Why ORCH

Every pain we already solve

14 real problems with multi-agent coordination. How ORCH handles each one. What's still in progress.

14 solved · 2 in progress · 2 coming soon

Coordination

Pain

Agents overwrite each other's files

Two agents edit the same file. Merge conflict. You spend 20 minutes fixing it manually. Work gets lost.

Solved

Git worktree isolation

Each agent works in its own git worktree on a separate branch. Parallel execution without file conflicts. Scope overlap detection prevents assignment of conflicting tasks.

shipped

Pain

Agent crashes at 2 AM. You don't notice.

An agent stalls or dies silently. You wake up to zero progress. Half a night wasted. Context lost on restart.

Solved

Auto-retry + stall detection

Exponential backoff retry with error context. Stall detection kills zombies and re-queues tasks. Previous failure context is injected into retry prompt.

shipped

Pain

Tab-switching between 3 terminals

Constantly switching between Claude, Codex, Cursor tabs. "Which one was doing what?" Cognitive overload. 40-60% of time spent routing, not building.

Solved

TUI dashboard

One terminal. All agents, tasks, goals, activity feed, token costs — real-time. Keyboard-driven. You see everything at a glance.

shipped

Pain

Copy-pasting context between agents

Agent A finished auth module. Agent B needs to know. You manually copy output from one terminal to another. Context gets truncated, lost, or stale.

Solved

Inter-agent messaging

Direct messages, team broadcasts, shared context store. Automatically injected into prompts at dispatch. orch msg send / orch msg broadcast

shipped

Pain

Tasks get lost, no process

No formal tracking. "Did that agent finish?" "Was the code reviewed?" Things slip through. Broken code ends up in main.

Solved

State machine governance

Every task flows through todo → in_progress → review → done. Mandatory review gate. Nothing merges without your approval. No task gets lost.

shipped

Pain

Close laptop — agents stop

ORCH runs in your terminal and stops when you close it. You want agents processing tasks 24/7 on a server — like a CI pipeline, but for AI work.

Solved

Headless daemon mode

orch serve — runs 24/7 with structured JSON logs for Datadog, Grafana, or jq. Deploy with pm2 or systemd. --once mode for CI/CD (exit 0/1). Graceful shutdown on SIGINT/SIGTERM.

shipped

Planning

Pain

"How do I split this into tasks for agents?"

You know you want auth. But how to break it into 5 independent tasks that agents can do in parallel? Wrong granularity = wasted time.

Solved

Goal decomposition

Describe your goal in plain English. CTO agent decomposes it into tasks with dependencies and priorities. You review the plan, then run. orch goal add

shipped

Pain

Manually assigning tasks to agents

You decide which agent gets which task. Waste time matching capabilities to tasks. Forget an agent is free.

Solved

Priority dispatch

P1-first, goal-linked tasks prioritized. Idle agents automatically pick up next available task by skill match. No manual assignment needed.

shipped

Pain

"I don't know where to start"

New tool, new concepts. Agents, tasks, teams, goals, adapters — too many things at once. You close the tab and go back to Claude.

Solved

10 org templates

orch org deploy startup-mvp — pre-built team with roles, prompts, and workflow. Also: security-dept, test-factory, content-agency, sales-machine, and 5 more.

shipped

Quality

Pain

Result is "almost right" — easier to fix yourself

Agent's code is 90% there. You need small tweaks. But re-prompting means re-explaining everything. Faster to just edit manually.

Solved

Rework loop

Reject task with feedback → agent retries with your notes injected. Full context preserved. No re-explanation. orch task reject -r "fix X"

shipped

Pain

Locked into one AI provider

You use Claude for everything because switching is friction. But Codex is better for tests, Cursor for multi-file edits, shell for scripts.

Solved

5 adapters, one state machine

Claude, OpenCode (Gemini/DeepSeek), Codex, Cursor, Shell — all coordinated through the same orchestrator. Best tool for each job.

shipped

Pain

Agent "done" but code doesn't compile

Agent claims task is finished. You merge. Tests fail. Typecheck broken. You find out after wasting 20 minutes debugging.

Solved

Staged fail-fast review

Automated review gates: typecheck → lint → test_pass. Stops on first failure — no wasted compute. Auto-approve or hold for human review.

shipped

Pain

Setting up agents takes 30 minutes

You need to configure roles, models, skills, approval policies for each agent. Too much friction before you see results.

Solved

Agent shop + org templates

15 pre-built agent templates via orch agent shop. 10 org templates deploy full teams in one command: orch org deploy startup-mvp

shipped

Pain

Managing ORCH from Claude Code is clunky

You switch between Claude Code and terminal to run orch commands. Context-switching kills flow. You want to stay in one place.

Solved

/orch skill for Claude Code

Type /orch deploy a team for my project in Claude Code. Natural language → CLI commands. Manage agents without leaving your editor.

shipped

In progress

Pain

Can't try without API keys

You install ORCH but need Claude/Codex API keys to see anything work. 15+ minutes before first "wow". Most people leave before that.

In progress

orch demo

Zero-key demo mode using shell agents. See 3 agents complete 5 tasks in 60 seconds. Full TUI experience without any API credentials.

building

Pain

Afraid to run on real project

"What if agents mess up my code?" Git worktree helps, but the fear of first run is real. You want to see what WOULD happen first.

In progress

--dry-run mode

Agents plan but don't execute. See "Agent would modify: src/auth.ts (+45, -12)" without touching your code. Try before you commit.

building

Coming soon

Pain

No idea how much this will cost

You leave agents running overnight. Morning surprise: $87 in API costs. No budget controls. No projection. No pause.

Coming soon

Budget caps + cost projection

max_cost_per_run in config. Auto-pause when budget is exhausted. TUI shows projected total cost based on remaining tasks.

planned

Pain

No one else is using this

Zero adopters listed. No case studies. No testimonials. "If it's good, why is no one using it?" Social proof is missing.

Coming soon

orch report --share

Beautiful shareable report after each run: "5 agents, 23 tasks, $14, 89h saved." Post to X/LinkedIn. Every successful run becomes social proof.

planned

14 solved. 4 more on the way.

ORCH already handles the hardest parts of multi-agent coordination — isolation, retry, state tracking, messaging, staged review, and 24/7 daemon mode. The rest is coming fast.

Try it on your project →

MIT license · zero cloud · one npm install