ForgeCode
ForgeCode is an open-source coding harness that runs in your terminal. Type a colon at the Zsh prompt and it plans, writes, refactors and commits code using OpenAI, Anthropic, OpenRouter, Groq, Bedrock or local models.
What is ForgeCode?
ForgeCode is a command-line coding harness published by Tailcall, Inc. Its own documentation frames it bluntly: think Claude Code, but with first-class support for many AI providers. It works with cloud models, open-weight models and models running locally, and it is open source under the Apache-2.0 licence, written in Rust, in the public repository tailcallhq/forgecode.
There are three ways to drive it. Running forge with no arguments opens an interactive terminal UI for multi-step work. Passing -p runs a single prompt and exits, which suits scripts and pipes. The third way is the one the product is built around: a Zsh plugin that turns a bare colon at your shell prompt into a message to the agent, while existing aliases and Oh My Zsh plugins keep working alongside it.
Underneath sits a multi-agent architecture with bounded context. Three named agents divide the work. forge reads and writes and implements changes, and is active by default. muse also reads and writes but is meant for planning and impact analysis. sage is read-only and is never selected by hand — the other two call it internally to research a codebase. Conversation and project context survive a switch between agents, and models are chosen separately from agents, so a reasoning model can plan while a fast model executes within the same session.
ForgeCode Services is an optional runtime layer on top. It brings a context engine that the vendor says beats the state of the art on retrieval benchmarks while using up to 93% fewer tokens, tool-call guardrails that catch and auto-correct malformed tool arguments, and a skill engine that picks the right SKILL.md workflow for the task at hand. Enabling it needs no API key — a browser sign-in with Google or GitHub — and running :sync indexes a project for semantic search.
The homepage claims first place on TermBench 2.0 with 81.8% accuracy and reports 38.1 billion tokens and 24.4 million lines of code handled per day. These figures are the vendor's own.
What it does
- Write, refactor and debug code autonomously from the terminal
- Plan a change with the muse agent, then implement it with the forge agent
- Search an entire codebase semantically once it has been indexed with :sync
- Switch LLM provider or model mid-session without restarting the conversation
- Generate a git commit message and commit in a single command
- Turn a plain-English description into a shell command with :suggest
- Connect external tools, APIs and services through MCP servers
When to use ForgeCode / When not to
A quick filter to help you decide if ForgeCode is the right fit.
When to use ForgeCode
- Developers who live in the terminal and do not want to switch editors to get AI assistance
- Teams that refuse vendor lock-in and want to mix models freely — a reasoning model to plan, a fast model to code, a large-context model for big files, all within one session
- Zsh and Oh My Zsh users, whose existing aliases and plugins keep working once the colon-prefix plugin is installed
- Engineers working in large repositories, where the context engine is claimed to navigate the codebase with up to 93% fewer tokens
- Privacy-conscious developers, since prompts, completions and conversation history never reach the publisher's servers
When not to use ForgeCode
- Anyone expecting a graphical interface: there is no web app, no mobile app and no standalone GUI — everything happens in the terminal
- Users on Windows without WSL or Git Bash, or on a machine where Zsh and a Nerd Font cannot be installed
- Teams wanting shared or team accounts on individual plans: account sharing is prohibited and usage is capped at two simultaneous devices
- Builders who want to embed the ForgeCode API in their own product — reselling access, integrating it into third-party software or using it as a backend service is banned outright
- Regulated organisations that need a signed DPA, a subprocessor list or a SOC 2 / ISO 27001 certificate: none of these is published
How to use ForgeCode
A typical end-to-end flow, from setup to results.
- Install a Nerd Font and make sure Zsh is installed and configured — both are hard prerequisites
- Install the binary with the one-line installer, then confirm it with forge --help
- Run forge zsh setup and follow the interactive wizard to install the Zsh plugin
- Restart the terminal, or run exec zsh — the colon trigger stays inactive until you do, which the documentation names as the most common failure
- Run :login to pick an LLM provider and enter an API key, or reuse an existing ChatGPT Plus or Claude subscription
- Run :model to browse and filter the model list; the choice is remembered across sessions
- Send a first prompt by typing a colon, a space and your request; press Tab straight after the colon to list every available command
- Optionally enable ForgeCode Services by running :login again and selecting ForgeServices, then authenticating in the browser with Google or GitHub
- Run :sync to index the project, :sync-status to follow progress and :tools to confirm sem_search is loaded
- Tune behaviour through AGENTS.md, SKILL.md, .mcp.json, .ignore, .forge.toml and permissions.yaml as needed
Pros & Cons
Pros
- Open source under Apache-2.0, in a public repository that anyone can audit
- Genuine provider independence, with six built-in providers plus custom endpoints and locally hosted models
- Bringing your own API keys is expressly allowed and lifts usage limits entirely
- A permanent free plan that the vendor explicitly distinguishes from a time-limited trial
- Stronger code confidentiality than most rivals: prompts, completions and conversations never touch the publisher's servers
- Semantic indexing is optional, reversible and user-deletable, and .ignore keeps chosen files from ever being uploaded
- Configurable guardrails and a restricted shell mode for teams that want to bound an autonomous agent
Cons
- No pricing page at all: the only published rates come from a blog post dated 27 July 2025
- No postal address, no contact page, no about page and no legal notice anywhere on the site
- No email address on the product's own domain — the sole contact is on the parent company's domain, with first-line support handled through Discord
- No DPA, no subprocessor list, no security certification and no trust or security page
- No Article 27 EU representative, no data protection officer and no quantified retention period
- Demanding prerequisites: Zsh, a Nerd Font, a mandatory terminal restart, and Windows only through WSL or Git Bash
- A two-device limit, monitored usage patterns, and suspensions or bans applied without refund
Pricing & Plans
A permanent free plan is available, described by the vendor as a free tier rather than a time-limited trial, with a dynamic allowance of roughly 10 to 50 AI requests per day depending on server load. The lowest paid entry point is the Pro plan at USD 20.00 per month, which raises the allowance to 1,000 AI requests per day. Users who supply their own provider API keys may run ForgeCode without a request limit, paying inference costs directly to that provider. These figures are drawn from a blog post published on 27 July 2025, the site having no pricing page.
- Free — permanent free tier with a dynamic limit that adjusts to server load
- typically 10 to 50 AI requests per day
- Pro — USD 20 per month
- up to 1
- 000 AI requests per day
- Max — USD 100 per month
- up to 5
- 000 AI requests per day
- BYOK — bring your own provider API keys for unlimited use of ForgeCode
- with inference billed by the provider
Data, GDPR & hosting
A consolidated view of how ForgeCode handles your data.
GDPR overview
The privacy policy carries a dedicated section 10, "EEA- and UK-Specific Disclosures", setting out seven GDPR rights: access, rectification, erasure, restriction of processing, portability, objection and withdrawal of consent. A separate section 9 covers the CCPA and states that no personal information of California residents is sold. Beyond that, implementation is thin. No lawful basis for processing is stated, no retention period is quantified, no Article 27 EU representative is named and no data protection officer is appointed. International transfers are announced, including to the United States, with "appropriate safeguards" that are never described. No data processing agreement is published or offered, and no subprocessor list exists. The single contact point for privacy questions is support@tailcall.run. The policy is dated 29 May 2026.
Who owns the data?
Under the privacy policy dated 29 May 2026, ForgeCode Services stores exactly two things on the publisher's servers: the raw chunks of your source files and their vector embeddings — nothing else. Prompts, completions and conversation history never arrive there; they travel from your machine straight to the LLM provider you chose. The publisher states it does not train models on that code, does not sell it and does not share it with any third party. Synced data stays until you delete it yourself with the :workspace command. Account information may be shared with vendors and consultants, on legal request, during a merger, or across group companies with your consent.
Reuse rights
The Fair Usage Policy is explicit about what you may do with what the tool produces. ForgeCode may be used for any coding project, personal or commercial, as well as for code generation, debugging, refactoring and learning or teaching. Nothing in the published terms claims ownership over the code you write with it, and no permission is required to reuse that output. The restrictions run the other way, targeting the service rather than its results: the ForgeCode API may only be used by the account holder inside the ForgeCode application, and powering another application with it, integrating it into third-party software, reselling access or running it as a backend service all lead to a permanent ban with no refund. Supplying your own provider API keys is expressly permitted and lifts usage limits.
Data retention & training
Hosting summary
The site names no hosting country or region for its data. The privacy policy says only that information may be transferred, processed and stored outside your home country, including in the United States, where data protection law may be less comprehensive than at home, and promises "appropriate safeguards" without describing them — no standard contractual clauses, no named mechanism. ForgeCode Services stores synced file chunks and their embeddings on the publisher's servers, again with no location given. What is stated clearly is what is not hosted there: prompts, completions and conversation history never reach that infrastructure, travelling instead from the developer's machine straight to the chosen LLM provider — so for conversation data, the jurisdiction that matters is the provider's, not ForgeCode's. No subprocessor list is published. The marketing site itself is served from United States infrastructure, which says nothing about where product data lives.
Things to keep in mind
Risks and trade-offs to weigh before adopting ForgeCode.
- The installer is a one-line remote script piped straight into a shell — a pattern many security policies forbid without prior review
- Restricted mode is off by default, and switching it on is not enough: the policy file ForgeCode generates allows every read, write, command and URL until you write rules of your own
- The forge agent writes files and runs shell commands on your machine; the documentation itself advises committing your work before letting it loose on anything significant
- Indexing a project uploads source chunks and embeddings to the publisher's servers, and they persist until you explicitly delete them — logging out does not remove them
- Delegating planning and reasoning to an agent day after day erodes the habit of holding a codebase in your own head, and the plan-then-execute split makes that delegation very easy to normalise
- The published prices are more than a year old and live in a blog post rather than on a pricing page, so budgeting on them carries real risk
- Account rules are strictly enforced: two devices maximum, monitored usage patterns, and permanent bans with no refund for API misuse or account sharing
Setup & Integrations
Technical difficulty
Moderate, and aimed squarely at people already at home in a terminal. The binary installs with a single command, but two prerequisites come first: Zsh installed and configured, and a Nerd Font enabled in the terminal. After the setup wizard runs, the terminal must be restarted before the colon trigger works — the documentation flags this as the most common failure. A provider API key or an existing ChatGPT Plus or Claude subscription is then needed, though ForgeCode Services alone requires neither. Windows needs WSL or Git Bash. Advanced tuning happens entirely in configuration files.
Deployment
Integrations
Behind ForgeCode
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement ForgeCode.
Frequently asked questions
Is ForgeCode open source?
Which operating systems does it run on?
Do I need an API key to get started?
Is there a free plan?
What do the paid plans cost?
Is my code used to train AI models?
Do my conversations pass through ForgeCode's servers?
How do I delete data that has already been indexed?
Can I stop the agent from running certain commands?
Can I share my account with my team?
Should you pick ForgeCode?
ForgeCode makes a clear bet: keep the developer in the terminal, and refuse to marry a single model vendor. That bet largely pays off. Six providers are supported out of the box, custom endpoints and self-hosted models are first-class citizens, and bringing your own API keys lifts usage limits altogether — an unusual degree of freedom in a market where most coding agents are the front end of one particular lab. The Apache-2.0 licence and the public Rust repository make that claim auditable rather than merely stated.
The privacy architecture deserves attention too. Because prompts and completions travel straight from the machine to the chosen provider, the publisher's servers see nothing of the conversation. Only the optional indexing step deposits source chunks and embeddings there, it is reversible, and the vendor commits to never training on that code.
The weaknesses are corporate rather than technical. There is no pricing page, no contact page, no about page and no postal address; the only rates on record sit in a blog post from July 2025, and the single published email belongs to the parent company's domain. Nothing resembling a DPA, a subprocessor list or a security certification exists, and no EU Article 27 representative is named — enough to stall a procurement review in a regulated environment. The Fair Usage Policy is unusually severe as well: two devices maximum, monitored usage, and bans applied without refund.
ForgeCode suits an individual developer or a small team at home in Zsh who want model choice without lock-in. Organisations with a compliance function should ask the publisher for what the site does not publish before committing. The domain was registered in May 2025: this is a young product moving fast, and its published paperwork has not yet caught up with its ambitions.
- Choosing a selection results in a full page refresh.
- Opens in a new window.