: runs tasks in isolated worktrees, exit 0 = PASS. No model-as-judge ever.
→Solo@k math: what would k plain attempts at the same token budget achieve? If advanced modes don't beat it, the answer is honest: just use Solo.
A SQLite-backed fact store you can actually inspect — facts, relationships, and reflections — all in one file. It remembers across sessions and injects relevant context just-in-time.
→Plain-text facts with an immutable revision trail — Google Memory Bank model, never deletes, only appends.
→Bi-temporal relationship graph: entities and typed edges that are invalidated (not destroyed) so history stays queryable.
→Vector search via sqlite-vec (native KNN) with a zero-dep pure-TS cosine fallback — identical ranking either way.
OB-1 builds a cross-reference graph of your codebase, ranks files by centrality, extracts symbols, and connects to language servers — so it knows your project before it touches a file.
→Cross-ref graph + PageRank centrality — finds the files that actually matter in your repo.
→Key symbol extraction via tree-sitter + LSP integration (diagnostics, go-to-definition).
→Auto-generates a project index (AGENTS.md) with repo summary and conventions.
→Git state tracking with branch analysis, worktree recovery and checkpointing.
→Available as the repo_map tool and the /map slash-command.
07
safety sandbox
OS-level containment for shell commands, configurable permission policies, and destructive-command detection — so agents can't brick your machine.
→macOS Seatbelt sandbox + Linux bubblewrap for all spawned processes.
# clone and install the launcher
git clone https://github.com/tashfeenahmed/ob1.git
cd ob1 && ./scripts/install.sh
# or compile a self-contained binary
./scripts/install.sh --binary
# then, in any project
cd ~/code/my-project
ob1
one binary. zero runtime deps. reads your repo, writes code you approve.