CodePlane¶
The control plane for AI coding agents.
Run any agent from anywhere. Own the data. Review with intelligence.
CodePlane wraps Claude Code and GitHub Copilot with a fleet dashboard and its own AI that watches while your agents work, catching stalls, tracking progress, and scoring the risk of every code change. When an agent finishes, you get a narrative explaining what changed and why, not just a diff. Run multiple agents from a kanban board, chat mid-run, approve from your phone. Or keep using your agents the way you already do and CodePlane picks them up automatically.
Works with Claude Code CLI and GitHub Copilot CLI · Open source, MIT license
Two Modes, One Pipeline¶
A
Launch through CodePlane¶
Write a prompt, pick a repo and model. The agent runs headless in an isolated worktree. When it finishes, you get a structured review — approve from your phone or desktop.
B
Mirror your native sessions¶
Run claude or copilot in your terminal as usual. CodePlane auto-discovers the session via file-tailing and ingests it — full dashboard, cost tracking, and trail enrichment with zero workflow change.
Either way, you get the same intelligence pipeline: trail enrichment, cost attribution, structural review, narrative story, and workspace memory.
| Agent | Managed (headless) | Mirrored (native CLI) |
|---|---|---|
| Claude Code | Yes | Yes |
| GitHub Copilot | Yes | Yes |
External agents can also orchestrate CodePlane programmatically through its built-in MCP server.
The Experience¶
Supervise¶
Launch headless jobs from a dashboard or discover native CLI sessions automatically. Watch live transcripts, plan progress, and cost as the agent works. Gate risky actions for approval — from your phone, tablet, or desktop. Push notifications when the agent needs you.
Review as a Story¶
When the agent finishes, you don't get a raw diff. You get a narrative that explains what changed and why — with verified file references traced back through the decision trail. Structural risk scoring tells you which changes are breaking, which are additive, and whether callers were verified. A merge confidence verdict before you read a single line of code.
Analyze¶
Every token, tool call, and dollar — attributed by activity, model, file, and phase. Waste patterns surface across jobs: rework, rereads, retry storms, cost escalation. The analysis compounds — your 50th job is more valuable than your first because the system has learned what your codebase costs.
What's Underneath¶
Behind the dashboard, CodePlane builds a normalized intelligence layer that no agent CLI provides on its own.
Provenance Data Layer¶
Every agent session — managed or mirrored — is normalized into a local SQLite database with precomputed cost per span, per file, per phase. Token counts, retry tracking, approval audit trails, intent graphs, and cross-job observations, all in relational tables. Query it with sqlite3, point a notebook at it, or let another agent analyze your agent history. The database is a standalone asset even if you never open the UI.
Sidecar Intelligence¶
CodePlane runs parallel LLM sessions alongside the agent — not hooks that fire after the fact, but persistent reasoning that observes and intervenes in real time. A stall detector that autonomously recovers stuck agents. A planner that infers the agent's strategy. An enricher that annotates what's happening as it happens. Custom sidecars you define in plain English.
Structural Code Analysis¶
Graph-based risk scoring, not LLM vibes. CodeRecon traces callers of every modified symbol, detects dependency cycles introduced by the change, measures coupling drift across module communities, and produces a merge confidence verdict grounded in code structure. A fundamentally different signal from LLM-based code review.
Decision Trail¶
Not a transcript — a structured intent graph. Each action is recorded as a trail node capturing what the agent did, why, and where it changed course. Nodes are enriched with rationale, purpose, and edit motivations by parallel LLM analysis. The trail feeds the narrative review and powers post-hoc debugging of agent behavior.
Action Policy Engine¶
The agent SDKs give binary allow/deny. CodePlane adds cost-aware tier escalation, batch approval (consecutive gate-tier actions become one prompt), session trust grants with TTL, protected path rules, and cost ceiling rules. The SDK is the valve; CodePlane is the control system.
MCP Server¶
Agents use MCP tools. CodePlane exposes itself as an MCP server — 7 tools for job orchestration, approval handling, workspace browsing, and repo management. External agents can delegate coding tasks to CodePlane and monitor them programmatically. Agent-to-agent orchestration through a control plane.
Who This Is For¶
- Solo devs using Claude Code or Copilot CLI who want cost visibility and review tools without changing their workflow
- Teams running many agent sessions who need cost forensics and behavioral pattern analysis across jobs, not just a billing page
- Reviewers of agent output who need structural risk triage and a narrative explanation, not just a raw diff
- Regulated environments where AI-generated code changes require decision provenance and an audit trail