A coding assistant that proves its work.
lex-code writes Lex and ends every task with a verdict from the type checker and from examples written before the code — not a transcript claiming success. Give it one task, or a brief for a whole package; run it on your own machine with a local model, or overnight, unattended.
A real, unedited run: the request is plain English, not Lex.
What it does
Verifies, doesn’t narrate
A typed issue is a contract: exact signatures plus the examples that decide whether it holds. The run ends with one machine-readable verdict, whatever the model claims.
Builds whole packages
From a brief: a checked plan, units built in dependency order, generated invariants, the real program started and attacked with the brief’s own requirements, and repair when it fails.
Runs while you sleep
lex-code-overnight resumes across
rounds, waits out a provider that goes away, restores the source file after a kill, and writes
a report.
Sandboxed by the language
Every session runs under a capability
grant enforced by the Lex VM. An effect it was not granted, it cannot perform.
--lex-os adds an outer, host-level perimeter.
Local-first, any provider
--ollama runs entirely on your
machine, no key. Cloud and OpenAI-compatible servers use the same flags.
Leaves a trail
Each session is a database of events, each issue is in a content-addressed store, and a running build has a read-only live dashboard.
How a package gets built and checked
- PlanAn agent drafts a graph of typed issues from your brief. Nothing is filed yet.
- Check the planUnbuildable examples, cycles, name collisions and invariants that can never fail are rejected before any code is written.
- BuildUnits are built in dependency order. Each is verified against its own examples; a later unit that breaks an earlier one is caught.
- HardenInvariants declared in the plan are checked on generated inputs; violations are fixed.
- AcceptThe real program is started and the brief’s requirements are replayed against it as requests.
- RepairFailures that only exist once the units share a file go back to a model as one task.
- ReportVerdict, per-unit attempts and tokens, and what is still wrong.
- Stop anywhereEvery stage is a command and a file you can read, edit and resume.
“Verified” is necessary, not sufficient: a unit’s examples are a finite list. That is why a build ends by running the program — and why you should attack the result yourself before exposing it.
Quick start
lex-code --ollama # interactive, local, no key lex-code --ollama "implement list.zip" # one task, then exit lex-code --plan --ollama # plan only, nothing written
Pick a provider with its flag; --ollama needs no key. See Providers.
lex-code --issue=<id> --ollama # implement a typed issue
# ends with [ISSUE_VERDICT]
A typed issue is exact signatures plus the examples that decide whether it holds.
Implementing a typed issue: iterate against the type checker and the declared acceptance, then one verdict line.
Whole packages, in stages you can stop at
lex-code --package "a small invoice REST API ..." --name=invoices --ollama # 1. plan; files nothing lex-code --package-check=invoices # 2. re-validate after you edit it lex-code --package-apply=invoices # 3. file the reviewed plan as issues lex-code --project=invoices --ollama > invoices.log 2>&1 & # 4. build, harden, accept, repair lex-code --dashboard=invoices # watch it live (127.0.0.1:7800)
Or all of it, unattended, with a report at the end:
lex-code-overnight --name=invoices --brief-file=brief.txt --ollama cat .lex/overnight/REPORT.md
If a unit gets stuck you can resume with --hint or write it yourself with
--patch: verified units are kept, and a patch goes through the same verification as a
model’s work. Read the packages guide →
Providers
Ollama
Fully local, no key. Default model qwen3.8:27b-mlx.
lex-code --ollama "..."OpenCode Go
Cloud access to several open-weight coding models behind one key.
export OPENCODE_API_KEY=...
lex-code --opencode "..."OpenAI-compatible server
Any vLLM-style endpoint, including a local inference engine.
export VLLM_BASE_URL=.../v1/chat/completions
export VLLM_MODEL=...
lex-code --vllm "..."Others
LiteLLM, lex-gpu, OpenAI, Mistral, Google, Vertex and Anthropic are implemented with less mileage. Only the three at left have been run end to end in this repository.
Provider guide →Safety model
- Capability grants, enforced by the VM. A session cannot perform an effect (
net,fs_write, …) it was not granted, whatever its prompt says. - A permission gate sits in front of the tools that change things.
- An outer sandbox on request.
--lex-osruns the whole session inside lex-os’s host-level perimeter. - Unattended runs execute shell commands. Run
lex-code-overnightin a dedicated project directory, or with--lex-os. - The dashboard listens on 127.0.0.1 only and is read-only.
Documentation
Status and limits
- Young, and mostly run with one local model. Whole-package builds have gone from a brief to a passing acceptance gate, and have also stalled; each cause found so far now has a check, and a new brief or library can still find another.
- A passing gate is not a security review. Acceptance scenarios come from the brief’s own requirements. Treat
doneas “meets the brief as tested”, and test the result before exposing it. - Local models are slow. A whole package takes hours on a laptop-class model; that is what the overnight supervisor is for.