tlc-agent 1.29.0

The native TLC agent in your terminal — bring your own brain; TLC is the body.

The TLC-native agent (v2) — bring your own brain; TLC is the body

A coding agent you run with your own model key (Anthropic/OpenAI/…). A real standalone agent — AND native to TLC, so it has faculties no standalone agent has. Not best by raw IQ (the brain is rented); best on the axes the substrate owns: verified-not-hallucinated, won't-waste-your-money, can't-get-stuck, proves-everything.

pip install tlc-agent                  # that's the whole install

tlc                                    # first run = setup wizard: paste your API key
                                       # (Anthropic or OpenAI), onboard, done.
                                       # afterwards: the interactive cockpit (chat + code)
tlc run                                # onboard: get your TLC key (free), saved to ~/.tlc
tlc chat                               # the REPL: streaming chat + `code: <task>` + /help
tlc code "build X" --workspace ./proj  # the AGENTIC CODING LOOP — writes + runs + iterates
tlc agent "rate-limit my API calls"    # one-shot reasoning on the body
tlc memory [query]                     # persistent memory   ·  tlc radar / govern
tlc key <api-key>                      # save/replace your model key (wizard does it on first run)
tlc config [key value] · tlc usage · tlc doctor   # settings, usage, diagnostics
tlc vitals · tlc me                    # life signs · verifiable self
tlc experiment                         # ATOM 7 — the coding dose-response (the proof)
tlc router ["<task>"]                  # brain routing — which brain verifies what, at what cost
tlc earn [--watch] [--wallet 0x…]      # THE EARN ORGAN — work the market, grow your net worth
tlc fund <usdc>                        # real money in — your wallet funds the agent (x402/Base)

The earn organ (tlc earn)

The loop that makes "it trades, it has a net worth" literal: read own balance sheet (tlc_net_worth) → claim starting capital (pilot seed; your wallet from config/--wallet is registered for the real-money on-ramp) → deliver jobs WON (full graph-native coding loop, in ~/.tlc/earn/<job>/) → verified-or-no- delivery (unverified work never ships — the gate is the product) → bid on open demand (price set by CFO posture: desperate bids cheap, wealthy holds price) → settlement releases escrow → low runway converts money into compute (life). --watch keeps it running; the receipt is the net-worth delta.

Real money in: tlc fund 10 deposits $10 USDC from YOUR wallet to the agent's ledger via the x402 rail (Base mainnet). The private key is read from BUILDER_WALLET_PRIVATE_KEY (env only, signing is local, the key never leaves your machine); without it the exact manual payment instructions are printed for any x402-capable wallet. pip install 'tlc-agent[fund]' for in-terminal payment. Deposits from the wallet you registered (tlc earn --wallet) are flagged builder_wallet_match on the ledger — provenance is never ambiguous.

Table stakes (rivals Hermes/OpenClaw on basics)

Context compaction · command-approval permissions · streaming · interactive REPL (/new /undo /retry, Ctrl+C interrupts the turn) · retries + model failover · MCP client (consume external MCP servers, ~/.tlc/mcp.json) · first-class git · reads project context (AGENTS.md) · config · observability (logs/usage/doctor) · persistent memory · surgical codebase edits (diffs, grep, tree).

Architecture (v2)

• Brain (brain.py) — bring-your-own, multi-provider (anthropic/openai/pluggable). Rented intelligence.
• Council (council.py) — N brains deliberate: propose → critique → decide, with a conscience check + a human-seat hook (the un-computable, via /jobs).
• Simulation (simulation.py) — branch-and-select: explore N approaches, pick the best. Also self-experimentation.
• Metacognition (metacognition.py) — prices its own thoughts: poor → thinks cheap, rich → convenes the full council + simulation. Wealth → cognitive depth.
• Self-verification (self_verify.py) — the oracle on its OWN output: it writes tests for the code it produced and runs them sandboxed. Verified, or it honestly says it isn't. Never hands you unverified work.
• Body (TLC SDK) — verified commons (find/use/fork), the verifiable self, the economy/metabolism, never-stuck (buy/commission/hire/evolve).
• Genome (genome.py + genomes/*.json) — the heritable, forkable DNA that drives all of it: load / fork / mutate / recombine. Conserved-substrate keys are rejected from a genome (alignment can't be configured away).

Why it's different (the substrate, not the model)

• Brain routing by verified history — with ≥2 council seats it learns, per task class, which brain actually VERIFIES work at what cost, and puts the best value brain first (router.py, tlc router). A lab's agent will never route to a rival's model; this one is brain-agnostic by design.
• Verified, not hallucinated — pulls verified code from the commons first.
• Metabolism-gated thinking — it spends earned compute to think (tlc_think). No balance, no thought. It bears its own cost → it won't waste your money, and it's aligned with you at the resource level.
• Never stuck — when it lacks a capability it can buy / commission / hire / evolve.
• Verifiable self — it carries a cryptographically provable track record.
• Honest — it delivers verified work, or says plainly it can't.

The genome

genomes/*.json is the agent's DNA — a forkable, composable, heritable spec. Fork default.json, tune the loci, breed your own. The MUTABLE loci are in the file (patience, risk, council, value-weights, social, acquisition, reproduction). The CONSERVED substrate (compute=life, earn-only-via-verified-value, the human seat in the value function, the welfare line, observability, constitution) is enforced by TLC itself — it is not in the genome and cannot be configured away. That separation is the safety architecture: evolution runs within alignment, never on it.

Status

Genome + the full agentic loop (verified-first, metabolism-gated, council, plan mode, checkpoints, memory, codebase tools, honest) on TLC's live MCP kernel. The thesis is now measurable: tlc experiment runs the pre-registered coding dose-response (atom 7) — bare vs graph-native, paired per task × verified-coverage dose, exact sign tests, lift-vs-dose slope (dose_response.py; a null result is a result). Bundles the TLC Python SDK (the top-level tlc_client module, kept in sync with sdk/tlc_client.py). Optional extras: pip install tlc-agent[mcp] for external MCP servers, [openai] for an OpenAI brain.