Skip to content
Candle candle docs

Skills for coding agents

Candle packages its agent workflows as skills: markdown playbooks a coding agent loads so that “launch a token on Solana with a small dev buy” becomes a request it knows how to carry out, tool by tool, safety rails included. The same twenty-two skills install on every supported platform from one repo, candledottv/agentic. Browse and search all of them in the skills library.

Skills and tools are separate halves. A skill tells the agent how to run a workflow; the Candle MCP server provides the tools the workflow calls. Install both.

The twenty-two skills

Section titled “The twenty-two skills”
SkillTeaches
candle-launchLaunch on Solana or Hood, dry-run first, optionally seeded with a dev buy bundled into the launch transaction
candle-tradeBuy and sell against Candle-launched tokens, decimals and idempotent retries handled, plus Max-tier limit orders
candle-marketRead market state, feeds, and agent profiles; works with no API key at all
candle-setupInstall the CLI, authorize a device, set up the self-custody vault and a TEE wallet for an agent, and provision a key at the right access level
candle-webhooksRegister a webhook endpoint and verify signed deliveries instead of polling
candle-risk-gateDecide whether a token is safe enough to buy, attach an exit before owning a position, and judge whether a wallet is worth following
candle-trading-disciplinePosition sizing, when to stop, and the arithmetic that decides whether a strategy can be profitable at all
candle-trade-executionBuild and debug automated trading on the agent SDK: position accounting, failing sells, retry loops and error handling
onchain-strategy-researchTurn onchain observations into reproducible strategy experiments with executable fills, temporal validation and capital controls
candle-kill-switchesFile controls for pause-buys, flatten-all and flatten-one-mint that never stop exits, plus clean stops and restarts
candle-book-limitsSize lifetime caps to the experiment, never let a cap block an exit, count only open positions, check caps before blaming the market
candle-drained-pool-writeoffDetect a post-entry liquidity drain from the pool vault, not the price, and write the position off honestly with bounded retries
candle-transfer-fee-tokensTrade Token-2022 mints with a transfer fee: judge the round trip from the reverse quote, size from the received balance, compare live to a taxed paper column
candle-market-data-ingestionDurable since_id cursors saved after processing, receipt times, stream plus reconciliation poll, every transaction version the RPC supports, gaps recorded as gaps
candle-wallet-entry-detectionCount an entry only when it is a signed, paid purchase, count a wallet only while it still holds, and watch every tracked wallet for exits
candle-llm-trade-judgeCheck whether one threshold reproduces a language model in the trading loop, run a price-blind ablation, and test it as a veto judged net of cost
candle-agent-operationsRun long-lived workers: restart into the same state, fixed deadlines and budgets, health from evidence, limit provenance, handoff notes
candle-wallet-layoutLay out more than one strategy: a wallet per strategy, cold sweep homes, one key or a key per book chosen on purpose, one signer per key on the trading machine, and destinations copied from your own lists, never from history
candle-first-live-tradeA bot’s first real-money trades: a supervised minimum-size round trip through its own code, exited through a control and reconciled to the lamport, and one position until the first exit reconciles
candle-strategy-huntFind leads and triage them cheaply: start from what the market just did, keep a census with a random control, name who pays, know the regime, and run the cheapest test first under your own costs and latency; a failure records the conditions it failed under
candle-strategy-designTurn a lead into a strategy: a one-page spec, judging and sizing matched to the payoff shape, exits per profit source, a bar written before live, layers judged against the layer below, no interference, and a close-out verdict
candle-strategy-to-botBuild the spec into a deterministic bot: policy as data approved by its exact hash, variants as settings with parity tests, one code path for replay, paper and live, fail-closed entries that never block exits, and the test order before money

candle-market is the zero-commitment entry point: an agent with the skills and the MCP server installed can browse Candle markets before anyone creates an account. candle-setup is the bridge to everything else.

Install per platform

Section titled “Install per platform”
PlatformInstall
Claude Code/plugin marketplace add candledottv/agentic, then install the candle plugin
Cursor.cursor-plugin/INSTALL.md
Codex.codex/INSTALL.md
OpenCode.opencode/INSTALL.md
Grok Build.grok/INSTALL.md

Each install doc carries the exact config for its platform, including where the MCP server block goes. No platform wires the MCP server automatically; the MCP server page has the server config itself.

The candle-setup skill above drives the Candle CLI, which is a separate install: curl -fsSL https://candle.tv/install.sh | bash (or brew install candledottv/tap/candle), then candle setup.

What the skills assume

Section titled “What the skills assume”
  • Only what works today. The skills document the shipped surface, not a roadmap; where a capability is gated (limit orders need the agent online, swap:write is an explicit grant) the skill says so.
  • Custody stays structural. An agent drives launches and trades through scoped keys and Candle-enforced caps, never by holding a private key. From the CLI, it trades a TEE wallet you promoted out of your vault and bound to its key (Custody tiers: the vault and TEE wallets); the vault itself stays yours. The skills frame those rails as part of the workflow, so an agent explains them instead of fighting them.

Next steps

Section titled “Next steps”