Run a coding harness through your signed-in chat browser. FreeChatCode connects Codewhale, OpenCode, or OpenClaw to a local OpenAI-compatible relay, then drives the provider's web UI with Playwright. The harness runs tools in your workspace; the browser supplies model replies.
Codewhale / OpenCode / OpenClaw
↓
authenticated loopback relay
↓
signed-in Chromium chat
You use your own chat account. Its subscription, usage limits, and availability still apply. No provider API key is required for the browser transport.
cargo install --git https://github.com/dixonSolutions/freechatcode --locked
freechatcode install # install Codewhale if needed
freechatcode # visible browser and CodewhaleSign in directly in the browser when prompted. Playwright's matching Chromium is downloaded automatically if missing. To install the browser beforehand:
cargo run --bin freechatcode-install-browserRun from the directory where the harness should work. Pass harness arguments
after --:
freechatcode launch codewhale -- --help
freechatcode launch opencode -- run "Inspect this project and run its tests"
freechatcode launch openclaw -- --json "Inspect this project and run its tests"OpenCode and official OpenClaw must already be installed. The Cargo package
openclaw-cli 0.1.0 is a community gateway CLI without an agent command; it
cannot run this coding adapter. Use official OpenClaw.
Watch the full OpenCode + DeepSeek demo: normalize whitespace, edit Python, and pass four tests. The terminal pane contains timestamped output from the real harness; the browser pane records the actual chat page. The fixture starts with failing tests and is checked independently after the harness exits. The publishing script refuses failed runs by default.
Record another example and render the capture:
cargo build --locked
python3 tools/record_demo.py --harness opencode --provider deepseek --scenario slug
python3 tools/render_demo.py /path/printed/by/recorder/capture.json| More examples | Scenario | Independent result |
|---|---|---|
| OpenCode + Gemini | Invoice fix with the explicit compact agent | 2 tests pass |
| Official OpenClaw + DeepSeek | Read, edit, execute | 2 tests pass |
| OpenClaw whitespace scenario | Trim, repeated spaces, tabs, empty input | 4 tests pass |
| OpenCode + Google AI Mode response | Small request with tools disabled | Exact response check passes |
| Google AI Mode diagnostic | Explicitly failed coding run | No tool call; 2 tests still fail |
Gemini's compact agent is an OpenCode project configuration, selected with
--compact-agent in the recorder. It permits bash only and explains the tool
reply shape in the harness's own prompt. The bridge adds no instructions.
See OpenCode's agent configuration.
| Provider | Browser responses | Coding tools |
|---|---|---|
| DeepSeek | Verified with a signed-in account | Real file edits and passing tests with all three harnesses |
| Gemini | Verified with a signed-in native browser | OpenCode compact-agent invoice scenario passed; default agents in all three harnesses returned prose |
| Google AI Mode | Verified through OpenCode with a small agent | Coding remains experimental; search fallbacks and context limits affect requests |
Provider model IDs describe page state. Gemini uses the model selected in its
web UI. DeepSeek exposes deepseek-chat and deepseek-pro; the latter engages
DeepThink. Provider web pages can change independently of this project.
freechatcode chatmodels list
freechatcode chatmodels configure --Gemini
freechatcode chatmodels configure --GoogleAIMode
freechatcode chatmodels set-default --Deepseek
freechatcode launch opencode --chatmodel=Gemini
freechatcode launch opencode --opt=allConfiguration checks the composer before saving the provider. That confirms
browser readiness; successful tool execution needs a live harness test.
Defaults persist. --opt=all exposes all configured models and warms their
browsers. See configuration for selectors and model
settings.
The default show mode opens a visible browser. silent runs headless; its
API attempt falls back to the page when the provider refuses it. For headless
browser transport without that attempt, use:
mode = "silent"
[transport]
mode = "gui"If Google rejects sign-in from automated Chromium, authenticate in a normal
supported browser first. Then restart that same profile with a loopback CDP
port and attach using --cdp-endpoint http://127.0.0.1:PORT. This flag also
works with chatmodels configure and doctor. Configure saves the endpoint
for that provider, so later launches can reuse its sign-in while that browser
is running. Other providers retain their own browser settings. Keep the same
browser application: copying cookies between Chromium applications can lose
authentication because their encryption differs. Credentials are entered by
you; the bridge checks the resulting composer.
Useful flags include --mode show|silent, --profile-dir, --cdp-endpoint,
--chatmodel, --model, --harness, and --harness-bin. For a managed browser,
--record-video DIR --record-video-size 960x800 records the driven page.
Attached browsers can be recorded with the capture helper.
The bridge forwards the harness's messages and tool catalog without adding instructions. It translates JSON tool calls and DeepSeek's native DSML calls to OpenAI tool calls. For native calls, the original selected reply's markdown preserves code whitespace that the rendered DOM would otherwise change. Page reasoning is excluded from replies, following the discussion poll.
A prose response ends the harness turn, including a promise to act or a claim that tools are unavailable. The bridge carries that response through; it does not invent a tool call. Live failures are tracked in issue #16.
Requests can identify independent conversations with
x-freechatcode-session-id and x-freechatcode-agent-id, or a conversation
object with session_id and agent_id. OpenCode's native session and parent
headers are recognized. Each identity gets its own tab, history, and replay
cache. Tabs expire after five idle minutes, with a limit of 64 per provider.
Harnesses without agent identity share the launch conversation; automatic
Codewhale subagent identification remains issue #1.
SQLite stores conversation URLs and message metadata for resume. Explicit image data URLs and local file parts upload through the provider's file input, up to 16 files and 8 MiB combined. Remote URLs and file IDs are forwarded as references. The harness retains ownership of workspace files and permissions.
Defaults are embedded from assets/config.default.toml.
User overrides live at ~/.codewhale/freechatcode/config.toml, or under
$CODEWHALE_HOME/freechatcode. CLI options override configuration.
freechatcode health
freechatcode turns --limit 20
freechatcode feedback --title "Reproducible failure" --body-file report.mdThe feedback command creates a GitHub issue. Include the harness/provider, steps, expected result, and relevant diagnostic output. Audit logs contain request and reply content; review them before sharing. The relay binds to loopback and uses a random bearer token; local audit and database files use owner-only permissions.
For detailed settings, recovery behavior, and timeouts, read the configuration reference. Historical measurements and test evidence live in STATUS.md. Ownership rules are in the design document.
cargo fmt --check
cargo test --locked
cargo clippy --locked --all-targets -- -D warningsLive tests are ignored by default and require signed-in provider profiles.
Run a selected live test with cargo test NAME -- --ignored --nocapture.
The demo recorder uses isolated workspaces, requires a failing baseline, and
checks the resulting files with an independent test process. The renderer
refuses failed captures by default; --diagnostic labels a failure explicitly.
Current validation: 99 active tests pass, with 23 live tests ignored by
default. See the live harness matrix for measured
results and remaining limitations.