Careful judgment: architectural changes, deep refactors and high-reasoning planning.
The right agent for every task. One pull request workflow for all of them.
Run Claude Code, Codex, Antigravity CLI, Mistral Vibe and OpenCode from one place. Pick the agent and model per task or per phase, on your own subscriptions or API keys, and get every result back as a GitHub pull request.
Task / PlanScope, context, branch
ClaudeDeep reasoning
CodexFast implementation
OpenCodeFlexible routing
One GitHub PR & LogReview, cost, follow-ups
Agent fit and access
Fast implementation and follow-ups on OpenAI's coding models.
Gemini, Claude and GPT-OSS models behind one Google login.
Mistral Medium and GLM models for routine reviews and implementation passes, on one Mistral account.
Open-source agent with free models for zero-cost passes and comparison baselines.
Which coding agent subscription should you use?
Pick the plan you already pay for, or the one whose models fit the work. Claude Code and Codex start at $20 a month; Antigravity CLI runs Gemini 3.8 Flash on a free Google account; Mistral Vibe Pro costs $14.99; OpenCode runs free models at no cost. Each agent page lists the provider’s tiers, included models and published limits, and every plan runs headless, which is how ProPR drives it on GitHub issues and pull requests.
| Agent | Entry plan | Top plan | Default model in ProPR | Plans and models |
|---|---|---|---|---|
| Claude Code | Claude Pro, $20 / month | Claude Ultra, $200 / month | Claude Opus 5.5 | Claude plans → |
| Codex | ChatGPT Plus, $20 / month | ChatGPT Pro, $500 / month | GPT-6 Astra | Codex plans → |
| Antigravity CLI | Free Google account; Google AI Pro, $19.99 / month | Google AI Ultra, $199.99 / month | Gemini 3.8 Flash | Antigravity plans → |
| Mistral Vibe | Vibe Free; Vibe Pro, $14.99 / month | Vibe Team, $24.99 / user / month | Mistral Medium 3.5 | Vibe plans → |
| OpenCode | Free models at $0; OpenCode Go, $10 / month | OpenCode Go Plus, $40 / month | Big Pickle | OpenCode plans → |
Provider prices exclude taxes; API keys bill per token on every agent. Verified November 2026 against each provider’s pricing page, linked from the agent pages.
What each agent runs.
Every agent plans, implements and reviews through the same PR path, as the guide to running Claude Code and its peers on GitHub issues shows end to end. Goals, the acting step in Automations and Agent Tank capacity tracking depend on the agent.
| Agent | Plan / implement / review | Goals | Automations acting step | Agent Tank capacity | Default model |
|---|---|---|---|---|---|
| Claude Code | Yes | Native, with live steering | Yes | Yes | Claude Opus 5.5 |
| Codex | Yes | Native, with live steering | Yes | Yes | GPT-6 Astra |
| Antigravity CLI | Yes | Resumable conversations | No | Yes | Gemini 3.8 Flash |
| Mistral Vibe | Yes | No | No | No | Mistral Medium 3.5 |
| OpenCode | Yes | No | No | No | Big Pickle |
Switch agents mid-task and keep one audit trail.
Start with one agent for broad context, hand the tricky implementation to Claude, and send the follow-up fix to Codex. ProPR keeps one record across all of them: what was planned, which agent ran, what it cost, and what landed in GitHub.
Plan
Claude Fable 5.1
Implement
Claude Opus 5.5
Review
GPT-6 Astra
Fix
GPT-6 Astra
Switch by phase
Pick the agent and model for each phase: planning, context analysis, implementation, review, follow-ups and repository suggestions.
Compare agents on the same task
Send one request to several agent and model combinations. Compare their pull requests side by side and merge the strongest one.
Track the cost
Every run records which model ran, how long it took and what it cost. Agent Tank adds subscription and rate-limit visibility for Claude Code, Codex and Antigravity.
Pool accounts behind one model
Put several agent accounts behind one virtual model. ProPR spreads runs across them by priority, round robin or remaining usage, and repositories and plans point at the pool.
Hand off a goal
Goals run on Claude Code and Codex with live steering, and on Antigravity through resumable conversations that take your corrections between steps.
Bring your own plan
Each agent runs on your provider subscription or API key. You pay the provider directly; ProPR never marks up tokens.
Let your assistant delegate to ProPR.
ProPR bundles its Operator Agent Skill for Codex, Claude, Antigravity and OpenCode. Once a human files or triages the issue, your assistant delegates the work through GitHub and ProPR. Humans keep every permission.
One issue. Two agent runs. Two pull requests.
In August 2026, Issue #1879 reported that propr setup with a custom GitHub App stalled on protected health checks before login. ProPR ran it through Codex with GPT-5.6 Sol and Claude Code with Claude Opus 4.8. Each run returned its own branch, summary and pull request, and both merged.
Bring a chat client to the agent workflow.
With MCP, a chat client requests work from your ProPR instance, and your coding agents run it on your infrastructure.
See how ProPR fits into the landscape.
Compare the shared PR workflow with hosted agent UIs, AI IDEs and direct CLI tools.
Compare ProPR against the alternatives →