ForgeCode - Worlds #1 Coding Agent

We’re working with Termbench to get our submission verified. As part of the process, the team has agreed to take down our older submissions. SPOILER ALERT: Once verified, this will be our highest score yet, beating all our previous submissions!

I built the open source codex app! One of my favorite features is “Workspaces” - reusable multi-chat layouts you can save and switch between depending on the task. Built on top of @forgecodehq

@AriesTheCoder @claudeusmaximus tbench.ai/news/leaderboa…

@claudeusmaximus @forgecodehq ** performative gasp ** @forgecodehq - you CHEATED on me?! Say it ain't so forge code. 💔

Days ago, I was checking out the once VERY relevant terminal bench. I kept seeing this Agent called @forgecodehq always in the top 5. Decided to try it and never looked back. It's now my daily ai-enhanced terminal. Feature packed, but not stuffed to the gills. I love it!

Have a look at the leaderboard in terminal bench @forgecodehq really shining. Codex with gpt5.5 hits 82% whereas forgecode hits 81.80% with gpt5.4.

This is why ForgeCode invests heavily in prompt caching. In one workspace over the last 7 days on Opus 4.7 - 407M: input tokens - 382.9M: cache-read tokens - 98.1%: cache read ratio - 22.9×: write amortization Amortization = tokens read back per token written to cache. So every 1 token cached was reused ~23 times. At public API pricing, that’s ~$2,035 without caching vs ~$333 with 5-minute caching. ~$1.7K saved in input-token cost alone.

Is the real bottleneck for AI agents the model—or the harness? Terminal-Bench 2.0 suggests it might be the latter. ForgeCode ranks #1 among open-source harnesses, showing how much performance you can unlock without changing the model—just by improving how it uses tools. In ForgeCode’s case, the gains come from better tool orchestration and execution. Learn more: tensorlake.ai/blog/forgecode…

@JustinPBarnett @tiberriver256 Please try with these Recommended Settings 1. GPT 5.5 (Medium). 2. GPT 5.4 (High) 3. GPT 5.3-codex (Xhigh)

@forgecodehq @tiberriver256 Tbh could just be my fault using gpt 5.5 xhigh. Not sure what yall would recommend as the thinking level

Anyone tried @forgecodehq? Would love some impressions

@JustinPBarnett @tiberriver256 Hi @JustinPBarnett can u tell us more about the slowness? Platform, model etc.

@tiberriver256 @forgecodehq I've played around with it a bit too, also was very slow for me

After 2 months our #1 rank on Termbench was finally broken by a worth competitor (by 0.2%) 🙌 If someone from @OpenAI can help us in getting unrestricted API access, that'd be great! We'd love to run to on @forgecodehq and share notes 😇

@im_hash_im Can you share a screenshot or join our discord, we'd be more than happy to help. Discord: discord.gg/kvQMYge4

Having some issues setting up @forgecodehq. I wanted to try and use anthropic, I put in my API key but the API requests are failing. Did the same thing in OpenCode and it worked. Couldn't figure it out and what's annoying is that the API response is truncated in the terminal

Configure a symbol + conversion rate to display costs in your local currency. Useful if you need a more accurate sense of the real value of the work being produced.

@ChamalkaAI Please share it with us, we'll fix it right away.

Tried to use ForgeCode's Forge Harness yesterday night, ran into a bunch or errors regarding the database being full and crashed out hard. Anyone know how to fix this? @forgecodehq

@mkXomj Great to hear that. Feel free to hop on to our discord, we'd be happy to help if you need anything.

あれ、やっぱりforgecode強いかも 2h動き続けてるし、buildやブラウザテストも勝手にやってる

Continuing our commitment to open-sourcing our TermBench improvements, we’re shipping another update. In `v2.8.0`, the `task` tool is now publicly available. `task` enables the main agent to delegate work to specialized, user-defined agents, keeping the context window focused and efficient. Example: hit a Rust compile error? Invoke a Rust-specific sub-agent to handle compiler runs, rules, and debugging in isolation. Enabled by default. github.com/antinomyhq/for…

@wiedymi Yes, and also made it faster! Feel free to hop on our discord, we'd be more than happy to help!

So forgecode instead of making another tui made a zsh plugin that make your prompt agentic?

@TheoLBorges @theo Looking forward to getting feedback.

@theo I think benchmarks have biases but yeah, still bad. Maybe this is why some people say to use Droid instead of CC for claude models. Currently I am trying ForgeCode. It has been a good experience, but not perfect. Next I'll try Droid.

Can't stop thinking about how Claude Code is in LAST PLACE on TerminalBench for harnesses using Opus 4.6. There are TEN separate harnesses that use Opus better than Claude Code

@devon_chaine @theo We talk about how we achieved SOTA in our blog forgecode.dev/blog/benchmark…

@theo If that were the case then everyone would just talk about and use forgecode all the time. But they don't because they optimize for the benchmarking tests. That's the problem with all these llm benchmarks. They can be tuned to pass the benchmarks

@Daniel_Kelly89 Join our discord, we'd be more than happy to help!

I've been using ForgeCode now for the past 24 hours. This is defiantly my new harness for the foreseeable future. Any veterans got some nice tips and tricks for me? #forgecode #agentic #AI

@_halshin `:config` currently shows u whats in `~/forge/.forge.toml`.

@forgecodehq This is even more confusing. Does :config show the current session configs or the global default? The video shows me using :reasoning-effort but :config stays the same

@forgecodehq Is reasoning effort not something I can configure?